A photovoltaic power generation system output prediction system and method

Abstract

This invention relates to the field of photovoltaic power generation prediction technology, and discloses a photovoltaic power generation system output prediction system and method. The method collects historical power generation data of the target photovoltaic power station, as well as environmental monitoring data including irradiance, ambient temperature, module temperature, and cloud cover. This data is time-aligned to form a synchronized data sequence, from which meteorological and power generation characteristics are extracted. A dynamic correlation model between meteorological and power generation characteristics is constructed, and the instantaneous impact weight of meteorological fluctuations on power generation is calculated. Based on this, sensitive and non-sensitive periods for power generation are defined. High-fluctuation meteorological characteristics are extracted during sensitive periods, and steady-state meteorological characteristics are extracted during non-sensitive periods. The high-fluctuation meteorological characteristics are input into a first prediction network to generate short-term power prediction values, and the steady-state meteorological characteristics are input into a second prediction network to generate long-term power prediction values. The two are then fused to generate a complete output power prediction curve.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation prediction technology, specifically to a photovoltaic power generation system output prediction system and method. Background Technology

[0002] With the ever-growing global demand for clean energy and increasing emphasis on environmental protection, photovoltaic (PV) power generation, as a sustainable and clean energy solution, has experienced rapid development worldwide. In recent years, various regions have significantly increased their investment in technological research and development for the PV industry, comprehensively promoting the popularization and application of PV power generation. Related data shows that over the past decade, global PV power generation capacity has continued to climb at a double-digit annual growth rate, with more and more large-scale PV power plants springing up around the world, and distributed PV power generation being widely applied in residential buildings, commercial buildings, and other fields. Today, PV power generation has become an indispensable and important component of the energy structure, playing a crucial role in achieving global energy transition and addressing climate change goals.

[0003] Despite the promising future of photovoltaic (PV) power generation, its output power characteristics face numerous challenges. PV power generation relies on solar radiation, and its output power is directly affected by sunlight intensity. Power generation is high during the day when there is ample sunlight, while power generation decreases significantly or even disappears at night or on cloudy days. Simultaneously, ambient temperature and module temperature also significantly impact the conversion efficiency of PV cells. Generally, increased temperature leads to a decrease in PV cell conversion efficiency, thus affecting power generation. Changes in cloud cover also cause instantaneous fluctuations in sunlight intensity, resulting in unstable PV power generation. The combined effect of these meteorological factors leads to significant fluctuations and intermittency in PV power output.

[0004] This volatility and intermittency pose a significant challenge to the stable operation of the power system. When a large amount of photovoltaic power is connected to the grid, if its output power cannot be accurately predicted, it is difficult to rationally arrange power dispatch, which may lead to grid supply and demand imbalance, causing problems such as voltage fluctuations and frequency instability, and in severe cases, even threatening the safe operation of the grid. In addition, for energy management departments, the inability to accurately grasp the output of photovoltaic power generation will also affect the scientific and rational nature of energy planning and decision-making. Summary of the Invention

[0005] The purpose of this invention is to provide a photovoltaic power generation system output prediction system and method to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for predicting the output of a photovoltaic power generation system, the method comprising:

[0007] Collect historical power generation data and environmental monitoring data of the target photovoltaic power station. The environmental monitoring data includes at least light intensity, ambient temperature, module temperature, and cloud cover.

[0008] Historical power generation data and environmental monitoring data are time-aligned to form a synchronous data sequence, and meteorological and power generation characteristics are extracted from the synchronous data sequence.

[0009] A dynamic correlation model between meteorological characteristics and power generation characteristics is constructed, and the instantaneous impact weight of meteorological fluctuations on power generation is calculated through the dynamic correlation model.

[0010] Based on the instantaneous impact weight, the power generation period is divided into sensitive and non-sensitive periods. High-fluctuation meteorological characteristics are extracted in the sensitive period and steady-state meteorological characteristics are extracted in the non-sensitive period.

[0011] High-fluctuation meteorological characteristics are input into the first prediction network to generate short-term power prediction values, and steady-state meteorological characteristics are input into the second prediction network to generate long-term power prediction values.

[0012] By combining short-term and long-term power forecasts, a complete output power forecast curve for the target photovoltaic power plant is generated.

[0013] Based on the deviation between the complete output power prediction curve and the actual power generation data, the instantaneous influence weights in the dynamic correlation model are dynamically adjusted.

[0014] The threshold for dividing sensitive and non-sensitive periods is updated based on the adjusted instantaneous impact weight;

[0015] The updated segmentation threshold is used to classify the subsequently collected environmental monitoring data by time period, and the complete output power prediction curve is iteratively optimized.

[0016] Preferably, the extraction of meteorological and power generation characteristics from the synchronized data sequence includes:

[0017] Identify abrupt changes and gradual transitions in light intensity within a synchronized data sequence;

[0018] Calculate the temperature gradient between ambient temperature and component temperature at different time scales;

[0019] Analyze the time lag between the rate of change in cloud cover and the rate of decrease in power generation;

[0020] The abrupt change point, gradual change range, temperature gradient, and lag time difference are combined to form meteorological characteristics;

[0021] The maximum, minimum, and fluctuation frequency of power generation within a preset time window are extracted as power generation characteristics.

[0022] Preferably, the construction of the dynamic correlation model between meteorological characteristics and power generation characteristics includes:

[0023] Establish a mapping table between the magnitude of changes in meteorological characteristics and the magnitude of fluctuations in power generation characteristics;

[0024] Calculate the power generation adjustment corresponding to a unit change in light intensity based on the mapping table;

[0025] The initial value of the instantaneous impact weight is determined based on the product of the power generation adjustment and the rate of change of ambient temperature.

[0026] The initial value of the instantaneous influence weight is corrected by the number of times meteorological characteristics and power generation characteristics are matched in historical data.

[0027] Preferably, the step of dividing the power generation period into sensitive and non-sensitive periods based on the instantaneous impact weight includes:

[0028] The cumulative change of instantaneous impact weights over a continuous time interval;

[0029] When the cumulative change exceeds the first preset threshold, the time interval is marked as a sensitive period.

[0030] When the cumulative change is below the second preset threshold and the duration exceeds the preset length, the time interval is marked as a non-sensitive period.

[0031] For transitional intervals that do not meet the standards for sensitive and non-sensitive time periods, interpolation is performed using the marking results of adjacent time periods.

[0032] Preferably, the extraction of highly fluctuating meteorological features during sensitive periods includes:

[0033] Extract data segments within a preset time range before and after the point of sudden change in light intensity during sensitive periods;

[0034] Calculate the covariance matrix of cloud cover and ambient temperature in the data segment;

[0035] The eigenvectors of the principal component directions are extracted from the covariance matrix as high-fluctuation meteorological features.

[0036] Preferably, the extraction of steady-state meteorological features during non-sensitive periods includes:

[0037] Calculate the moving average of light intensity during non-sensitive periods;

[0038] Fit a linear regression equation between ambient temperature and component temperature;

[0039] The moving average value is combined with the slope of the linear regression equation to form a steady-state meteorological characteristic.

[0040] Preferably, the fusion of short-term power prediction and long-term power prediction includes:

[0041] Calculate the first and last differences of the short-term power forecast at the boundary of the sensitive period;

[0042] The amplitude of the long-term power forecast at the corresponding time point is adjusted based on the difference between the first and last values.

[0043] A weighted average algorithm is used to smoothly transition the adjusted long-term power forecast values ​​with the short-term power forecast values ​​in the overlapping interval.

[0044] Preferably, the dynamic adjustment of the instantaneous influence weights in the dynamic correlation model includes:

[0045] Calculate the root mean square error between the complete output power prediction curve and the actual power generation data;

[0046] The instantaneous impact weights are scaled proportionally based on the ratio of the root mean square error to a preset error threshold.

[0047] When the scaled instantaneous impact weight exceeds the historical maximum value, the backup weight in the backup weight library is used for replacement.

[0048] Preferably, the iterative optimization of the complete output power prediction curve includes:

[0049] The updated classification threshold will be applied to the latest collected environmental monitoring data;

[0050] Based on the latest classification results, high-fluctuation meteorological characteristics and steady-state meteorological characteristics were re-extracted;

[0051] A new round of short-term and long-term power predictions are generated using the re-extracted features.

[0052] Repeat the fusion and adjustment steps until the fluctuation range of the complete output power prediction curve tends to stabilize.

[0053] Preferably, the present invention also includes a photovoltaic power generation system output prediction system, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the photovoltaic power generation system output prediction method described above.

[0054] Compared with the prior art, the beneficial effects of the present invention are:

[0055] Historical power generation data and environmental monitoring data of the target photovoltaic power station were collected. The environmental monitoring data covered key elements such as solar irradiance, ambient temperature, module temperature, and cloud cover. Based on this, time alignment processing was performed on these data. This process effectively solved the problem of asynchronous data sources, ensuring that each power generation data point could be accurately matched with the corresponding meteorological and environmental data. This greatly guaranteed data consistency and reliability, providing a high-quality data foundation for subsequent in-depth analysis and modeling. After forming a synchronized data sequence, meteorological and power generation characteristics were extracted using scientific methods. These characteristics comprehensively and meticulously reflect the operating status and influencing factors of the photovoltaic power generation system. For example, meteorological characteristics such as the trend of solar irradiance variation and the range of temperature fluctuations, and power generation characteristics such as peak and valley values ​​of power output and the rate of power change, provide rich information for a deeper understanding of the operating rules and influencing factors of the photovoltaic power generation system.

[0056] A dynamic correlation model between meteorological characteristics and power generation characteristics is constructed, which can deeply reveal the complex dynamic relationship between meteorological factors and power generation characteristics. In actual operation, meteorological conditions are constantly changing, and traditional static models are difficult to accurately describe the impact of such dynamic changes on power generation. The dynamic correlation model of this invention, through learning and analysis of a large amount of historical data, can capture the impact of minute changes in meteorological factors on power generation. By calculating the instantaneous impact weight of meteorological fluctuations on power generation, it provides a crucial basis for photovoltaic power generation prediction. For example, when rapid cloud movement causes a significant instantaneous change in sunlight intensity, the model can quickly calculate the instantaneous impact weight of this change in sunlight intensity on power generation based on historical data and real-time monitored meteorological information. This allows for a more accurate prediction of power generation changes, making the prediction results more consistent with actual conditions and providing a more reliable reference for power dispatch and energy management.

[0057] Based on the instantaneous impact weights calculated using a dynamic correlation model, power generation is divided into sensitive and non-sensitive periods. During sensitive periods, even minor fluctuations in meteorological factors can trigger significant changes in power generation. In these periods, focusing on extracting highly volatile meteorological features allows for a more accurate grasp of dynamic changes in power generation. For example, in situations involving rapid cloud movement or sudden weather changes, meteorological factors such as sunlight intensity and temperature fluctuate dramatically. By focusing on these highly volatile meteorological features, rapid changes in power generation can be detected in a timely manner, providing accurate forecasting information for short-term power dispatch. During non-sensitive periods, meteorological conditions are relatively stable, and power generation is less affected by meteorological factors. Extracting steady-state meteorological features during these periods allows for a more effective analysis of the long-term trends and stability of power generation. This targeted feature extraction approach, based on the characteristics of different time periods, significantly improves the relevance and accuracy of forecasts, enabling the forecast results to better meet the needs of power dispatch and energy management at different times. Attached Figure Description

[0058] Figure 1 This is a schematic diagram illustrating the working principle of a photovoltaic power generation system output prediction system and method according to the present invention.

[0059] Figure 2 This is a flowchart for extracting meteorological and power generation characteristics from a synchronized data sequence;

[0060] Figure 3 A flowchart for constructing a dynamic correlation model between meteorological characteristics and power generation characteristics. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Please see Figure 1This invention provides a photovoltaic power generation system output prediction system and method. The method includes comprehensively processing historical data and real-time monitoring information to achieve accurate power prediction. Historical power generation data and environmental monitoring data of the target photovoltaic power station are collected. The environmental monitoring data includes at least solar irradiance, ambient temperature, module temperature, and cloud cover. The historical power generation data and environmental monitoring data are time-aligned to form a synchronous data sequence. Meteorological features and power generation features are extracted from the synchronous data sequence. A dynamic correlation model between the meteorological features and power generation features is constructed, and the instantaneous impact weight of meteorological fluctuations on power generation is calculated using the dynamic correlation model. Based on the instantaneous impact weight, power generation is divided into sensitive and non-sensitive periods. High-fluctuation meteorological features are extracted in the sensitive periods, and steady-state meteorological features are extracted in the non-sensitive periods. The high-fluctuation meteorological features are input into a first prediction network to generate short-term power prediction values, and the steady-state meteorological features are input into a second prediction network to generate long-term power prediction values. The short-term and long-term power prediction values ​​are fused to generate a complete output power prediction curve for the target photovoltaic power station. The deviation between the complete output power prediction curve and the actual power generation data is used to dynamically adjust the instantaneous impact weight in the dynamic correlation model. The adjusted instantaneous impact weights update the thresholds for distinguishing between sensitive and non-sensitive periods. These updated thresholds are then applied to subsequent environmental monitoring data for period classification, and the complete output power prediction curve is optimized through an iterative process. The data acquisition process involves multiple sensors and recording devices to ensure the integrity and accuracy of historical power generation data and environmental monitoring data. Time alignment employs a timestamp matching algorithm to eliminate temporal discrepancies between different data sources, and synchronized data sequences are stored at fixed time intervals. Meteorological feature extraction focuses on the changing patterns of light intensity, ambient temperature, component temperature, and cloud cover, while power generation features reflect the dynamic behavior of power output. A dynamic correlation model establishes the mathematical relationship between meteorological variables and power generation, and instantaneous impact weights quantify the immediate effects of meteorological changes. Period division is based on accumulated weights; sensitive periods correspond to periods of severe meteorological fluctuations, while non-sensitive periods correspond to stable conditions. The feature extraction method adapts to the characteristics of different time periods; high-fluctuation meteorological features capture short-term disturbances, while steady-state meteorological features describe long-term trends. The prediction network adopts a machine learning architecture, with the first prediction network handling high-frequency changes and the second handling low-frequency changes. A fusion algorithm coordinates prediction results across different time scales, and a bias feedback mechanism enables adaptive adjustment of model parameters. Threshold updates ensure that time period classifications match real-time data, and iterative optimization gradually improves the reliability of the prediction curve.

[0063] Example 1: See Figure 2The synchronous data sequence is formed by time-aligning historical power generation data and environmental monitoring data. The environmental monitoring data includes light intensity, ambient temperature, module temperature, and cloud cover. Extraction operations aim to capture the dynamic relationship between meteorological variables and power generation, providing input for subsequent model construction. Identifying abrupt changes and gradual transitions in light intensity within the synchronous data sequence is the initial step. Abrupt changes refer to moments when the light intensity value changes significantly within a very short time. Abrupt change detection uses a difference algorithm to compare the intensity values ​​of consecutive sampling points; when the difference between adjacent points exceeds a preset threshold, it is marked as an abrupt change location. The preset threshold is set based on the historical data fluctuation range. Gradual transitions describe the stages of slow increase or decrease in light intensity. Gradual transitions are divided using a sliding window averaging method, with the window size adjusted according to the data sampling frequency to ensure smooth changes within the interval. The temperature gradient between ambient temperature and module temperature is calculated at different time scales. The temperature gradient reflects the rate and direction of temperature change. Different time scales include minute, hour, and day. The minute gradient captures short-term fluctuations, the hourly gradient captures medium-term trends, and the daily gradient reflects long-term patterns. The calculation process involves temporal differences between the ambient temperature sequence and the component temperature sequence. For each time point, the derivative of the temperature difference is calculated using numerical differentiation methods such as the central difference method to reduce edge errors. The temperature gradient values ​​are stored as multi-dimensional vectors, with the vector dimensions corresponding to the number of time scales, facilitating subsequent feature integration.

[0064] This study analyzes the lag time difference between the rate of change in cloud cover and the rate of decrease in power generation, as the lag time difference indicates the delayed effect of meteorological changes on power generation. The rate of change in cloud cover is derived from the cloud cover sequence by calculating the difference in cover at consecutive time points and dividing by the time interval. The rate of decrease in power generation is extracted from the power generation sequence, identifying the power decrease segment and calculating the slope of the decrease. Lag time difference analysis uses a cross-correlation function to compare the time offsets of the cloud cover change rate sequence and the power generation decrease rate sequence, finding the offset corresponding to the maximum value of the cross-correlation function as the lag time difference. The cross-correlation calculation covers a reasonable time range to avoid excessive delay. The abrupt change in light intensity, the gradual change interval in light intensity, the temperature gradient between ambient temperature and component temperature, and the lag time difference between the rate of change in cloud cover and the rate of decrease in power generation are combined into meteorological features. This combination process involves the construction of feature vectors. Information on abrupt change in light intensity includes the abrupt change location and timestamp, while information on gradual change intervals in light intensity includes the start point, end point, and average rate of change of the interval. The gradient values ​​of ambient temperature and component temperature difference are arranged by time scale, and the time difference between the rate of change in cloud cover and the rate of decrease in power generation are used as scalar values. After normalization, all elements are concatenated into a fixed-dimensional feature vector, and the vector length is adjusted according to the data resolution to ensure feature consistency.

[0065] The maximum, minimum, and fluctuation frequency of power generation within a preset time window are extracted as power generation features. The size of the preset time window is dynamically set according to forecasting needs; for example, a 5-minute window is used for short-term forecasting, and a 1-hour window is used for long-term forecasting. The maximum and minimum power generation values ​​are directly obtained from the power sequence within the window, identifying the power output range. The fluctuation frequency is calculated by counting the number of times the power generation value crosses the average line or by using Fourier transform to extract the dominant frequency component, quantifying the frequency of power changes. Power generation features and meteorological features are stored synchronously to form a paired dataset for model training. The data preprocessing stage ensures the quality of the synchronous data sequence. Outlier detection uses the Z-score method to remove outliers, and data smoothing applies moving averages or low-pass filters to reduce noise. The feature extraction algorithm is implemented as a software module. The module takes the synchronous data sequence as input and outputs meteorological feature vectors and power generation feature vectors. The feature extraction process is executed in real time, adapting to streaming data input. In the detection of abrupt changes in light intensity, the threshold of the difference algorithm is adaptively adjusted according to the season and weather pattern to avoid false detections. The division of gradual light intensity change intervals is combined with clustering algorithms such as K-means to classify intervals with similar change patterns, improving interval consistency. The calculation of the temperature gradient between ambient temperature and component temperature takes into account sensor errors, and the gradient is calculated only after temperature data is calibrated. When integrating gradients at different time scales, weight allocation is based on scale importance, with higher weights for minute-scale data to emphasize short-term changes. In the lag time difference analysis of the rate of change in cloud cover and the rate of decrease in power generation, the cross-correlation function calculation optimizes computational efficiency, and a Fast Fourier Transform is used for acceleration. The lag time difference values ​​are verified through backtesting of historical events to ensure physical rationality.

[0066] When combining meteorological features, feature selection techniques such as principal component analysis are applied to high-dimensional vectors to retain the components with the largest variance, reducing computational burden. In power generation feature extraction, the preset time window sliding step size is matched with the sampling interval to avoid information loss. Zero-crossing detection or wavelet analysis is used to calculate fluctuation frequencies, adapting to non-stationary power sequences. Feature data is formatted into standard structures, such as CSV or JSON, for easy reading by subsequent models. The entire extraction process is integrated into the photovoltaic power generation system output prediction workflow. The software implementation uses Python or MATLAB, and the feature extraction module interfaces with the data acquisition system. In real-time processing, a data pipeline ensures low latency, and feature extraction time is constrained within the prediction period. Feature effectiveness is evaluated through cross-validation, using historical data to train the prediction model and test feature importance. Algorithm parameters for detecting abrupt changes in light intensity, such as the difference threshold, are optimized through grid search to balance sensitivity and specificity. In processing gradual light intensity ranges, the window size is automatically adjusted based on data variance; when the variance is high, the window shrinks to capture details.

[0067] The calculation of the temperature difference gradient between ambient temperature and component temperature incorporates a temperature compensation factor to compensate for sensor drift. The calculation of cloud cover change rate considers the influence of cloud type, distinguishing between thick and thin cloud patterns. Lag time difference analysis extends to multivariate scenarios, simultaneously handling the lag relationship between multiple meteorological factors and power generation. The meteorological feature combination algorithm supports dynamic dimensions, allowing seamless integration of new features such as humidity or wind speed. Power generation feature extraction enhances robustness, employing interpolation methods to handle missing data. The feature storage database adopts a time-series database structure, such as InfluxDB, supporting efficient queries. The feature extraction module logs operational status for easy debugging and monitoring. The overall implementation emphasizes scalability, with the feature pipeline designed as a microservice architecture for independent deployment. Light intensity abrupt change point detection integrates a machine learning classifier, such as a support vector machine, to distinguish between real abrupt changes and noise. Light intensity gradual change interval analysis combines trend fitting, with a linear regression model fitting the interval change trend. The temperature difference gradient calculation incorporates a time decay factor, with higher weights for recent gradients. Lag time difference analysis uses a dynamic time warping algorithm to handle non-uniformly sampled data.

[0068] After combining meteorological features, feature scaling applies min-max normalization to ensure feature values ​​are within a uniform range. In power generation feature extraction, parallel algorithms are used to calculate power statistics within the window, improving processing speed for large datasets. Fluctuation frequency analysis, combined with power spectral density estimation, identifies periodic patterns. Feature extraction output matches the input format of the dynamic correlation model, ensuring end-to-end compatibility. Implementation details are documented, with code comments detailing the logic of each step. Test cases cover various weather conditions to verify the stability of feature extraction. Continuous integration automatically runs regression tests, ensuring code changes do not affect functionality. Feature extraction performance optimization involves algorithm selection, such as using C++ to write core modules to accelerate computation. Memory management optimization reduces overhead when processing large sequences. Real-time detection of abrupt changes in light intensity requires high performance; sliding window incremental calculation is used to avoid reprocessing the entire sequence. Light intensity gradient interval division uses an online clustering algorithm with dynamically updated interval boundaries. Temperature gradient calculation is accelerated using GPUs, enabling parallel processing across multiple time scales. Lag time difference analysis caches cross-correlation results to reduce redundant calculations. Vectorized operations for meteorological feature combination are implemented efficiently using NumPy or similar libraries. The power generation feature extraction module integrates a signal processing toolkit to accurately calculate fluctuation frequencies. The entire implementation emphasizes code readability and maintainability, with a modular design for easy future expansion. The feature extraction module interfaces with the prediction network in a standardized manner, supporting multiple data sources. A monitoring system tracks feature quality metrics, such as feature variance and correlation, triggering re-extraction when these metrics are abnormal. A version control system manages code changes, ensuring implementation traceability.

[0069] Example 2: See Figure 3The dynamic correlation model is constructed based on the mapping relationship between meteorological characteristics and power generation characteristics. The mapping relationship table records the correspondence between the magnitude of changes in meteorological characteristics and the fluctuation magnitude of power generation characteristics in historical data. Regression analysis is used to establish the mapping relationship table, with meteorological characteristic changes such as the magnitude of sudden changes in light intensity and temperature gradient values ​​as independent variables, and the fluctuation magnitude of power generation as the dependent variable for fitting. The least squares method is used to solve the coefficient matrix during the fitting process. The mapping relationship table is stored in a two-dimensional table format, with row indices representing the levels of meteorological characteristic changes and column indices representing the levels of power generation characteristic fluctuation magnitudes. The table cells are filled with the probability or weight value of the corresponding relationship. The power generation adjustment corresponding to a unit change in light intensity is calculated based on the mapping relationship table. A unit change in light intensity is defined as the change in light intensity per 100 watts per square meter. The power generation adjustment is obtained by finding the average power fluctuation magnitude corresponding to a 1-unit change in light intensity in the mapping relationship table. The calculation process involves taking the arithmetic mean of the power response to all unit change events in historical data. The power generation adjustment is stored as a scalar parameter in the model parameter library. The initial value of the instantaneous influence weight is determined by multiplying the power generation adjustment amount with the rate of change of ambient temperature. The rate of change of ambient temperature is obtained from the first derivative of the temperature series. The product operation reflects the combined effect of multiple meteorological factors. The initial value calculation formula is in the form of a linear combination, and the coefficients are calibrated through training data.

[0070] The initial value of the instantaneous impact weight is corrected by matching the number of times meteorological features and power generation features are matched in historical data. The number of matches is counted as the number of times meteorological feature changes and power generation feature fluctuations occur simultaneously within a sliding time window. The correction operation uses a weighted average algorithm. The initial value is multiplied by the matching number weight and then merged with the historical weight value. The matching number weight is positively correlated with the number of matches, and high-frequency matching improves the reliability of the weight value. The dynamic association model is implemented in real time in matrix form, and the model outputs the instantaneous impact weight time series. The cumulative change of the instantaneous impact weight is calculated in a continuous time interval with a fixed length of 15 minutes. The cumulative change is obtained by numerically integrating the instantaneous impact weight value within the interval using the trapezoidal rule. The cumulative change reflects the severity of the weight change. When the cumulative change exceeds a first preset threshold, the corresponding time interval is marked as a sensitive period. The first preset threshold is set according to the percentile of the historical weight fluctuation range, usually the 75th percentile value. The sensitive period is marked as a binary flag sequence. When the cumulative change is lower than the second preset threshold and the duration exceeds the preset length, the time interval is marked as a non-sensitive period. The second preset threshold is the 25th percentile of the historical weight fluctuation range. The preset length is set to a continuous interval of 30 minutes according to the power plant's operating characteristics. Non-sensitive periods are marked and stored independently.

[0071] For transitional intervals that do not meet the standards for sensitive and non-sensitive time periods, interpolation is performed using the labeling results of adjacent time periods. The transitional interval refers to the area where the cumulative change is between a first and a second preset threshold. A linear interpolation algorithm is used, calculating the gradual label value of the transitional interval based on the labels of adjacent sensitive or non-sensitive time periods. This interpolation ensures the continuity of time period division. The time period division results are stored as time-label pairs to guide subsequent feature extraction. The mapping relationship table is maintained using a dynamic update mechanism. Newly collected data is validated and the table content is incrementally updated. The update strategy uses a rolling time window, retaining the data relationships of the most recent three months. The power generation adjustment corresponding to a unit change in light intensity is periodically recalculated, and the adjustment value is correlated with seasonal factors, establishing a seasonal correction coefficient table. The calculation of the ambient temperature change rate optimizes the sensor data sampling frequency, and an anti-aliasing filter is used to preprocess the temperature data. A temperature compensation factor is introduced into the calculation of the initial value of the instantaneous influence weight to compensate for the differences in influence between different temperature ranges.

[0072] Match count statistics utilize a hash table structure to accelerate queries, with the hash key being a combined encoding of meteorological and power generation characteristics. The cumulative change calculation employs an optimized numerical integration algorithm, using Simpson's rule to improve accuracy. An adaptive adjustment mechanism is implemented for the first and second preset thresholds, dynamically updating the threshold size based on recent weight distribution. Sensitive period marker sequences undergo morphological filtering to eliminate isolated marker points. Stability checks are added for non-sensitive period detection, requiring that the variance of weights within the interval simultaneously be below a specified value. Spline interpolation is introduced for transition interval interpolation to smooth transition boundaries. The dynamic association model construction and period segmentation are integrated through software modules, with inputs being meteorological and power generation characteristic sequences and output being period segmentation markers. Model parameters are stored in a relational database, supporting transaction rollback and version management. Instantaneous impact weight calculation is threaded to meet real-time requirements. Period segmentation results are visualized to assist maintenance personnel in verifying the rationality of the segmentation. Robust statistical methods are used to handle abnormal data during the mapping relationship table establishment process, eliminating the influence of outliers. Power generation adjustment calculations are modeled separately for different weather types, distinguishing between sunny, cloudy, and rainy scenarios.

[0073] The calculation of the rate of change in ambient temperature integrates data from multiple temperature sensors, and improves reliability through weighted averaging. A time decay factor is added to the instantaneous impact weighting process, with higher weight given to recent matching frequencies. Cumulative change statistics are normalized to eliminate the influence of dimensions. A delayed confirmation mechanism is added to the sensitive period marking logic to avoid misjudgments due to instantaneous fluctuations. Non-sensitive period detection is combined with power generation stability indicators for double verification of stability. The boundary handling of the transition interval interpolation algorithm is optimized to prevent overflow of marked values. The entire implementation process emphasizes a balance between computational efficiency and accuracy; the core algorithm is written in C++, and the business logic is encapsulated in Python. Dynamic association model version control is implemented using Git management, ensuring complete traceability of model updates. The time period segmentation results are compared and analyzed with real-time power generation data, with segmentation accuracy used as a model optimization indicator. A machine learning classification algorithm is introduced into the mapping relationship table to automatically identify important feature combinations. The calculation of power generation adjustment corresponding to unit light intensity change is modeled by time period, distinguishing between low-intensity periods in the morning and evening and high-intensity periods at noon. The calculation of the rate of change in ambient temperature considers the terrain shading effect and incorporates azimuth correction.

[0074] The initial value calculation of instantaneous impact weights incorporates component temperature change factors and expands into a multi-dimensional weight vector. Distributed computing is implemented for matching frequency statistics to handle large-scale historical data. Cumulative change calculation uses a sliding window incremental update to reduce computational complexity. An expert rule engine is used to set the first and second preset thresholds, incorporating domain knowledge. Kalman filtering is applied after marking sensitive periods to smooth time series fluctuations. Weather forecast information is integrated for detecting non-sensitive periods to predict the duration of stable periods. Machine learning prediction models are introduced for transition interval interpolation to improve interpolation accuracy. The system is deployed on a cloud computing platform to achieve elastic scaling and load balancing. Model construction focuses on mining quantitative relationships between features, and time period division emphasizes the cumulative effect of changes. A multi-level optimization strategy is adopted during implementation, with quality control throughout the process from data preprocessing to result post-processing. Dynamically correlated model parameters are periodically retrained to adapt to changes in the power plant's operating status. Time period division thresholds are self-learned and updated to maintain classification accuracy.

[0075] Example 3: Extraction of highly fluctuating meteorological features within sensitive periods begins with extracting data segments within a preset time range before and after a sudden change in light intensity during the sensitive period. The preset time range is set to 5 minutes before and after the sudden change point based on meteorological change characteristics, and the data segment contains a complete fluctuation cycle. The covariance matrix of cloud cover and ambient temperature in the data segment is calculated. The covariance matrix quantifies the cooperative relationship between the two meteorological variables during the sensitive period, and the matrix element values ​​reflect the strength of the linear correlation between the variables. Eigenvectors of the principal component directions are extracted from the covariance matrix as highly fluctuating meteorological features. The eigenvectors correspond to the largest eigenvalue of the covariance matrix, representing the direction of the greatest data variation. The data segment extraction operation uses a sliding window algorithm, with the window center aligned with the timestamp of the sudden change in light intensity. The window size is dynamically adjusted according to historical fluctuation patterns. Covariance matrix calculation uses vectorized operations, calculating the expected value of the outer product after centering the cloud cover and ambient temperature sequences. Principal component analysis uses a power iteration method to solve for the eigenvectors, ensuring computational efficiency meets real-time requirements. The highly fluctuating meteorological features are stored as normalized vectors, with the vector dimension consistent with the rank of the covariance matrix. Extracting steady-state meteorological features during non-sensitive periods involves calculating a moving average of light intensity during these periods. The moving average is calculated using an exponentially weighted moving average algorithm, with a decay factor set to 0.2 to balance smoothness and response speed. A linear regression equation is fitted between ambient temperature and component temperature. The linear regression uses the least squares method to estimate the slope and intercept parameters, and the equation is a linear function of component temperature with respect to ambient temperature. The moving average and the slope of the linear regression equation are combined to form the steady-state meteorological features. The moving average represents the baseline level of light intensity, and the slope reflects the stability of the temperature relationship. The moving average calculation uses a recursive formula for streaming processing, avoiding the need to store complete historical data. A regularization term is introduced into the linear regression fitting to prevent overfitting; the regularization coefficient is determined through cross-validation. The steady-state meteorological features are combined into a two-dimensional vector, with the first dimension storing the moving average and the second dimension storing the regression slope. The feature values ​​are scaled to a uniform dimension for easier processing by the subsequent prediction network.

[0076] The high-fluctuation meteorological feature extraction process optimizes the data segment truncation strategy, with segment length adaptively adjusted based on the intensity of abrupt changes, assigning longer segments to strong abrupt changes. Robustness processing is added to the covariance matrix calculation, employing a minimum covariance decision algorithm to eliminate the influence of outliers. Principal component analysis is extended to an incremental update mode, with feature vectors updated online as new data arrives. Feature vector selection criteria balance interpretability and physical meaning, retaining the top k principal components. In steady-state meteorological feature extraction, the moving average calculation optimizes weight allocation, giving higher weight to recent data. Linear regression equation fitting incorporates a time variable as a covariate to capture the temporal evolution of temperature relationships. Correlation analysis is performed before feature combination, removing highly redundant feature dimensions. The overall feature extraction process is parallelized, with threads running independently for sensitive and non-sensitive periods. The data preprocessing stage standardizes meteorological data sampling frequency, and missing values ​​are filled using spline interpolation. The feature extraction algorithm is implemented as a C++ extension module with a Python encapsulated interface. The high-fluctuation meteorological feature extraction module integrates real-time anomaly detection, automatically discarding outlier segments. The steady-state meteorological feature extraction module incorporates trend testing, triggering recalculation during non-stationary periods. Feature storage employs a columnar database to optimize query efficiency, and feature vectors support fast retrieval. The covariance matrix calculation in high-fluctuation meteorological feature extraction utilizes SIMD instructions to accelerate matrix operations. Principal component analysis uses a stochastic PCA algorithm to process high-dimensional data. The moving average calculation in steady-state meteorological feature extraction employs a double-buffer mechanism to ensure data consistency.

[0077] The system integrates abrupt change detection and feature extraction modules, dynamically adjusting the confidence threshold for these changes. Covariance matrix calculation considers the nonlinear relationships of meteorological variables, incorporating a kernel function method. Principal component eigenvectors are visualized to aid feature interpretability analysis. Moving average calculation incorporates seasonal patterns, using different decay factors for different seasons. Variance inflation factor detection is added to the linear regression equation fitting to eliminate multicollinearity. Feature selection is performed after feature combination, prioritizing features with high information content based on mutual information ranking. The high-fluctuation meteorological feature extraction module outputs feature importance scores to guide model optimization. The steady-state meteorological feature extraction module integrates change point detection, automatically identifying changes in feature patterns. The system deploys a distributed computing framework, scheduling feature extraction tasks for parallel execution across multiple nodes. The high-fluctuation meteorological feature extraction algorithm optimizes memory usage, employing a sparse matrix to store the covariance structure. The steady-state meteorological feature extraction algorithm uses incremental learning updates to adapt to data distribution drift. The feature extraction pipeline monitors feature quality indicators, triggering alarms for abnormal features. High-fluctuation meteorological features capture rapid change patterns during sensitive periods, while steady-state meteorological features describe stable patterns during non-sensitive periods. The extraction process emphasizes both feature representativeness and computational efficiency, while feature design considers both statistical properties and physical meaning. The implementation phase employs a multi-level optimization strategy, comprehensively considering real-time performance and accuracy requirements from algorithm selection to engineering implementation.

[0078] The covariance matrix in high-fluctuation meteorological feature extraction is calculated using the following formula to characterize the relationship between variables:

[0079]

[0080] in: The covariance matrix representing cloud cover and ambient temperature. This represents the total number of sampling points in the data segment. This represents the cloud cover value at the i-th sampling point. This represents the average cloud cover within a data segment. This represents the ambient temperature value at the i-th sampling point. This represents the average ambient temperature within the data segment. This represents the matrix transpose operation. The covariance matrix calculation formula quantifies the degree of coordinated change between cloud cover and ambient temperature during sensitive periods. The eigenvectors of the covariance matrix are used to construct highly volatile meteorological features. The feature extraction module implements unit testing covering various meteorological scenarios to verify feature stability. Highly volatile meteorological feature extraction integrates multi-scale analysis, with features from different time scales being complementary. Steady-state meteorological feature extraction introduces sliding window verification to ensure feature consistency. The entire feature engineering pipeline has a fault-tolerant mechanism, and single-point failures do not affect the overall system operation. The feature extraction results strictly match the input format of the prediction network, and the data flow is seamlessly connected. The implementation documentation records parameter configurations and algorithm flows in detail, facilitating maintenance and optimization.

[0081] Example 4: The fusion calculation of short-term and long-term power predictions begins with calculating the difference between the beginning and end of the short-term power prediction at the sensitive time period boundary. The sensitive time period boundary is determined by the time point marked by the time period segmentation module. The difference is defined as the arithmetic difference between the short-term power prediction at the beginning and end of the sensitive time period. The amplitude of the long-term power prediction at the corresponding time point is adjusted based on this difference using a linear scaling algorithm. The long-term power prediction at the sensitive time period boundary is scaled proportionally to the difference. A weighted average algorithm is used to smoothly transition the adjusted long-term and short-term power predictions within the overlapping interval, which is set as a 10-minute time window before and after the sensitive time period boundary. The weighting coefficients of the weighted average algorithm change over time: the weight of the short-term power prediction decreases linearly from 1.0 at the beginning of the overlapping interval to 0.0 at the end, while the weight of the long-term power prediction increases linearly from 0.0 to 1.0. The fusion result generates a complete output power prediction curve, and the curve data is stored in time series format.

[0082] The instantaneous impact weights in the dynamic correlation model are dynamically adjusted based on the deviation analysis between the complete output power prediction curve and the actual power generation data. The root mean square error (RMSE) between the two data points is calculated, with the RMSE formula being the square root of the average of the squares of the differences between the predicted and actual values ​​at each time point. The instantaneous impact weights are scaled proportionally to the ratio of the RMSE to a preset error threshold. The preset error threshold is set to 5% of the rated power based on the historical operating performance of the photovoltaic power plant, and the scaling factor is equal to the quotient of the RMSE and the preset error threshold. When the scaled instantaneous impact weight exceeds the historical maximum value, a backup weight from the backup weight library is used for replacement. The historical maximum value records the peak instantaneous impact weight over the past 30 days, and the backup weight library stores the 10 best-performing weight values ​​recently. The replacement strategy uses the least recently used algorithm to select backup weights. The weight adjustment cycle is set to 15 minutes, synchronized with the data acquisition cycle.

[0083] The algorithm for calculating the first and last differences in the fusion calculation of short-term and long-term power forecasts is optimized to improve boundary point detection. Boundary point localization is combined with gradient detection to avoid misjudgments. A nonlinear correction factor is introduced for the amplitude adjustment of long-term power forecasts, dynamically calculated based on the rate of change of meteorological conditions. The weight allocation curve of the weighted average algorithm is replaced with a Sigmoid function to achieve a smooth transition and avoid abrupt changes in linearity. The fusion calculation module implements a real-time verification mechanism to check the physical rationality of short-term and long-term power forecasts. A moving average filter is used in the post-processing of the complete output power forecast curve to eliminate high-frequency noise introduced by fusion. The root mean square error calculation optimizes numerical stability processing, using the Kahan summation algorithm to reduce floating-point errors. A seasonal adaptive adjustment is implemented for the preset error threshold, with the summer threshold increased by 10% compared to winter. Boundary protection is added to the weight scaling operation, limiting the scaling factor to the range of 0.5 to 2.0. A backup weight library update mechanism incorporates weight performance evaluation, recording the recent prediction accuracy for each weight group. A rolling verification step is added to the weight replacement strategy; new weights must pass temporary testing before being officially used. Detailed logs are recorded during the dynamic adjustment process, supporting fault tracing and analysis.

[0084] The overlap interval length in the fusion calculation of short-term and long-term power forecasts is dynamically adjusted based on forecast uncertainty, extending the overlap interval when uncertainty is high. The weighted average weight function is configurable, supporting triangular window or Hanning window functions. A confidence interval calculation is added to the complete output power forecast curve, generated based on historical forecast error distribution. Root mean square error calculation differentiates between different weather types, using different error thresholds for sunny and cloudy days. A momentum factor is introduced in weight scaling to avoid excessive oscillation of weight values. A backup weight library is implemented with distributed storage and weight data synchronization across multiple nodes. The dynamic adjustment algorithm implements a degradation strategy, reverting to conservative weight values ​​when data is abnormal. The fusion calculation module for short-term and long-term power forecasts adopts a microservice architecture, supporting horizontal scaling. Fusion parameters are stored in a configuration center and updated in real-time. The weight adjustment task scheduling uses a time-wheel algorithm to ensure timing accuracy. The system monitors the data flow quality during the fusion process, triggering alarms in case of anomalies. The fusion algorithm emphasizes the smooth connection of forecast results at different time scales, and the weight adjustment mechanism ensures the model's adaptability to real-time data. Multi-level verification and protection measures are adopted during implementation to ensure stable system operation. Refer to Table 1, which shows the configuration of the weight coefficients in the weighted average algorithm over time. The data in the table is used to control the smoothness of the transition between short-term and long-term power predictions in the overlapping interval.

[0085] Table 1: Weighting coefficient configuration for weighted average algorithm

[0086] Time offset (minutes) Short-term forecast weighting coefficient Long-term forecast weighting coefficient -10 1.00 0.00 -8 0.95 0.05 -6 0.85 0.15 -4 0.70 0.30 -2 0.50 0.50 0 0.30 0.70 2 0.15 0.85 4 0.05 0.95 6 0.00 1.00

[0087] The time offset column in the table represents the time difference relative to the boundary of sensitive periods; negative values ​​indicate the time point before the boundary, and positive values ​​indicate the time point after the boundary. The sum of the short-term and long-term prediction weight coefficients is always 1, and the weight change curve ensures the continuity of fusion. The weight coefficient configuration can be adjusted according to the characteristics of the power plant to adapt to different operating scenarios. A quality assessment step is added to the fusion calculation of short-term and long-term power prediction values, and energy conservation verification is performed before and after fusion. The weight coefficient configuration table is version-managed, with different configuration versions corresponding to different seasons. The fusion algorithm test covers various boundary conditions, including the fusion of prediction values ​​under extreme weather conditions. The dynamic weight adjustment module implements an A / B testing framework, with the new algorithm and the old algorithm running in parallel to compare their effects. The complete output power prediction curve undergoes a rationality check before release, and the curve shape must conform to the physical characteristics of photovoltaic power generation. A weight discount factor is added to the root mean square error calculation, with recent errors having a higher weight than earlier errors. The weight scaling algorithm introduces smoothing processing to avoid drastic changes in the scaling factor. The spare weight library implements an automatic cleanup mechanism to discard weights that have not been used for a long time. The time synchronization accuracy in the fusion calculation of short-term and long-term power forecasts reaches the millisecond level, ensuring accurate alignment of data from different sources. The weighted average algorithm supports non-linear weight allocation to address fusion requirements under special meteorological conditions. The complete output power forecast curve is stored with accompanying metadata, recording fusion parameters and weight versions. A dynamically adjusted fault tolerance mechanism ensures that single-point failures do not affect basic forecasting functionality.

[0088] Example 5: Iterative optimization of the complete output power prediction curve begins by applying the updated segmentation threshold to the latest collected environmental monitoring data. The updated segmentation threshold comes from the new threshold set generated in the previous round of dynamic adjustment. The latest collected environmental monitoring data is acquired in real time through a sensor network deployed in the photovoltaic power station. The environmental monitoring data includes key parameters such as light intensity, ambient temperature, module temperature, and cloud cover, and the data continuously flows into the processing system at one-minute intervals. The segmentation threshold application process uses an online classification algorithm to divide the real-time data stream into continuous time segments, each segment being marked as a sensitive or non-sensitive period according to threshold rules. Based on the latest segmentation results, high-fluctuation meteorological features and steady-state meteorological features are re-extracted. The feature extraction module receives the time segmentation label sequence and performs the high-fluctuation meteorological feature extraction process on the intervals marked as sensitive periods. The high-fluctuation meteorological feature extraction focuses on identifying abrupt changes in light intensity and covariance analysis. After preprocessing, the data segments within the sensitive periods are used to calculate the co-variance characteristics of cloud cover and ambient temperature. For the intervals marked as non-sensitive periods, the steady-state meteorological feature extraction process is performed, calculating the moving average of light intensity and fitting a temperature relationship curve. The feature extraction process adopts an incremental calculation mode, processing only the changed parts when new data arrives, thus improving computational efficiency.

[0089] New short-term and long-term power predictions are generated using newly extracted features. Short-term predictions are generated by a first prediction network employing a Long Short-Term Memory (LSTM) neural network architecture, taking a highly volatile meteorological feature vector as input and outputting a power prediction sequence for the next 15 minutes. Long-term predictions are generated by a second prediction network employing a gated recurrent unit (GRU) network, taking a steady-state meteorological feature vector as input and outputting a power prediction sequence for the next 4 hours. Model quantization is used to accelerate the prediction network inference process, converting floating-point calculations to fixed-point operations to meet real-time requirements. The fusion and adjustment steps are repeated until the fluctuation amplitude of the complete output power prediction curve stabilizes. The fusion step aligns the short-term and long-term power predictions temporally and synthesizes the complete output power prediction curve using a weighted average algorithm. The adjustment step calculates the deviation between the newly generated complete output power prediction curve and the actual power generation data, dynamically adjusting the instantaneous influence weights in the dynamic correlation model based on the deviation magnitude. The iteration termination condition is set to the rate of change of the fluctuation amplitude of the complete output power prediction curve being less than a set threshold over three consecutive iterations. The fluctuation amplitude is measured by calculating the curve variance.

[0090] Taking the actual operating data of a photovoltaic power station on June 15, 2023 as an example, the weather conditions on June 15 showed a typical pattern of cloudy in the morning, sunny at noon, and cloudy again in the evening. At 6:00 AM, the data acquisition system recorded that the light intensity dropped sharply from 85 W / m² to 35 W / m², and the environmental monitoring data was transmitted to the processing system in real time. The updated threshold marked this period as a sensitive period, and the feature extraction module identified the point of sudden change in light intensity and extracted the data segment between 6:00 and 6:10. In the data segment, the cloud cover increased from 30% to 65%, and the ambient temperature dropped slightly from 21.5℃ to 21.2℃. The covariance matrix eigenvector was calculated to be [0.874, 0.486]. The high-fluctuation meteorological features were input into the first prediction network to generate short-term power prediction values ​​for 6:10-6:25, which showed that the power would decrease from 28% to 19% of the rated capacity. Simultaneously, steady-state feature extraction for the non-sensitive period between 5:30 and 6:00 yielded a moving average of 72 W / m² for light intensity, with a regression slope of 1.15 between ambient temperature and component temperature. The steady-state meteorological features were input into the second prediction network to generate long-term power predictions for 6:00-10:00, with the prediction curve showing a steady upward trend. The fusion module assigned weight coefficients in the overlapping interval between 6:00 and 6:10, with the short-term prediction weight linearly decreasing from 1.0 to 0.0, and the long-term prediction weight increasing accordingly.

[0091] After the complete output power prediction curve is generated, it is compared with the actual power generation data from 6:00 to 10:00. The calculated root mean square error is 3.7% of the rated power. The instantaneous impact weight is adjusted according to the error ratio, scaling the weight value from 0.82 to 0.79. The new iteration uses the updated weight to re-divide the time period; the 6:00-6:15 interval is still marked as a sensitive period, but the boundary is slightly adjusted. The feature extraction and prediction process is repeated, and the variance of the complete output power prediction curve generated in the second iteration decreases from 0.124 to 0.098. The third iteration continues to optimize the weight parameters; the curve variance change rate is 0.021, which is below the preset threshold of 0.03, and the iteration terminates. The iterative optimization module in the system implementation adopts a fault-tolerant design, automatically reverting to the previous stable version when a single iteration fails. Each iteration records the complete parameter change trajectory, supporting analysis and review of the optimization process. The stability judgment of the prediction curve introduces a multi-indicator comprehensive evaluation, including variance, entropy value, and autocorrelation coefficient. The iteration cycle is dynamically adjusted according to computing resources, with extended iteration intervals during high-load periods. A real-time data processing pipeline ensures low latency for iterative optimization, with a data buffer queue depth set to 5 minutes of data. Feature extraction and prediction network calls are asynchronous to avoid blocking the main data stream. Intermediate results during iterative optimization are cached in an in-memory database, accelerating data access across multiple iterations. The system monitoring module tracks performance metrics for each iteration, including computation time, memory usage, and prediction accuracy. The implementation of iterative optimization of the complete output power prediction curve emphasizes a closed-loop feedback mechanism, with each iteration using the latest actual data to correct model parameters. The optimization process follows a progressive refinement principle, with initial iterations focusing on rapid response and subsequent iterations gradually improving accuracy. The complete output power prediction curve after iteration termination is published to the power plant control system for power generation planning and grid dispatching. Historical iteration data is archived and analyzed to optimize algorithm parameters and improve system performance.

[0092] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0093] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the output of a photovoltaic power generation system, characterized in that, include: Collect historical power generation data and environmental monitoring data of the target photovoltaic power station. The environmental monitoring data includes at least light intensity, ambient temperature, module temperature, and cloud cover. Historical power generation data and environmental monitoring data are time-aligned to form a synchronous data sequence, and meteorological and power generation characteristics are extracted from the synchronous data sequence. A dynamic correlation model between meteorological characteristics and power generation characteristics is constructed, and the instantaneous impact weight of meteorological fluctuations on power generation is calculated through the dynamic correlation model. Based on the instantaneous impact weight, the power generation period is divided into sensitive and non-sensitive periods. High-fluctuation meteorological characteristics are extracted in the sensitive period and steady-state meteorological characteristics are extracted in the non-sensitive period. High-fluctuation meteorological characteristics are input into the first prediction network to generate short-term power prediction values, and steady-state meteorological characteristics are input into the second prediction network to generate long-term power prediction values. By combining short-term and long-term power forecasts, a complete output power forecast curve for the target photovoltaic power plant is generated. Based on the deviation between the complete output power prediction curve and the actual power generation data, the instantaneous influence weights in the dynamic correlation model are dynamically adjusted. The threshold for dividing sensitive and non-sensitive periods is updated based on the adjusted instantaneous impact weight; The updated segmentation threshold is used to classify the subsequently collected environmental monitoring data by time period, and the complete output power prediction curve is iteratively optimized. The construction of the dynamic correlation model between meteorological characteristics and power generation characteristics includes: Establish a mapping table between the magnitude of changes in meteorological characteristics and the magnitude of fluctuations in power generation characteristics; Calculate the power generation adjustment corresponding to a unit change in light intensity based on the mapping table; The initial value of the instantaneous impact weight is determined based on the product of the power generation adjustment and the rate of change of ambient temperature. The initial value of the instantaneous influence weight is corrected by the number of times meteorological characteristics and power generation characteristics are matched in historical data; The process of dividing power generation into sensitive and non-sensitive periods based on instantaneous impact weights includes: The cumulative change of instantaneous impact weights over a continuous time interval; When the cumulative change exceeds the first preset threshold, the time interval is marked as a sensitive period. When the cumulative change is below the second preset threshold and the duration exceeds the preset length, the time interval is marked as a non-sensitive period. For transitional intervals that do not meet the standards for sensitive and non-sensitive time periods, interpolation is performed using the marking results of adjacent time periods.

2. The photovoltaic power generation system output prediction method according to claim 1, characterized in that, The extraction of meteorological and power generation characteristics from the synchronized data sequence includes: Identify abrupt changes and gradual transitions in light intensity within a synchronized data sequence; Calculate the temperature gradient between ambient temperature and component temperature at different time scales; Analyze the time lag between the rate of change in cloud cover and the rate of decrease in power generation; The abrupt change point, gradual change range, temperature gradient, and lag time difference are combined to form meteorological characteristics; The maximum, minimum, and fluctuation frequency of power generation within a preset time window are extracted as power generation characteristics.

3. The photovoltaic power generation system output prediction method according to claim 1, characterized in that, The extraction of highly fluctuating meteorological features during sensitive periods includes: Extract data segments within a preset time range before and after the point of sudden change in light intensity during sensitive periods; Calculate the covariance matrix of cloud cover and ambient temperature in the data segment; The eigenvectors of the principal component directions are extracted from the covariance matrix as high-fluctuation meteorological features.

4. The photovoltaic power generation system output prediction method according to claim 3, characterized in that, The extraction of steady-state meteorological features during non-sensitive periods includes: Calculate the moving average of light intensity during non-sensitive periods; Fit a linear regression equation between ambient temperature and component temperature; The moving average value is combined with the slope of the linear regression equation to form a steady-state meteorological characteristic.

5. The photovoltaic power generation system output prediction method according to claim 4, characterized in that, The fusion of short-term and long-term power predictions includes: Calculate the first and last differences of the short-term power forecast at the boundary of the sensitive period; The amplitude of the long-term power forecast at the corresponding time point is adjusted based on the difference between the first and last values. A weighted average algorithm is used to smoothly transition the adjusted long-term power forecast values ​​with the short-term power forecast values ​​in the overlapping interval.

6. The photovoltaic power generation system output prediction method according to claim 5, characterized in that, The dynamic adjustment of instantaneous influence weights in the dynamic correlation model includes: Calculate the root mean square error between the complete output power prediction curve and the actual power generation data; The instantaneous impact weights are scaled proportionally based on the ratio of the root mean square error to a preset error threshold. When the scaled instantaneous impact weight exceeds the historical maximum value, the backup weight in the backup weight library is used for replacement.

7. The photovoltaic power generation system output prediction method according to claim 6, characterized in that, The iterative optimization of the complete output power prediction curve includes: The updated classification threshold will be applied to the latest collected environmental monitoring data; Based on the latest classification results, high-fluctuation meteorological characteristics and steady-state meteorological characteristics were re-extracted; A new round of short-term and long-term power predictions are generated using the re-extracted features. Repeat the fusion and adjustment steps until the fluctuation range of the complete output power prediction curve tends to stabilize.

8. A photovoltaic power generation system output prediction system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the photovoltaic power generation system output prediction method according to any one of claims 1 to 7.

Images