A photovoltaic power station power prediction method in a step-by-step manner
By adopting a step-by-step photovoltaic power plant power prediction method, and combining LSTM and random forest models to handle multiple nonlinear interference factors of photovoltaic power plants, a minute-level power prediction sequence is generated. This solves the prediction bias problem of photovoltaic power plants under rapidly changing weather conditions and achieves high-precision power prediction and grid dispatch support.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHAANXI HYDROPOWER DEVELOPMENT GROUP CO LTD
- Filing Date
- 2026-05-13
- Publication Date
- 2026-08-04
AI Technical Summary
Existing photovoltaic power plant power prediction methods are difficult to accurately predict under rapidly changing weather conditions. In particular, under the influence of factors such as severe radiation fluctuations caused by cloud movement, temperature changes, and equipment aging, the prediction deviation is large and the accuracy is insufficient at the minute level, which cannot meet the grid's rapid response requirements.
A step-by-step photovoltaic power plant power prediction method is adopted. By acquiring real-time radiation data and historical meteorological conditions, multiple interference factors are fused using LSTM and random forest models to generate minute-level power prediction sequences. The influence of cloud movement is refined through random forest regression, and finally an optimized power output curve is formed.
It significantly improves the minute-level forecasting accuracy and stability of photovoltaic power stations under complex weather conditions, supporting grid dispatch and efficient energy utilization.
Smart Images
Figure CN122203958B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of photovoltaic power generation technology, and in particular to a method for predicting the power of a photovoltaic power plant using a step-by-step approach. Background Technology
[0002] Photovoltaic power plants are playing an increasingly important role in the new energy system, and the accuracy of their power forecasting directly affects the stable dispatch of the power grid, the efficient utilization of energy, and the large-scale absorption capacity of renewable energy. Against the backdrop of deepening electricity market liberalization, accurate forecasting of photovoltaic power generation has become one of the key links in ensuring the safe and economical operation of the system.
[0003] Current forecasting methods have revealed significant shortcomings in practical applications. Most methods tend to treat the problem from a single perspective, leading to significant fluctuations in forecast bias when different weather conditions and operational scenarios change. Especially when ground weather changes rapidly, or when cloud movement causes severe radiation fluctuations, the forecast results often deviate significantly from the actual output, making it difficult to meet the grid's demand for minute-level or even second-level response. This instability makes dispatchers unable to trust the forecast values, thus limiting the deep involvement of photovoltaic power plants in the power system.
[0004] The core technical challenge in photovoltaic power prediction lies in the complex, multi-layered relationships among the factors influencing power generation. While the theoretical value of solar radiation can be precisely calculated using astronomical geometry, the actual radiation reaching the module surface is affected by various environmental disturbances such as atmospheric scattering, cloud cover, and dust accumulation. The real-time changes of these disturbances cannot be fully described by a single physical law. Furthermore, historical data contains factors such as equipment aging, efficiency degradation due to temperature increases, and the nonlinear effects of cloud type and wind speed on power output at different times, further compounding the complexity of the mapping from theoretical radiation to actual output. It is precisely because multiple dynamic interference layers exist between theoretical calculations and actual effects, and these interferences act together at different time and spatial scales, that it is difficult to capture the complete causal chain using any single fixed method, leading to systematic deviations in predictions under certain typical weather conditions. For example, in cloudy weather, the slow movement of clouds in the morning causes the power to slowly climb, while in the afternoon, a large area of thick clouds suddenly passes by rapidly, causing the power to drop sharply from a high level to near zero within minutes. This drastic and nonlinear jump process makes it difficult to simultaneously take into account the continuity of the trend and the agility of the sudden change in forecasting.
[0005] Therefore, how to effectively integrate the dynamic characteristics of multiple interference factors while preserving the clarity of physical laws, and at the same time grasp the multi-layered progressive mapping relationship from theoretical radiation to actual output, has become a key issue in improving the reliability of photovoltaic power plant power prediction. Summary of the Invention
[0006] To address the shortcomings of existing technologies, this invention provides a step-by-step method for predicting the power of photovoltaic power plants. This method solves the problems of large power prediction deviations and insufficient minute-level accuracy caused by multiple nonlinear interference factors such as severe radiation fluctuations due to rapid cloud movement, superimposed temperature changes, and equipment aging.
[0007] To achieve the above objectives, the technical solution of the present invention is as follows: A step-by-step method for predicting the power output of a photovoltaic power plant, comprising the following steps: Step S1: Obtain real-time radiation data and historical meteorological condition records of the photovoltaic power station, integrate interference factors to calculate the initial radiation fluctuation amplitude, and determine the initial deviation fluctuation range. Step S2: Based on the preliminary deviation fluctuation range, the nonlinear change sequence is processed using a Long Short-Term Memory (LSTM) model to extract the cloud movement pattern from the historical meteorological condition records and obtain the dynamic interference adjustment coefficient. Step S3: If the dynamic interference adjustment coefficient exceeds the preset threshold, random forest classification is performed on the radiation jitter to determine the current meteorological condition type and obtain the power prediction correction value after classification. Step S4: Based on the classified power prediction correction value, obtain the current temperature influence parameters and equipment aging parameters, integrate the temperature influence parameters and equipment aging parameters into the nonlinear change model, and determine the comprehensive interference compensation vector; Step S5: Based on the comprehensive interference compensation vector, the deviation fluctuation estimate is iteratively updated using the LSTM model, and the time scale requirement is extracted from the power grid dispatch demand to generate a minute-level power prediction sequence. Step S6: If the minute-level power prediction sequence shows severe radiation jitter, then the influence of cloud movement is refined by random forest regression, the superposition effect of interference factors is determined, and the optimized power output curve is obtained. Step S7: Based on the optimized power output curve, obtain energy utilization feedback data, integrate the energy utilization feedback data into the overall prediction model, and determine the final photovoltaic power generation prediction result.
[0008] Furthermore, step 1 specifically includes: Real-time radiation data and historical meteorological condition records are obtained from the monitoring system of the photovoltaic power station, and wind speed parameters, temperature parameters and cloud cover parameters are extracted from the historical meteorological condition records. The real-time radiation data is aligned point-to-point with the extracted wind speed, temperature, and cloud cover parameters. The radiation jitter amplitude is calculated for each aligned dataset. The radiation jitter amplitude is defined as the absolute value of the difference between the radiation value at the current moment and the radiation value at the previous moment. All the radiation jitter amplitudes are arranged in chronological order to form a jitter amplitude sequence. By statistically analyzing the distribution characteristics of the differences between adjacent values in the jitter amplitude sequence, the initial deviation fluctuation range is determined. If the difference between adjacent values exceeds three times the median of the jitter amplitude sequence, it is marked as an abnormal data point. Remove the radiation jitter amplitudes corresponding to the abnormal data points, reorder the remaining radiation jitter amplitudes to form a new jitter amplitude sequence, and calculate the standard deviation of the new jitter amplitude sequence; By multiplying the standard deviation by a preset coefficient, the upper and lower limits of the initial deviation fluctuation range are obtained, respectively.
[0009] Furthermore, step 2 specifically includes: Meteorological parameters related to cloud movement are extracted from the historical meteorological condition records, and the extracted meteorological parameters are organized into an ordered cloud movement dataset by time series analysis. Based on the ordered cloud movement dataset, the LSTM model is used to model the nonlinear change characteristics and determine the periodic pattern of cloud movement. Based on the periodic pattern of cloud movement, the changes in external environmental parameters related to dynamic interference are analyzed. If the changes in external environmental parameters exceed a preset threshold, they are marked as high interference periods, and the distribution characteristics of high interference periods are obtained. Based on the distribution characteristics of the high-interference period, the deviation fluctuation data of the corresponding period is extracted, the dynamic change trend of the deviation fluctuation range is calculated, and the adjustment benchmark of the deviation fluctuation range is determined. Based on the adjustment benchmark of the deviation fluctuation range, and combined with the distribution characteristics of dynamic interference, the correction value of the adjustment coefficient is calculated to obtain the interference adjustment parameters adapted to the current scenario. Based on the interference adjustment parameters, the nonlinear change sequence is corrected point by point. It is determined whether the corrected sequence meets the preset stationarity condition. If not, the interference adjustment parameters are iteratively updated until a corrected sequence that meets the stationarity condition is obtained. The adjustment parameters corresponding to the corrected sequence are the dynamic interference adjustment coefficients.
[0010] Furthermore, step 3 specifically includes: When the radiation jitter amplitude exceeds the preset range, the radiation jitter sequence at the current moment is extracted. Based on the radiation jitter sequence, meteorological observation records of the same period are extracted from the historical database and arranged in chronological order to form a meteorological condition sequence. The meteorological condition sequence is classified using a random forest classification model to determine the current meteorological condition category, and the power actual value sequence under the same meteorological condition category is extracted from a preset historical dataset. The actual power value sequence is compared with the reference power prediction value for the same period to calculate the deviation sequence corresponding to the meteorological condition category. The deviation sequence is statistically analyzed to determine the average offset direction and amplitude range of the meteorological condition category. Based on the average offset direction and amplitude range, a power prediction correction value adapted to the current meteorological condition category is calculated.
[0011] Furthermore, step 4 specifically includes: The temperature sequence and cumulative equipment aging sequence within the current operating cycle of the photovoltaic power station are obtained. Based on the temperature sequence and cumulative equipment aging sequence, the actual power generation records of the same period are extracted from the historical database and arranged in chronological order to form the actual power generation sequence. The actual power generation sequence is classified using a random forest classification model to determine its current temperature-aging combination category and obtain a classification label. Based on the classification labels, the baseline prediction deviation sequence under the same temperature-aging combination category is extracted from the historical dataset; Statistical analysis is performed on the benchmark prediction deviation sequence to determine the average deviation direction and amplitude range of the temperature-aging combination category. If the average deviation direction is positive and the amplitude range exceeds the preset range, the corresponding temperature weighting coefficient and aging decay coefficient are extracted from the nonlinear change model. Based on the temperature weighting coefficient and aging attenuation coefficient, combined with the average deviation direction and amplitude range, a comprehensive interference compensation vector is calculated.
[0012] Furthermore, step 5 specifically includes: Based on the comprehensive interference compensation vector, and combined with the interference vector, the comprehensive interference compensation vector is decomposed using a preset mapping rule to obtain the preliminary deviation fluctuation range. Based on the initial deviation fluctuation range, the LSTM model is used for fluctuation analysis. Through multiple iterations, the dynamic trend of deviation fluctuation is determined. Based on the dynamic trend of the deviation fluctuation, and combined with the time scale constraints of power grid dispatch, the corresponding minute-level time window is extracted from the power grid dispatch data. Based on the minute-level time window and combined with the preset prediction sequence generation rules, the estimation results of the LSTM model are time-aligned to obtain a preliminary minute-level power prediction sequence. The preliminary minute-level power prediction sequence is compared with the power grid dispatch data to determine whether there is a time deviation. If there is a time deviation, it is determined whether time scale adjustment is needed based on a preset threshold. Based on the time scale adjustment results, the preliminary minute-level power prediction sequence is finally corrected to obtain a power prediction sequence that meets the minute-level requirements of power grid dispatch.
[0013] Furthermore, step 6 specifically includes: Based on the minute-level power prediction sequence, using preset filtering rules, key fluctuation points of radiation jitter are extracted from the time series, and the jitter degree parameters corresponding to the key fluctuation points are obtained. Based on the jitter level parameter and combined with real-time cloud movement data, a preset threshold is used to determine whether there is a significant superposition effect of interference factors. If the superposition effect of interference factors exceeds the preset threshold, the interference time window is separated from the minute-level power prediction sequence to determine the specific interval affected by cloud movement interference. For the specific interval affected by cloud movement, a random forest regression model is used to model and analyze the correlation between cloud movement and radiative jitter, and to obtain preliminary correction values for power output. Based on the initial power output correction value, and combined with the time series continuity constraint, the power output data is smoothed to obtain an intermediate power output curve that is close to the actual operating state. Based on the fluctuation characteristics of the intermediate power output curve and the dynamic changes of interference factors, the power output is finally corrected to determine the optimized power output curve.
[0014] Furthermore, step 7 specifically includes: Energy utilization feedback data is extracted from the optimized power output curve, and the timestamps of the energy utilization feedback data are matched with the time series of the optimized power output curve to obtain the energy utilization feedback values corresponding to each time point. For each time point, the energy utilization feedback value is compared with the power prediction value at that time point to determine whether the energy utilization feedback value exceeds a preset range; if it exceeds the preset range, the original prediction component for that time point is extracted from the overall prediction model. A linear regression method was used to establish a correlation model between the original predicted components and energy utilization feedback values at time points exceeding the preset range, and correction coefficients were obtained. Based on the correction coefficient, the power prediction values at corresponding time points in the overall prediction model are multiplied and adjusted to obtain the adjusted power prediction sequence. Based on the continuity of the adjusted power prediction sequence, the entire sequence is weighted and smoothed to determine the final photovoltaic power generation prediction result.
[0015] The technical effects and advantages of this invention are as follows: This application provides a step-by-step photovoltaic power plant power prediction method. First, it calculates the initial radiation fluctuation amplitude and determines the preliminary deviation fluctuation range based on real-time radiation and historical meteorological data. Then, it uses LSTM to mine cloud movement patterns from historical sequences and generate dynamic interference adjustment coefficients. When the coefficient exceeds a threshold, it introduces random forest to classify meteorological conditions to obtain power prediction correction values. Next, it integrates the current temperature influence and equipment aging parameters into a nonlinear model to form a comprehensive interference compensation vector. Then, it iteratively updates the deviation estimate through LSTM and generates a high-resolution prediction sequence in combination with the minute-level requirements of grid dispatch. When severe radiation fluctuations are detected, random forest regression is further used to refine the cloud superposition effect. Finally, it outputs an optimized power curve and incorporates energy utilization feedback to form a closed-loop correction, thereby significantly improving the minute-level prediction accuracy and stability of photovoltaic power generation under complex meteorological conditions, effectively supporting grid dispatch and efficient energy utilization. Attached Figure Description
[0016] Figure 1 This is a flowchart of the photovoltaic power plant power prediction method using a step-by-step approach according to the present invention; Figure 2 This is a flowchart illustrating the process of determining the initial deviation fluctuation range in an embodiment of the present invention; Figure 3 This is a schematic diagram of the process for determining the integrated interference compensation vector in an embodiment of the present invention. Detailed Implementation
[0017] The present invention will be further described in detail below with reference to the embodiments given in the accompanying drawings.
[0018] like Figures 1-3 As shown, the photovoltaic power plant power prediction method using a step-by-step approach in this embodiment may specifically include the following steps: Step S1 involves acquiring real-time radiation data and historical meteorological records from the photovoltaic power station, calculating the initial radiation fluctuation amplitude by integrating interference factors, and determining the initial deviation fluctuation range. The specific implementation process is as follows: First, real-time radiation data and historical meteorological condition records of the photovoltaic power station are collected, and wind speed, temperature and cloud cover parameters are extracted from the historical meteorological condition records. The real-time radiation data and the extracted wind speed, temperature and cloud cover parameters are then aligned point-to-point in time to ensure the matching of each parameter in the same time dimension.
[0019] Secondly, the radiative jitter amplitude is calculated for each aligned dataset. The formula for calculating the radiative jitter amplitude is as follows:
[0020] in, Let be the amplitude of the radiation jitter at time t. Let be the real-time radiation intensity at time t. Let be the real-time radiation intensity at time t-1.
[0021] Subsequently, all the radiated jitter amplitudes were arranged in chronological order to form a jitter amplitude sequence. By statistically analyzing the distribution characteristics of the differences between adjacent values in the jitter amplitude sequence, the initial deviation fluctuation range is determined; wherein, the difference between adjacent jitter amplitudes is... ,all Forming a difference sequence .
[0022] Calculate the median of the difference sequence. If a certain difference satisfy Then mark the corresponding time. These are abnormal data points; after removing the radiation jitter amplitudes corresponding to the abnormal data points, the remaining radiation jitter amplitudes are reordered to form a new jitter amplitude sequence. .
[0023] Calculate the average value of the new jitter amplitude sequence. and standard deviation The formula for calculating the average value is:
[0024] Standard deviation The calculation formula is:
[0025] in, For the new jitter amplitude sequence The Middle One effective radiated jitter amplitude value, This represents the number of samples for the remaining effective radiative jitter amplitude.
[0026] Finally, by multiplying the standard deviation by a preset coefficient k (the preset coefficient k is calibrated according to the actual operation scenario of the photovoltaic power station, and usually takes a value range of 2~3), the upper limit value U and the lower limit value L of the preliminary deviation fluctuation range are obtained respectively. The calculation formula is as follows:
[0027]
[0028] The initial deviation fluctuation range is output for subsequent photovoltaic power plant power prediction.
[0029] In the actual operation of the photovoltaic power station, the real-time radiation data is collected by radiation sensors deployed on-site at the photovoltaic power station, with a collection frequency of once per minute. The collected data is transmitted to the central monitoring system of the power station via a wireless network. The historical meteorological conditions are recorded from the meteorological department's database and the power station's own meteorological monitoring station, covering parameters such as wind speed, temperature, and cloud cover over the past few years, and are stored in hourly units.
[0030] For example, taking a 100MW desert photovoltaic power station as an example, the historical monitoring data of the power station includes hourly average wind speed, ambient temperature and cloud cover percentage, which are stored in CSV format to support the analysis of seasonal meteorological characteristics.
[0031] Specifically, historical meteorological records were first imported, and valid data columns within a specified time period were filtered using a data processing script to extract time-series data on wind speed, temperature, and cloud cover from January to December 2023. Wind speed was measured in m / s, temperature in °C, and cloud cover was represented as 0–100% coverage. After standardization, a structured dataset was formed, such as wind speed sequences [4.2, 5.1, 3.8], temperature sequences [20.5, 22.3, 19.7], and cloud cover sequences [10, 15, 20], providing meteorological correlation basis for subsequent radiation characteristic analysis.
[0032] Real-time radiation data is aligned point-to-point with wind speed, temperature, and cloud cover data to unify the timestamps of each parameter. For data with inconsistent sampling granularity or missing values, linear interpolation and neighbor mean filling methods are used to complete and align the data, forming a synchronized matching dataset. For example, a single set of aligned data includes: radiation intensity 500W / m², wind speed 4.5m / s, temperature 25℃, and cloud cover 12%, to ensure the accuracy of radiation jitter calculation.
[0033] The radiation jitter amplitude is calculated point by point for the time-aligned dataset. Assuming the radiation sequence is [480, 500, 490] W / m², the radiation jitter amplitudes are 20 and 10 respectively, which can accurately characterize the short-term fluctuation characteristics of solar radiation, and are especially suitable for scenarios with severe radiation disturbances caused by rapid cloud movement.
[0034] All radiation jitter amplitudes are arranged chronologically to form a jitter amplitude sequence. The deviation fluctuation range is determined based on the statistical distribution of the differences between adjacent jitter amplitudes, and outliers are marked using the 3x median method. For the jitter amplitude sequence [20, 10, 50, 15], the adjacent differences are -10, 40, and -35, and the sorted difference sequence is [-35, -10, 40], with a median Med of -10. If the following conditions are met... If the absolute value of the difference is greater than 30, then the corresponding jitter amplitude of 50 is marked as an outlier, which can effectively eliminate noise data introduced by sensor malfunctions, etc.
[0035] After removing outliers, the jitter amplitude sequence is reconstructed, for example, [20, 10, 50, 15] is purified to [20, 10, 15]. The mean and standard deviation of the purified sequence are calculated, and the standard deviation amplification method is used to determine the initial deviation fluctuation range. If the mean of the sequence [20, 10, 15] is... Standard deviation ,Pick The upper limit is 25 and the lower limit is 5. Substituting the obtained preliminary deviation fluctuation range into the power prediction model can suppress the prediction deviation caused by radiation fluctuations, improve the stability of the prediction output, and support the refined dispatching of the power grid.
[0036] This method can significantly improve the reliability of meteorological and radiation data. In complex meteorological areas such as high wind speed and high cloud cover, it can accurately predict short-term power drop trends, optimize the response strategy of energy storage systems, and improve the operational stability and grid acceptance capacity of photovoltaic power plants.
[0037] Step S2: Based on the initial deviation fluctuation range, a long short-term memory model is used to process the nonlinear change sequence, extracting the cloud movement pattern from the historical meteorological condition records to obtain the dynamic interference adjustment coefficient. The specific implementation process is as follows: First, meteorological parameters related to cloud movement are extracted from the historical meteorological condition records. These meteorological parameters include satellite cloud image sequence data, cloud base height data collected by ground cloud height meters, and cloud movement speed data. The extracted meteorological parameters are then processed into a time series to form an ordered cloud movement dataset.
[0038] Secondly, based on the ordered cloud movement dataset, the LSTM model is used to model its nonlinear variation characteristics. The gating mechanism of the LSTM model is used to capture long-term dependencies in the cloud movement time series, thus determining the periodicity of cloud movement. This periodicity is identified using the periodogram method, and its periodogram power calculation formula is as follows:
[0039] in, The power of the periodogram corresponding to frequency f is represented; T represents the length of the historical data time series. f represents the hidden layer output of the LSTM model at time t; f represents the frequency (period). ); This represents a complex exponential function, where j is the imaginary unit ( ); The square operation of the modulus of a complex number; : The phase term of a complex exponential function.
[0040] Based on the periodicity of cloud movement, the changes in external environmental parameters related to dynamic interference are analyzed. These external environmental parameters include the rate of change of the pressure gradient and radar echo intensity. If the changes in these external environmental parameters exceed a preset threshold, they are marked as high-interference periods, and the distribution characteristics of these high-interference periods are obtained. The formula for calculating the rate of change of the pressure gradient is:
[0041] when And radar echo intensity When this period is marked as a high-interference period; in, This represents the rate of change of the pressure gradient at time t; This represents the air pressure value at time t; Indicates the time interval (10 minutes in this case); This represents the safe threshold for the rate of change of the air pressure gradient (obtained based on historical data statistics, typically taken as 0.05 hPa / min). Indicates radar echo intensity; This represents the radar echo intensity threshold (usually taken as 35 dBZ, as the criterion for judging strong convective cloud clusters).
[0042] Based on the distribution characteristics of the high-interference periods, the deviation fluctuation data (i.e., the difference sequence between the actual output power of the photovoltaic power station and the predicted power under theoretical clean sky conditions) of the corresponding periods are extracted, and the dynamic trend of the deviation fluctuation range is calculated. A linear fitting method is used to fit the dynamic trend, and the fitting formula is:
[0043] in Let be the fluctuation intensity at time t. k The fitting slope (characterizing the rate of change of wave intensity) is used for fitting. b The fitting intercept (characterizing the fluctuation intensity at the start of the disturbance) is used for fitting. t The time point within the duration of the interference event; the fluctuation intensity corresponding to the inflection point of the fitted trend line (the moment when the slope changes from positive to negative or from negative to positive) is used as the adjustment benchmark for the deviation fluctuation range. .
[0044] By adjusting the fluctuation range as a benchmark and combining the distribution characteristics of dynamic interference, the correction value of the adjustment coefficient is calculated to obtain the interference adjustment parameters applicable to the current scenario.
[0045] Adjustment benchmark based on the aforementioned deviation fluctuation range Based on the distribution characteristics of dynamic interference (including interference intensity, duration, and direction of arrival), the correction value of the adjustment coefficient is calculated to obtain the interference adjustment parameter K adapted to the current scenario. The calculation formula is as follows:
[0046] in, Basic adjustment factor ( , (The maximum value of the fluctuation intensity during the period of high interference). Time decay correction factor ( , (This is the attenuation coefficient, with an empirical value of 0.02). The azimuth correction factor ( , The angle between the direction of the incoming interference and the normal direction of the power station array. ); This represents the maximum value of the power deviation fluctuation intensity during periods of high interference. The adjustment benchmark for the fluctuation range (unit: kW); This is the ratio of the maximum interference intensity to the adjustment reference. It is an exponentially decaying function; This is the time decay coefficient; t The duration of the interference; The angle between the direction of the incoming interference and the normal direction of the power station array, with a value range of [value missing]. ; Angle The sine function value; this parameter K The aim is to quantify the impact of external disturbances on power plant output fluctuations. K The larger the value, the stronger the interference, and the greater the correction required.
[0047] Based on the interference adjustment parameter K, the nonlinear variation sequence is corrected point by point, and the correction formula is as follows:
[0048] in, The power value after correction at time t. The original power change sequence value at time t. K is the average value of the original power sequence; K(t) represents the disturbance adjustment parameter at time t.
[0049] The ADF unit root test is used to determine whether the corrected sequence meets the preset stationarity condition. If it does not meet the condition, the iterative formula is used instead. Iteratively update the interference adjustment parameters; wherein, The adjustment parameters for the (n+1)th iteration. The adjustment parameters for the nth iteration are... Let L be the learning rate, with an empirical value of 0.01, and L be the loss function. Represents the loss function Adjusting parameters for the nth iteration The partial derivatives are calculated until a correction sequence that satisfies the stationarity condition is obtained. The adjustment parameter corresponding to the correction sequence is the dynamic disturbance adjustment coefficient.
[0050] Step S3: If the dynamic interference adjustment coefficient exceeds a preset threshold, random forest classification is performed on the radiative jitter to determine the current meteorological condition type and obtain the corrected power prediction value after classification. The specific implementation process is as follows: When the radiation jitter amplitude exceeds the preset range, the radiation jitter sequence at the current moment is extracted (30 minutes in length, with a collection interval of one minute); based on the radiation jitter sequence, meteorological observation records of the same period are extracted from the historical database and arranged in chronological order to form a meteorological condition sequence, which covers parameters such as temperature, humidity, wind speed, and cloud cover.
[0051] For example, when a photovoltaic power station detects abnormal fluctuations in solar radiation intensity, and the amplitude of the radiation jitter exceeds a preset threshold of 10%, it is determined that the strong interference is caused by rapid cloud movement or atmospheric turbulence. The system immediately collects the current radiation intensity time-series data through on-site radiation sensors, constructing a 30-minute radiation jitter sequence with a sampling interval of 1 minute, for radiation disturbance characteristic analysis. For the radiation jitter sequence, the system retrieves historical databases and extracts meteorological observation records that match the current time period and the same month. These observation records include parameters such as temperature, humidity, wind speed, and cloud cover, arranged in an orderly manner along the time axis to form a meteorological condition time-series sequence, ensuring coverage of the complete period from one hour before the radiation jitter occurred to the current moment, providing input basis for meteorological type classification.
[0052] A random forest classification model is used to classify the meteorological condition sequence. This model consists of multiple decision trees, constructed by randomly sampling data and features. A majority voting mechanism is used to determine the current meteorological condition category, resulting in a classification label. The meteorological condition categories include typical categories such as sunny, cloudy, overcast, and haze. The real-time meteorological condition sequence is input into the trained random forest classification model, which outputs the current meteorological condition classification label, for example, classifying it as partly cloudy.
[0053] Based on the classification labels, a sequence of actual power values under the same meteorological condition category is extracted from a preset historical dataset; the sequence of actual power values is compared with the predicted baseline power value for the same period to calculate the deviation sequence corresponding to that meteorological condition category. The formula for calculating the deviation value is:
[0054] in, Let be the deviation value at time i. Let be the reference power prediction value at time i, which is calculated based on an ideal, disturbance-free model, and the formula is: S represents the total area of the photovoltaic modules. Where I is the photovoltaic module conversion efficiency, K is the system loss factor (default value is 0.95), and I is the current radiation intensity. Let be the actual power value at time i.
[0055] Perform statistical analysis on the deviation sequence and calculate the average offset value of the deviation sequence. and standard deviation Determine the average offset direction and amplitude range of this meteorological condition category; among which, At that time, the offset direction is downward (the actual power is lower than the predicted power). When the offset direction is upward (actual power is higher than predicted power), the amplitude range is expressed as follows: .
[0056] For example, for the partly cloudy to partly sunny category, the average offset value is obtained statistically. kW, standard deviation kW, meaning the actual power is on average 75kW lower than the predicted value, with a deviation range of 60kW to 90kW, corresponding to a deviation ratio of 5% to 20%.
[0057] Based on the average offset direction and amplitude range, a power prediction correction value adapted to the current meteorological condition category is calculated, and the correction formula is as follows:
[0058] in, This is a power prediction correction value. This represents the average value of the predicted baseline power for the same period. This is the absolute value of the average offset.
[0059] Step S4: Based on the classified power prediction correction value, obtain the current temperature influence parameters and equipment aging parameters, and integrate the temperature influence parameters and equipment aging parameters into the nonlinear variation model to determine the comprehensive interference compensation vector. The specific implementation process is as follows: The temperature sequence (collected at hourly intervals, covering the entire operating cycle) and the cumulative equipment aging sequence (cumulative running time of photovoltaic modules since installation and total hours of high-temperature exposure) within the current operating cycle of the photovoltaic power station are obtained. Based on the temperature sequence and the cumulative equipment aging sequence, the actual power generation records of the same period are extracted from the historical database and arranged in chronological order to form the actual power generation sequence.
[0060] For example, when a photovoltaic power station enters a new operating cycle, the system collects temperature data from the photovoltaic module panels using on-site temperature sensors. Sampling is done at 1-hour intervals, covering the period from 6:00 AM to 6:00 PM daily, constructing a temperature time-series sequence covering the entire operating cycle (e.g., one week). Simultaneously, based on the cumulative operating time of the photovoltaic modules since commissioning and the cumulative exposure time to high-temperature environments, a cumulative aging sequence is calculated. For instance, if a photovoltaic module has been in continuous operation for 5 years, with a total cumulative operating time and high-temperature exposure time of 18,000 hours, the system can characterize the real-time impact of temperature on power generation performance and quantify the long-term degradation effect of the equipment. This data acquisition process ensures the real-time nature, completeness, and accuracy of the temperature and aging sequences, providing a data foundation for subsequent temperature-aging coupled analysis.
[0061] A random forest classification model is used to classify the actual power generation sequence. This model is an ensemble learning model that achieves classification through joint voting by multiple decision trees, which can effectively avoid overfitting and improve generalization ability. The model uses the mean, variance, and trend change of power generation as input features, and divides the combined categories according to temperature range (low temperature, medium temperature, high temperature) and aging range (low aging, medium aging, high aging), and outputs the classification label of the current operating condition.
[0062] Based on the classification labels, the baseline prediction deviation sequence under the same temperature-aging combination category is extracted from the historical dataset; Statistical analysis is performed on the baseline prediction deviation sequence to determine the average deviation direction and amplitude range of the temperature-aging combination category.
[0063] For example, if the mean of the calculated sequence is positive 5kW, then the direction is positive. The amplitude range is obtained by the minimum and maximum values, such as from 2kW to 10kW. This reflects the systematic offset caused by temperature and aging. In business, this can optimize the calibration of the prediction model and improve the operating efficiency of the power plant.
[0064] If the average deviation direction is positive and the amplitude range exceeds the preset range, then the corresponding temperature weighting coefficient is extracted from the nonlinear change model. and aging degradation coefficient The expression for the nonlinear change model is:
[0065] in, To predict power generation using a nonlinear model, For standard operating conditions ( , Rated power generation under ) Temperature weighting coefficient (unit: °C) - ¹, Example value is 0.02). Aging degradation coefficient (unit: h) - ¹, Example value is T represents the current panel temperature, and A represents the cumulative aging value.
[0066] Based on the temperature weighting coefficient Aging attenuation coefficient Based on the average deviation direction and amplitude range, a comprehensive interference compensation vector is calculated. Its expression is:
[0067] in, This is the minimum value of the deviation sequence. This represents the maximum value of the deviation sequence.
[0068] Step S5: Based on the comprehensive interference compensation vector, the LSTM model is used to iteratively update the deviation fluctuation estimate, and the time scale requirement is extracted from the power grid dispatch demand to generate a minute-level power prediction sequence. The specific implementation process is as follows: Based on the comprehensive interference compensation vector, and combined with the interference vector (capturing interference factors such as sudden weather changes and equipment failures), the comprehensive interference compensation vector is decomposed using a preset mapping rule. The decomposition yields a preliminary deviation fluctuation range. The preset mapping rule is a linear transformation or nonlinear mapping framework built based on historical experience, expressed by the following formula: ;in For the interference vector, Preset mapping function: This represents the initial deviation fluctuation range.
[0069] Based on the initial deviation fluctuation range, the LSTM model is used for fluctuation analysis. The model is initialized with a hidden state of zero, and the model weights are adjusted through multiple iterations (the number of iterations is preset to 10) to minimize the mean square error and determine the dynamic trend of deviation fluctuation.
[0070] Based on the dynamic trend of the deviation fluctuation, and combined with the time scale constraint of power grid dispatch (limited to a daily 24-hour analysis window), the corresponding minute-level time window (usually 5-10 minutes) is extracted from the power grid dispatch data; based on the minute-level time window, and combined with the preset prediction sequence generation rule (using linear interpolation method), the estimation results of the LSTM model are time-aligned to obtain a preliminary minute-level power prediction sequence.
[0071] The preliminary minute-level power prediction sequence is compared with the power grid dispatch data to determine whether there is a time deviation. If there is a time deviation, it is determined whether time scale adjustment is needed based on a preset threshold (usually 1 minute). Based on the time scale adjustment result, the preliminary minute-level power prediction sequence is shifted and corrected to obtain a power prediction sequence that meets the minute-level requirements of power grid dispatch.
[0072] Step S6: If the minute-level power prediction sequence shows severe radiation jitter, then random forest regression is used to refine the effect of cloud movement, determine the superposition effect of interfering factors, and obtain the optimized power output curve. The specific implementation process is as follows: Based on the minute-level power prediction sequence, using a preset screening rule (setting the radiation jitter amplitude threshold to 0.05), key fluctuation points of radiation jitter are extracted from the time series, and the jitter degree parameter corresponding to the key fluctuation points is obtained (quantized as an exponent of 0 to 1, where 0 represents stability and 1 represents severe jitter). The calculation formula for the jitter degree parameter is as follows: ;in, for The parameter for the degree of jitter at any given time. for The amplitude of constant radiation jitter. This represents the maximum amplitude of radiation jitter within the observation period.
[0073] Based on the jitter level parameter, combined with real-time cloud movement data (including cloud speed, coverage, and direction of movement), and through comparison with a preset threshold (superposition effect index greater than 0.3), it is determined whether there is a significant superposition effect of interfering factors. The formula for determining the superposition effect of interfering factors is as follows: ,in, for The exponential effect of time-based interference. These are weighting coefficients. For cloud coverage, This represents the speed of cloud movement. If the effects of superimposed interference factors are considered... If a significant interference superposition effect is found, the sliding window method is used to separate the interference time window from the minute-level power prediction sequence to determine the specific interval affected by cloud movement interference.
[0074] For the specific interval affected by cloud movement, a random forest regression model is used to model and analyze the correlation between cloud movement and radiative jitter. The random forest regression model consists of 100 decision trees, and the input features are cloud speed, coverage, and radiative jitter value. The training dataset is derived from historical records of the past month, and the output power output is a preliminary correction value.
[0075] Based on the initial power output correction value, and considering the time series continuity constraint (adjacent point deviation not exceeding 5%), a moving average filter is used to smooth the power output data, resulting in an intermediate power output curve that approximates the actual operating state. Specifically, the intermediate power output curve is obtained using a moving average filter, and the formula is as follows: ;in, for Intermediate power value after time smoothing This is an initial correction value for the output power. The length of the sliding window. The time variable is used for summation and traversal.
[0076] Based on the fluctuation characteristics of the intermediate power output curve (extracting the main frequency component through Fourier transform), and combined with the dynamic changes of interference factors, the power output is finally corrected (weighted fusion of the intermediate curve and interference increment) to determine the optimized power output curve, defined by the formula: ,in, To optimize the power output value, These are weighting coefficients. This is the interference increment compensation term.
[0077] Step S7: Based on the optimized power output curve, obtain energy utilization feedback data, integrate the energy utilization feedback data into the overall prediction model, and determine the final photovoltaic power generation prediction result. The specific implementation process is as follows: Energy utilization feedback data is extracted from the actual operating data corresponding to the optimized power output curve. The energy utilization feedback data comes from the actual power generation value collected by the inverter or electricity meter, with an accurate timestamp.
[0078] The timestamps of the energy utilization feedback data are precisely matched with the time series of the optimized power output curve (using a time alignment algorithm) to obtain the energy utilization feedback values corresponding to each time point.
[0079] For each time point, the energy utilization feedback value is compared with the power prediction value at that time point to determine whether the energy utilization feedback value exceeds a preset range (based on ±5% of the prediction value); if it exceeds the preset range, the original prediction component (including the basic irradiance prediction component, temperature influence component, and equipment attenuation component) for that time point is extracted from the overall prediction model.
[0080] A linear regression method was used to establish a correlation model between the original predicted components and the energy utilization feedback values at time points exceeding the preset range. The original predicted components were used as independent variables X, and the energy utilization feedback values were used as dependent variables Y to construct a linear equation. , where the slope a is the correction coefficient.
[0081] Based on the correction coefficient, the power prediction values at corresponding time points in the overall prediction model are multiplied and adjusted to obtain the adjusted power prediction sequence. Based on the continuity of the adjusted power prediction sequence, the entire sequence is weighted and smoothed using an exponentially weighted moving average method to ensure that the prediction curve conforms to the physical characteristics of natural changes in photovoltaic power, and finally the final photovoltaic power generation prediction result is determined.
[0082] This invention addresses the unique operational scenario of photovoltaic power plants, which suffers from large power prediction deviations and insufficient minute-level accuracy due to multiple nonlinear interference factors such as severe radiation fluctuations caused by rapid cloud movement, superimposed temperature changes, and equipment aging. It solves the problem through a step-by-step, progressive integration approach, thereby significantly improving the minute-level prediction accuracy and stability of photovoltaic power generation under complex weather conditions, effectively supporting grid dispatch and efficient energy utilization.
[0083] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.
Claims
1. A method for power prediction of a photovoltaic power plant in a stepwise manner, characterized in that, The method includes the following steps: Step S1: Obtain real-time radiation data and historical meteorological condition records of the photovoltaic power station, integrate interference factors to calculate the initial radiation fluctuation amplitude, and determine the initial deviation fluctuation range. Step S2: Based on the preliminary deviation fluctuation range, a long short-term memory model is used to process the nonlinear change sequence, extracting the cloud movement pattern from the historical meteorological condition records to obtain the dynamic interference adjustment coefficient; this step specifically includes: Meteorological parameters related to cloud movement are extracted from the historical meteorological condition records, and the extracted meteorological parameters are organized into an ordered cloud movement dataset by time series analysis. Based on the ordered cloud movement dataset, the LSTM model is used to model the nonlinear change characteristics and determine the periodic pattern of cloud movement. Based on the periodic pattern of cloud movement, the changes in external environmental parameters related to dynamic interference are analyzed. If the changes in external environmental parameters exceed a preset threshold, they are marked as high interference periods, and the distribution characteristics of high interference periods are obtained. Based on the distribution characteristics of the high-interference period, the deviation fluctuation data of the corresponding period is extracted, the dynamic change trend of the deviation fluctuation range is calculated, and the adjustment benchmark of the deviation fluctuation range is determined. Based on the adjustment benchmark of the deviation fluctuation range, and combined with the distribution characteristics of dynamic interference, the correction value of the adjustment coefficient is calculated to obtain the interference adjustment parameters adapted to the current scenario. Based on the interference adjustment parameters, the nonlinear change sequence is corrected point by point. It is determined whether the corrected sequence meets the preset stationarity condition. If not, the interference adjustment parameters are iteratively updated until a corrected sequence that meets the stationarity condition is obtained. The adjustment parameters corresponding to the corrected sequence are the dynamic interference adjustment coefficients. Step S3: If the dynamic interference adjustment coefficient exceeds a preset threshold, perform random forest classification on the radiative jitter to determine the current meteorological condition type and obtain the classified power prediction correction value; this step specifically includes: When the radiation jitter amplitude exceeds the preset range, the radiation jitter sequence at the current moment is extracted. Based on the radiation jitter sequence, meteorological observation records of the same period are extracted from the historical database and arranged in chronological order to form a meteorological condition sequence. The meteorological condition sequence is classified using a random forest classification model to determine the current meteorological condition category, and the power actual value sequence under the same meteorological condition category is extracted from a preset historical dataset. The actual power value sequence is compared with the reference power prediction value for the same period to calculate the deviation sequence corresponding to the meteorological condition category. The deviation sequence is statistically analyzed to determine the average offset direction and amplitude range of the meteorological condition category. Based on the average offset direction and amplitude range, a power prediction correction value adapted to the current meteorological condition category is calculated; Step S4: Based on the classified power prediction correction value, obtain the current temperature influence parameters and equipment aging parameters, and integrate the temperature influence parameters and equipment aging parameters into the nonlinear change model to determine the comprehensive interference compensation vector; this step specifically includes: The temperature sequence and cumulative equipment aging sequence within the current operating cycle of the photovoltaic power station are obtained. Based on the temperature sequence and cumulative equipment aging sequence, the actual power generation records of the same period are extracted from the historical database and arranged in chronological order to form the actual power generation sequence. The actual power generation sequence is classified using a random forest classification model to determine its current temperature-aging combination category and obtain a classification label. Based on the classification labels, the baseline prediction deviation sequence under the same temperature-aging combination category is extracted from the historical dataset; Statistical analysis is performed on the benchmark prediction deviation sequence to determine the average deviation direction and amplitude range of the temperature-aging combination category. If the average deviation direction is positive and the amplitude range exceeds the preset range, the corresponding temperature weighting coefficient and aging decay coefficient are extracted from the nonlinear change model. Based on the temperature weighting coefficient and aging attenuation coefficient, combined with the average deviation direction and amplitude range, a comprehensive interference compensation vector is calculated. Step S5: Based on the comprehensive interference compensation vector, the deviation fluctuation estimate is iteratively updated using an LSTM model. Time scale requirements are extracted from the power grid dispatch demand to generate a minute-level power prediction sequence. This step specifically includes: Based on the comprehensive interference compensation vector, and combined with the interference vector, the comprehensive interference compensation vector is decomposed using a preset mapping rule to obtain the preliminary deviation fluctuation range. Based on the initial deviation fluctuation range, the LSTM model is used for fluctuation analysis. Through multiple iterations, the dynamic trend of deviation fluctuation is determined. Based on the dynamic trend of the deviation fluctuation, and combined with the time scale constraints of power grid dispatch, the corresponding minute-level time window is extracted from the power grid dispatch data. Based on the minute-level time window and combined with the preset prediction sequence generation rules, the estimation results of the LSTM model are time-aligned to obtain a preliminary minute-level power prediction sequence. The preliminary minute-level power prediction sequence is compared with the power grid dispatch data to determine whether there is a time deviation. If there is a time deviation, it is determined whether time scale adjustment is needed based on a preset threshold. Based on the time scale adjustment results, the preliminary minute-level power prediction sequence is finally corrected to obtain a power prediction sequence that meets the minute-level requirements of power grid dispatch. Step S6: If the minute-level power prediction sequence shows severe radiation jitter, then the influence of cloud movement is refined through random forest regression to determine the superposition effect of interfering factors, and the optimized power output curve is obtained; this step specifically includes: Based on the minute-level power prediction sequence, using preset filtering rules, key fluctuation points of radiation jitter are extracted from the time series, and the jitter degree parameters corresponding to the key fluctuation points are obtained. Based on the jitter level parameter and combined with real-time cloud movement data, a preset threshold is used to determine whether there is a significant superposition effect of interference factors. If the superposition effect of interference factors exceeds the preset threshold, the interference time window is separated from the minute-level power prediction sequence to determine the specific interval affected by cloud movement interference. For the specific interval affected by cloud movement, a random forest regression model is used to model and analyze the correlation between cloud movement and radiative jitter, and to obtain preliminary correction values for power output. Based on the initial power output correction value, and combined with the time series continuity constraint, the power output data is smoothed to obtain an intermediate power output curve that is close to the actual operating state. Based on the fluctuation characteristics of the intermediate power output curve and combined with the dynamic changes of interference factors, the power output is finally corrected to determine the optimized power output curve. Step S7: Based on the optimized power output curve, obtain energy utilization feedback data, integrate the energy utilization feedback data into the overall prediction model, and determine the final photovoltaic power generation prediction result.
2. A method for power prediction of a photovoltaic power plant in a stepwise manner according to claim 1, characterized in that, Step 1 specifically includes: Real-time radiation data and historical meteorological condition records are obtained from the monitoring system of the photovoltaic power station, and wind speed parameters, temperature parameters and cloud cover parameters are extracted from the historical meteorological condition records. The real-time radiation data is aligned point-to-point with the extracted wind speed, temperature, and cloud cover parameters. The radiation jitter amplitude is calculated for each aligned dataset. The radiation jitter amplitude is defined as the absolute value of the difference between the radiation value at the current moment and the radiation value at the previous moment. All the radiation jitter amplitudes are arranged in chronological order to form a jitter amplitude sequence. By statistically analyzing the distribution characteristics of the differences between adjacent values in the jitter amplitude sequence, the initial deviation fluctuation range is determined. If the difference between adjacent values exceeds three times the median of the jitter amplitude sequence, it is marked as an abnormal data point. Remove the radiation jitter amplitudes corresponding to the abnormal data points, reorder the remaining radiation jitter amplitudes to form a new jitter amplitude sequence, and calculate the standard deviation of the new jitter amplitude sequence; By multiplying the standard deviation by a preset coefficient, the upper and lower limits of the initial deviation fluctuation range are obtained, respectively.
3. The method for power prediction of a photovoltaic power plant in a stepwise manner according to claim 1, characterized in that, Step 7 specifically includes: Energy utilization feedback data is extracted from the optimized power output curve, and the timestamps of the energy utilization feedback data are matched with the time series of the optimized power output curve to obtain the energy utilization feedback values corresponding to each time point. For each time point, the energy utilization feedback value is compared with the power prediction value at that time point to determine whether the energy utilization feedback value exceeds a preset range; if it exceeds the preset range, the original prediction component for that time point is extracted from the overall prediction model. A linear regression method was used to establish a correlation model between the original predicted components and energy utilization feedback values at time points exceeding the preset range, and correction coefficients were obtained. Based on the correction coefficient, the power prediction values at corresponding time points in the overall prediction model are multiplied and adjusted to obtain the adjusted power prediction sequence. Based on the continuity of the adjusted power prediction sequence, the entire sequence is weighted and smoothed to determine the final photovoltaic power generation prediction result.