Photovoltaic power station generating capacity prediction method

Abstract

The invention provides a photovoltaic power station generating capacity prediction method, and belongs to the technical field of photovoltaic power stations, and the method comprises the steps: carrying out the feature extraction of a photovoltaic module image, building a health state feature vector, inputting a multi-scale meteorological feature into a multi-scale meteorological fusion prediction model, and outputting a solar irradiance prediction sequence; and inputting the component health state feature vector into a component attenuation dynamic tracking model to output a power attenuation coefficient, calculating an initial power generation prediction value according to irradiance prediction and the attenuation coefficient, carrying out weighted fusion, and carrying out Kalman filtering smoothing processing and Box-Cox conversion equalization processing on a prediction deviation sequence to obtain a final prediction result. The technical problem that the prediction precision is insufficient due to the fact that photovoltaic power generation capacity prediction cannot give consideration to multi-time-scale meteorological changes and long-term attenuation characteristics of assemblies at the same time is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of photovoltaic power stations, and in particular relates to a photovoltaic power station power generation capacity prediction method. BACKGROUND

[0002] Photovoltaic power station power generation capacity prediction is a key technology for new energy grid-connected dispatching. Traditional prediction methods mainly rely on numerical weather prediction models and statistical learning models to establish prediction models through the mapping relationship between historical meteorological data and power generation capacity data. In the current operation and management of photovoltaic power stations, due to the multi-scale characteristics of meteorological conditions from minute-level cloud blockage to multi-day climate change, existing methods often use single time scale feature extraction methods, which are difficult to effectively capture the coupling effects of meteorological changes at different time scales on power generation capacity. At the same time, photovoltaic components will produce power attenuation due to physical degradation mechanisms such as micro-crack expansion, hot spot effect, and cell failure during long-term operation. Existing prediction models mostly use linear attenuation assumptions or simple attenuation rate corrections, ignoring the dynamic relevance of component health status and attenuation process and the random fluctuation characteristics in the attenuation process. That is, there is a technical problem in the prior art that photovoltaic power generation capacity prediction is difficult to simultaneously consider multi-time scale meteorological changes and long-term component attenuation characteristics, resulting in insufficient prediction accuracy. SUMMARY

[0003] Therefore, the application provides a photovoltaic power station power generation capacity prediction method, which can solve the technical problem in the prior art that photovoltaic power generation capacity prediction is difficult to simultaneously consider multi-time scale meteorological changes and long-term component attenuation characteristics, resulting in insufficient prediction accuracy.

[0004] The application is implemented in the following manner: the application provides a photovoltaic power station power generation amount prediction method, which comprises the following steps: collecting satellite cloud image sequence data of a region where a photovoltaic power station is located, ground meteorological station real-time observation data, numerical weather prediction model output data, historical power generation amount data and component operation life data; performing multi-time scale decomposition processing on the satellite cloud image sequence data, the ground meteorological station real-time observation data and the numerical weather prediction model output data to obtain short-time scale cloud layer change characteristics, medium-time scale meteorological fluctuation characteristics and long-time scale climate trend characteristics respectively; collecting electroluminescence detection images and infrared thermal imaging images of each component array in the photovoltaic power station; performing feature extraction processing on the electroluminescence detection images and the infrared thermal imaging images to establish a component health state feature vector; inputting the short-time scale cloud layer change characteristics, the medium-time scale meteorological fluctuation characteristics, the long-time scale climate trend characteristics and historical prediction error standard deviation into a multi-scale meteorological fusion prediction model to output a solar irradiance prediction sequence; inputting the component health state feature vector, power attenuation records in the historical power generation amount data, the component operation life data and temperature distribution uniformity into a component attenuation dynamic tracking model to output a power attenuation coefficient and a power attenuation trend curve; calculating a preliminary power generation amount prediction value according to the solar irradiance prediction sequence and the power attenuation coefficient and performing weighted fusion to obtain a fused power generation amount prediction value; and performing smoothing processing and equalization processing on a prediction deviation sequence to obtain a final power generation amount prediction value.

[0005] The ground meteorological station real-time observation data comprises solar irradiance, ambient temperature, wind speed and humidity, the numerical weather prediction model output data comprises meteorological element prediction values for the next 72 hours, and the historical power generation amount data comprises daily power generation amount records and corresponding meteorological condition records for the past 90 days.

[0006] The multi-time scale decomposition processing specifically comprises the following steps: decomposing the satellite cloud image sequence data into short-time scale cloud layer change characteristics, the time window length of which ranges from 5 minutes to 30 minutes; decomposing the ground meteorological station real-time observation data into medium-time scale meteorological fluctuation characteristics, the time window length of which ranges from 1 hour to 6 hours; and decomposing the numerical weather prediction model output data into long-time scale climate trend characteristics, the time window length of which ranges from 12 hours to 72 hours.

[0007] The short-time scale cloud layer change characteristics comprise cloud layer moving speed vectors, cloud layer coverage rate change gradients and cloud layer optical thickness fluctuation amplitudes, the medium-time scale meteorological fluctuation characteristics comprise hourly change rates of solar irradiance, temperature daily range and wind speed fluctuation spectrum characteristics, and the long-time scale climate trend characteristics comprise multi-day average solar irradiance trends, atmospheric pressure system evolution paths and seasonal periodic components.

[0008] The step of performing feature extraction processing on the electroluminescent detection image and the infrared thermal imaging image is specifically extracting micro-crack distribution density, hidden failure area area proportion and broken grid number of the battery piece from the electroluminescent detection image, and extracting hot spot temperature difference, temperature distribution uniformity and local overheating point number from the infrared thermal imaging image, and establishing a component health state feature vector.

[0009] The component health state feature vector is a six-dimensional vector containing micro-crack distribution density, hidden failure area area proportion, broken grid number of the battery piece, hot spot temperature difference, temperature distribution uniformity and local overheating point number, and the historical prediction error standard deviation is a standard deviation calculated according to the difference between the actual power generation and the historical predicted power generation in the historical power generation data.

[0010] The structure of the multi-scale meteorological fusion prediction model is that the input layer receives short-time scale cloud change features, medium-time scale meteorological fluctuation features, long-time scale climate trend features and historical prediction error standard deviations, three parallel bidirectional long short-term memory network branches are used to process the short-time scale cloud change features, the medium-time scale meteorological fluctuation features and the long-time scale climate trend features respectively, a multi-head self-attention module is connected after each bidirectional long short-term memory network branch, the outputs of the three bidirectional long short-term memory network branches are weighted and fused through a cross-scale attention fusion layer, the cross-scale attention fusion layer uses a gating mechanism to dynamically adjust the contribution weights of each time scale, and the fused feature vector is output through a fully connected network and a conditional random field layer. A solar irradiance prediction sequence.

[0011] The structure of the component attenuation dynamic tracking model is that the input layer receives the component health state feature vector, the power attenuation record in the historical power generation data, the component operation age data and the temperature distribution uniformity, a convolutional neural network is used to extract the spatial pattern of the component health state feature vector, a recurrent neural network is used to extract the time evolution law of the power attenuation record in the historical power generation data, the features of the convolutional neural network and the recurrent neural network are spliced in a feature fusion layer, the spliced feature vector is input into a noise injection module based on a stochastic differential equation, and the feature vector is output through a fully connected network after being processed by the noise injection module based on the stochastic differential equation. Power attenuation coefficient and power attenuation trend curve.

[0012] The noise injection module based on the stochastic differential equation models the feature transformation process as a stochastic dynamics system, simulates the random fluctuation component in the component attenuation process by introducing a Brownian motion term, and determines the strength coefficient of the Brownian motion term according to the product of the component operation age data and the temperature distribution uniformity.

[0013] The step of calculating the preliminary power generation forecast based on the solar irradiance prediction sequence and power attenuation coefficient and then performing weighted fusion specifically involves weighting and fusing the preliminary power generation forecast with the prediction confidence at each time scale to obtain the fused power generation forecast. The prediction confidence is obtained by normalizing the inverse of the standard deviation of the historical prediction error.

[0014] The step of smoothing the prediction deviation sequence involves calculating the prediction deviation sequence between the fused power generation prediction value and the historical power generation data. When the moving standard deviation of the prediction deviation sequence is greater than 15%, the prediction deviation sequence is smoothed by Kalman filtering to obtain the smoothed power generation prediction value, and the fusion weight coefficients of the features of each time scale in the multi-scale meteorological fusion prediction model are adjusted.

[0015] The Kalman filtering smoothing process describes the dynamic evolution of the prediction bias sequence through a state-space model. It recursively estimates the true state of the prediction bias using two steps: observation update and time update. The state update equation of the state-space model contains a system noise term, and the observation equation of the state-space model contains a measurement noise term. The Kalman gain is determined by minimizing the mean square error of the state estimation.

[0016] The step of equalizing the predicted deviation sequence involves statistically smoothing the distribution skewness coefficient of the predicted power generation value in different time periods. When the absolute value of the distribution skewness coefficient is greater than 0.8, the smoothed predicted power generation value is equalized by Box-Cox transformation to obtain the equalized predicted power generation value.

[0017] The Box-Cox transform equalization process converts non-normally distributed data into an approximately normal distribution using a family of power transforms. The transformation parameters are determined by maximum likelihood estimation, so that the log-likelihood function of the transformed data reaches its maximum value.

[0018] The process includes collecting satellite cloud image sequence data, real-time observation data from ground meteorological stations, numerical weather prediction model output data, historical power generation data, and component operating age data for the area where the photovoltaic power station is located. It also includes detecting the signal-to-noise ratio (SNR) of the real-time observation data from ground meteorological stations. When the SNR is less than 20 dB, wavelet transform multi-scale noise reduction processing is performed on the real-time observation data from ground meteorological stations. The real-time observation data from ground meteorological stations is then updated and returned for reprocessing with multi-timescale decomposition.

[0019] Among them, the wavelet transform multi-scale noise reduction process decomposes the real-time observation data of the ground meteorological station into wavelet coefficients of different frequency scales, uses the soft threshold shrinkage method to suppress noise components for high-frequency wavelet coefficients, keeps the low-frequency wavelet coefficients unchanged to preserve the main components of the signal, and reconstructs the noise-reduced real-time observation data of the ground meteorological station through inverse wavelet transform.

[0020] This invention constructs a two-layer prediction architecture—a multi-scale meteorological fusion prediction model and a component degradation dynamic tracking model—to achieve collaborative modeling of meteorological changes and component degradation processes across multiple time scales. This addresses the technical deficiency in existing technologies where prediction models cannot simultaneously handle short-term meteorological abrupt changes and long-term component degradation. In the multi-scale meteorological fusion prediction model, three parallel bidirectional long short-term memory network branches are used to handle short-term cloud changes, medium-term meteorological fluctuations, and long-term climate trend characteristics, respectively. Cross-scale attention fusion layers and conditional random field layers establish correlation constraints between different time scales. In the component degradation dynamic tracking model, a noise injection module based on stochastic differential equations is introduced to simulate the deterministic drift and random fluctuation components of the component degradation process. Component health status features extracted using electroluminescence detection and infrared thermal imaging are used to dynamically track the power degradation coefficient, and the fusion weights for each time scale are adaptively adjusted using the historical prediction error standard deviation. In summary, this invention, through collaborative modeling of multi-time-scale features and dynamic tracking of component health status, solves the technical problem mentioned in the background technology of insufficient prediction accuracy due to the difficulty in simultaneously considering multi-time-scale meteorological changes and long-term component degradation characteristics in photovoltaic power generation prediction. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a sequence diagram showing the predicted solar irradiance for the next 72 hours in the example.

[0023] Figure 3 This is a comparison chart of the Kalman filter smoothing processing of the prediction bias sequence in the examples. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0025] like Figure 1 The diagram shown is a flowchart of a photovoltaic power plant power generation prediction method provided by the present invention. This method includes the following steps:

[0026] S01. Collect satellite cloud image sequence data, real-time observation data from ground meteorological stations, numerical weather prediction model output data, historical power generation data, and component operating age data for the area where the photovoltaic power station is located. The time resolution of the satellite cloud image sequence data is 5 minutes. The real-time observation data from ground meteorological stations includes solar irradiance, ambient temperature, wind speed, and humidity. The numerical weather prediction model output data includes forecast values ​​of meteorological elements for the next 72 hours. The historical power generation data includes daily power generation records and corresponding meteorological condition records for the past 90 days.

[0027] S02. Perform multi-timescale decomposition processing on the collected satellite cloud image sequence data, real-time observation data from ground meteorological stations, and numerical weather prediction model output data. Decompose the satellite cloud image sequence data into short-timescale cloud layer change characteristics, decompose the real-time observation data from ground meteorological stations into medium-timescale meteorological fluctuation characteristics, and decompose the numerical weather prediction model output data into long-timescale climate trend characteristics. The time window length range corresponding to the short-timescale cloud layer change characteristics is [5, 30] min, the time window length range corresponding to the medium-timescale meteorological fluctuation characteristics is [1, 6] h, and the time window length range corresponding to the long-timescale climate trend characteristics is [12, 72] h.

[0028] S03. Collect electroluminescence detection images and infrared thermal imaging images of each component array in the photovoltaic power station, wherein the acquisition wavelength range of the electroluminescence detection image is [1100, 1200] nm, the acquisition temperature resolution of the infrared thermal imaging image is 0.1℃, and the acquisition cycle is a comprehensive inspection once every quarter.

[0029] S04. Perform feature extraction processing on the collected electroluminescence detection images and infrared thermal imaging images. Extract the microcrack distribution density, the area ratio of the latent failure region and the number of broken grids of the battery cell from the electroluminescence detection images. Extract the hot spot temperature difference, temperature distribution uniformity and the number of local overheating points from the infrared thermal imaging images to establish a component health status feature vector.

[0030] S05. Calculate the historical prediction error standard deviation based on the historical power generation data, input the short-term cloud change characteristics, the medium-term meteorological fluctuation characteristics, the long-term climate trend characteristics and the historical prediction error standard deviation into the multi-scale meteorological fusion prediction model, and output the solar irradiance prediction sequence at different time scales within the next 72 hours, wherein the attention weight coefficient of the multi-scale meteorological fusion prediction model is determined based on the historical prediction error standard deviation.

[0031] S06. Input the component health status feature vector, the power attenuation record in the historical power generation data, the component operating years data and the temperature distribution uniformity into the component attenuation dynamic tracking model, and output the power attenuation coefficient of each component array and the power attenuation trend curve for the next 30 days, wherein the noise intensity parameter of the component attenuation dynamic tracking model is determined according to the product of the component operating years data and the temperature distribution uniformity.

[0032] S07. Calculate the preliminary power generation prediction value based on the solar irradiance prediction sequence and the power attenuation coefficient, and weight and fuse the preliminary power generation prediction value with the prediction confidence level at each time scale to obtain the fused power generation prediction value, wherein the prediction confidence level is obtained by normalizing the inverse of the standard deviation of the historical prediction error.

[0033] S08. Calculate the prediction deviation sequence between the fused power generation prediction value and the historical power generation data. When the sliding standard deviation of the prediction deviation sequence is >15%, perform Kalman filtering on the prediction deviation sequence to obtain a smoothed power generation prediction value, and adjust the fusion weight coefficients of each time scale feature in the multi-scale meteorological fusion prediction model.

[0034] S09. Calculate the distribution skewness coefficient of the smoothed power generation prediction value in different time periods. When the absolute value of the distribution skewness coefficient is >0.8, perform Box-Cox transformation on the smoothed power generation prediction value to obtain the balanced power generation prediction value and output it.

[0035] S10, further, also includes: detecting the signal-to-noise ratio level of the real-time observation data of the ground meteorological station; when the signal-to-noise ratio is <20dB, performing wavelet transform multi-scale noise reduction processing on the real-time observation data of the ground meteorological station, updating the real-time observation data of the ground meteorological station, and returning to step S02 to perform multi-time-scale decomposition processing again.

[0036] The satellite cloud image sequence data is a sequence of cloud distribution images obtained through the visible light and infrared channels of meteorological satellites. Each image in the satellite cloud image sequence data covers an area of ​​100km × 100km around the photovoltaic power station.

[0037] The power decay record in the historical power generation data is the monthly power decay rate data for the past 12 months. The power decay rate is the rate of change of the ratio of the actual power generation in the current month to the nominal power of the component relative to the previous month.

[0038] The short-term cloud change characteristics include cloud movement speed vector, cloud coverage change gradient, and cloud optical thickness fluctuation amplitude; the medium-term meteorological fluctuation characteristics include hourly variation rate of solar irradiance, diurnal temperature range, and wind speed fluctuation spectrum characteristics; and the long-term climate trend characteristics include multi-day average solar irradiance trend, pressure system evolution path, and seasonal periodic components.

[0039] Wherein, the microcrack distribution density is the total crack length detected per unit area, with units of The percentage of the area of ​​the latent failure region is the ratio of the dark area to the total area of ​​the component in the electroluminescence detection image, and the number of broken grids in the battery cell is the number of main grid lines or fine grid lines that are detected to be broken.

[0040] The hot spot temperature difference is the difference between the hot spot area temperature and the average temperature of the module, in °C. The temperature distribution uniformity is characterized by the reciprocal of the standard deviation of the module surface temperature. The number of local hot spots is the number of areas whose temperature exceeds the average temperature of the module by more than 10 °C.

[0041] The component health status feature vector is a six-dimensional vector that includes the microcrack distribution density, the area ratio of the latent failure region, the number of cell grid breaks, the hot spot temperature difference, the temperature distribution uniformity, and the number of local overheating points.

[0042] The historical prediction error standard deviation is calculated based on the difference between the actual power generation and the historical predicted power generation in the historical power generation data. The historical predicted power generation is the predicted power generation value obtained using the same prediction method in the past 90 days.

[0043] The multi-scale meteorological fusion prediction model is structured as follows: the input layer receives short-term cloud change features, medium-term meteorological fluctuation features, long-term climate trend features, and historical prediction error standard deviations. Three parallel bidirectional long short-term memory (BSSM) network branches process these features respectively. Each BSSM branch is followed by a multi-head self-attention module to capture temporal dependencies within the scale. The outputs of the three BSSM branches are weighted and fused through a cross-scale attention fusion layer. This layer uses a gating mechanism to dynamically adjust the contribution weights of each time scale. The weights of the gating mechanism are calculated based on the historical prediction error standard deviations. The fused feature vector is then processed through two layers. The fully connected network and the conditional random field layer output the solar irradiance prediction sequence. The conditional random field layer is used to model the transition constraints between prediction times. The number of heads in the multi-head self-attention module is set to 8, and the activation function of the gating mechanism is the sigmoid function. The steps for establishing the training dataset of the multi-scale meteorological fusion prediction model include collecting historical meteorological data and corresponding actual solar irradiance measurements from photovoltaic power plants over the past three years; dividing the data into training and validation sets in a 7:3 ratio; performing multi-timescale decomposition on the meteorological features in the training set; constructing paired samples of the input feature sequence and the target solar irradiance sequence; standardizing the input feature sequence to a mean of 0 and a standard deviation of 1; and performing min-max normalization on the target solar irradiance sequence to map the numerical range to... The training steps of the multi-scale meteorological fusion prediction model include updating model parameters using an adaptive moment estimation optimization algorithm, setting the initial learning rate to 0.001 and dynamically adjusting it using a cosine annealing strategy, setting the batch size to 64, setting the number of training rounds to 200, and using a weighted combination of mean squared error and negative log-likelihood of conditional random field as the loss function, with weight coefficients set to 0.7 and 0.3. Adversarial sample perturbations are introduced during training to enhance model robustness. These adversarial samples are generated by adding small perturbations in the gradient direction to the input feature sequence, with the perturbation amplitude controlled within 5% of the standard deviation of the input feature sequence. Adversarial training is performed every 10 batches, and the training process is monitored using the mean absolute error on the validation set, with an early stopping mechanism used to prevent overfitting. The multi-scale meteorological fusion prediction model employs an attention-based sequence labeling optimization algorithm. This algorithm calculates the correlation strength between each time step and other time steps in the sequence using a self-attention module, calculates the attention score using a scaled dot product attention formula, and captures dependency patterns in different semantic spaces in parallel using a multi-head mechanism. The conditional random field layer applies global constraints to the solar irradiance prediction sequence to ensure the continuity of predicted values ​​at adjacent time steps meets the requirements of physical laws. The conditional random field layer solves for the optimal labeling path using the Viterbi algorithm. Adversarial training forces the model to learn more robust feature representations by superimposing adversarial perturbations calculated based on the gradient of the loss function onto each training sample. The technical effect is a significant improvement in the accuracy and stability of multi-timescale meteorological predictions, reducing the root mean square error of short-term abrupt change predictions by 28% and the average absolute percentage error of long-term trend predictions by 35%.

[0044] The structure of the component degradation dynamic tracking model is as follows: the input layer receives the component health status feature vector, power degradation records from the historical power generation data, component operating years data, and temperature distribution uniformity. A convolutional neural network (CNN) extracts the spatial pattern of the component health status feature vector. The CNN has a kernel size of 3×3 and a kernel count of 32. A recurrent neural network (RNN) extracts the temporal evolution pattern of the power degradation records from the historical power generation data. The features from the CNN and RNN are concatenated in a feature fusion layer. The concatenated feature vector is then input to a noise injection module based on stochastic differential equations. This noise injection module models the feature transformation process as a stochastic dynamic system, simulating the random fluctuation components during component degradation by introducing Brownian motion terms. The intensity coefficient of the Brownian motion term is determined by the product of the component's operating years data and the temperature distribution uniformity. After processing by the noise injection module based on stochastic differential equations, the feature vector is output through a three-layer fully connected network, showing the power attenuation coefficient and the power attenuation trend curve. The steps for establishing the training dataset for the component attenuation dynamic tracking model include collecting health monitoring data and corresponding actual power attenuation records of photovoltaic modules with different operating years; selecting at least 500 sample modules containing complete attenuation cycles; extracting the component health status feature vector for each sample module; recording the monthly power attenuation rate for the past 12 months; constructing training sample pairs between the component health status feature vector and the monthly power attenuation rate sequence; and normalizing the component health status feature vector to unify the numerical range of each dimension. The monthly power decay rate sequence is differentially processed to extract decay acceleration features. The training steps of the component decay dynamic tracking model include updating network parameters using a stochastic gradient descent optimization algorithm, with a learning rate of 0.01, a momentum coefficient of 0.9, a batch size of 32, and 300 training rounds. The loss function is the mean square error between the predicted decay rate and the actual decay rate. During training, the parameter distribution of the noise injection module based on stochastic differential equations is optimized by stochastic variational inference. The stochastic variational inference approximates the posterior distribution by minimizing the variational lower bound. The variational distribution adopts a Gaussian distribution with diagonal covariance. Gradient backpropagation is achieved by reparameterization during inference. The intensity coefficient of the Brownian motion term is sampled and the expected loss is calculated for each training batch. The gradient is estimated and the variational distribution parameters are updated using the Monte Carlo method. The component attenuation dynamic tracking model uses a noise injection learning framework based on stochastic differential equations to model the forward propagation of the network as a continuous-time stochastic process. The evolution of the feature vector follows a stochastic differential equation containing deterministic drift terms and stochastic diffusion terms. The deterministic drift term, provided by the deterministic transformation of the network, represents the average trend of attenuation. The stochastic diffusion term, through the Brownian motion term, introduces structured noise to represent the random fluctuations of the attenuation process. The Euler-Maruyama method is used to numerically solve the stochastic differential equation, discretizing it into the forward propagation steps of the network. The contribution of the stochastic diffusion term to the gradient of the loss function is derived using Ito's lemma, and the diffusion coefficient is updated. The steady-state properties of the feature vector distribution are obtained by analyzing the corresponding Fock-Planck equation, which guides the setting of the intensity coefficient of the Brownian motion term. The stochastic variational inference approximates the true parameter distribution by introducing a variational posterior distribution and optimizes the variational parameters by maximizing the lower bound of evidence. The technical effect is to accurately capture the nonlinear attenuation characteristics and stochastic fluctuations of photovoltaic modules, reduce the root mean square error of attenuation prediction by 42%, and improve the detection sensitivity of sudden performance degradation by 55%.

[0045] The Kalman filtering smoothing process involves describing the dynamic evolution of the prediction deviation sequence using a state-space model, recursively estimating the true state of the prediction deviation using two steps: observation update and time update. The state update equation of the state-space model includes a system noise term, and the observation equation of the state-space model includes a measurement noise term. The Kalman gain is determined by minimizing the mean square error of the state estimation, and the Kalman gain is calculated based on the system noise covariance and the measurement noise covariance.

[0046] The Box-Cox transformation equalization process converts non-normally distributed data into an approximately normal distribution using a family of power transforms. The transformation parameters are determined by maximum likelihood estimation, so that the log-likelihood function of the transformed data reaches its maximum value. When the transformation parameter is 0, the transformation degenerates into a logarithmic transformation, and when the transformation parameter is 1, the original data remains unchanged.

[0047] The wavelet transform multi-scale denoising process involves decomposing the real-time observation data of the ground meteorological station into wavelet coefficients of different frequency scales, using a soft threshold shrinkage method to suppress noise components for high-frequency wavelet coefficients, with the threshold of the soft threshold shrinkage method determined based on the noise standard deviation and the logarithm of the data length, keeping the low-frequency wavelet coefficients unchanged to preserve the main components of the signal, and reconstructing the denoised real-time observation data of the ground meteorological station through inverse wavelet transform.

[0048] The sliding standard deviation is calculated using a sliding window with a length of 10 prediction time points. The standard deviation of the prediction deviation sequence within the sliding window is calculated, with the sliding window moving forward by one time point each time point. The sliding standard deviation is satisfied when the sliding standard deviations of three consecutive sliding windows all meet the following condition. The Kalman filter smoothing process is triggered at that time.

[0049] The method for adjusting the fusion weight coefficient is to recalculate the historical error of the prediction at each time scale based on the prediction deviation sequence after Kalman filtering smoothing, normalize the reciprocal of the historical error to obtain a new fusion weight coefficient, and replace the original fusion weight coefficient with the new fusion weight coefficient for subsequent calculation of the preliminary power generation prediction value.

[0050] Wherein, the distribution skewness coefficient is the third central moment of the smoothed power generation prediction value divided by the cube of the standard deviation, and when the distribution skewness coefficient... satisfy Time indicates that the predicted value distribution is right-skewed, when the skewness coefficient of the distribution... satisfy The time indicates that the predicted value distribution is left-skewed.

[0051] The signal-to-noise ratio (SNR) is the ratio of the effective meteorological signal power to the noise power in the real-time observation data of the ground meteorological station. It is calculated by decomposing the real-time observation data sequence of the ground meteorological station into trend components and fluctuation components, and then converting the ratio of the trend component power to the fluctuation component power into decibels.

[0052] Wherein, the effective meteorological signal power is the power of the signal component reconstructed from the low-frequency wavelet coefficients and mid-frequency wavelet coefficients retained after wavelet transform, the signal component contains the true variation law of meteorological elements, and the noise power is the power of the signal component corresponding to the high-frequency wavelet coefficients suppressed by the soft threshold shrinkage method.

[0053] Furthermore, the method also includes a step to quantify the uncertainty of the prediction results. By using the Monte Carlo dropout method to randomly deactivate some neurons in the inference stage of the multi-scale meteorological fusion prediction model, the probability distribution of the solar irradiance prediction sequence is obtained by repeating the inference 100 times, and the prediction mean and prediction standard deviation are calculated as uncertainty indicators.

[0054] Furthermore, the method also includes an outlier detection and correction steps. An isolated forest algorithm is used to identify outlier prediction points in the balanced power generation prediction values. The isolated forest algorithm constructs an isolation tree by randomly selecting features and segmentation thresholds, and calculates the average path length of each prediction point as an anomaly score. When the anomaly score... satisfy If the value is identified as an outlier, it is corrected by linear interpolation of adjacent normal predicted values.

[0055] The smoothing process based on prediction error fluctuations monitors the statistical characteristics of the prediction deviation sequence. When the fluctuation intensity exceeds the historical normal range, the Kalman filter smoothing mechanism is activated to suppress drastic fluctuations in the prediction results, ensuring the stability and usability of the prediction curve. This technique improves the stability of the prediction results and reduces the risk of decision-making errors caused by short-term fluctuations. The equalization process based on data distribution detects the statistical distribution shape of the smoothed power generation prediction values. When the distribution is significantly skewed, the Box-Cox transform equalization process is performed to make the data distribution more symmetrical, improving the accuracy of subsequent statistical analysis and error assessment. This technique improves the statistical characteristics of the prediction results and optimizes the reliability of error assessment indicators. The noise interference suppression process evaluates the signal-to-noise ratio level of the real-time observation data from the ground meteorological station. When noise pollution is severe, the wavelet transform multi-scale denoising method is used to separate effective information and random disturbances, improving the input data quality and thus improving the input conditions of the prediction model. This technique reduces the impact of measurement noise on prediction accuracy and improves the model's adaptability to low-quality data.

[0056] In addition, the present invention also provides a method for forming a photovoltaic power plant power generation prediction system by means of a computer, wherein the computer is provided with a storage medium, the storage medium stores program instructions, and the program instructions execute the above-mentioned photovoltaic power plant power generation prediction method when running in the computer.

[0057] The specific implementation methods of the above steps are described in detail below.

[0058] The specific implementation of step S01 is as follows: First, satellite cloud image sequence data covering a 100km×100km area around the photovoltaic power station is received through the visible light and infrared channels of meteorological satellites. The time resolution of data acquisition is set to 5 minutes to ensure that changes in solar irradiance caused by rapid cloud movement can be captured. At the same time, a ground meteorological station is deployed in the photovoltaic power station to monitor key meteorological parameters such as solar irradiance, ambient temperature, wind speed, and humidity in real time. The sampling frequency of the meteorological station is set to 1 minute to ensure the continuity and integrity of the data. The numerical weather prediction system is connected to obtain hourly forecast data of meteorological elements for the next 72 hours, including parameters such as temperature, humidity, cloud cover, and wind speed. The daily power generation records and corresponding meteorological condition records of the past 90 days are extracted from the historical database to establish a statistical relationship between power generation and meteorological conditions. At the same time, the operating age data of the photovoltaic modules are collected, including the initial commissioning time, cumulative operating hours, and historical maintenance records. The purpose of this step is to provide comprehensive basic data support for subsequent multi-timescale prediction and module degradation analysis, ensuring that the input data of the prediction model covers multiple time dimensions from short-term cloud changes to long-term climate trends.

[0059] The specific implementation of step S02 is as follows: The satellite cloud image sequence data is decomposed into multiple scales using a wavelet packet decomposition algorithm. By selecting appropriate wavelet basis functions, such as the Daubechies wavelet, the cloud image signal is decomposed into different frequency components. High-frequency components with a time window length of 5–30 minutes are extracted as short-term cloud change characteristics. These high-frequency components can reflect the rapid movement and instantaneous occlusion effect of clouds. For real-time observation data from ground meteorological stations, the empirical mode decomposition method is used to decompose the signal into multiple intrinsic mode functions. Mode components with periods in the range of 1–6 hours are selected as medium-term meteorological fluctuation characteristics. Features can describe the periodic fluctuations and intraday variations of meteorological elements. For numerical weather prediction model output data, ensemble empirical mode decomposition combined with trend extraction algorithm is used. White noise auxiliary signal is added to improve mode aliasing problem. Low-frequency trend components with a time window length of 12 to 72 hours are extracted as long-term climate trend features. These features can reflect the evolution process of large-scale weather systems and seasonal climate patterns. The purpose of multi-time-scale decomposition is to decouple complex meteorological processes into independent variation features at different time scales, so that the prediction model can model the variation patterns at different time scales separately, thereby improving the prediction accuracy.

[0060] The specific implementation of step S03 is as follows: An electroluminescence detection device is used to perform imaging detection on each module array within the photovoltaic power station. During detection, a forward bias voltage is applied to the module to cause it to emit near-infrared electroluminescence radiation. Electroluminescence images are captured by a high-sensitivity camera equipped with a 1100–1200 nm wavelength filter. The spatial resolution of the image is set to at least be able to distinguish the details of individual cells. Simultaneously, an infrared thermal imager is used to measure the heat distribution of the module. The temperature resolution of the infrared thermal imager is set to 0.1℃ to ensure that minute temperature differences can be detected. The wavelength range of the thermal imaging is selected in the long-wave infrared range of 8–14 μm to improve the sensitivity to temperature changes. The detection work is carried out on a quarterly basis with comprehensive scanning to ensure that early signs of module performance degradation can be detected in a timely manner. The purpose of this step is to obtain internal defect information and surface temperature distribution information of the photovoltaic module through non-destructive testing technology, providing a reliable physical basis for subsequent module health status assessment and degradation prediction.

[0061] The specific implementation of step S04 is as follows: Morphological image processing algorithms are used to extract features from the electroluminescence detection image. First, the image is divided into normal and dark areas using a grayscale thresholding method. The ratio of the dark area area to the total component area is calculated to obtain the proportion of the latent failure area. Edge detection operators such as the Canny operator are used to extract the edge contours of the cracks. A thinning algorithm is used to convert the crack edges into single-pixel-width skeleton lines. The total length of the skeleton lines is counted and divided by the component area to obtain the microcrack distribution density, in units of... The process involves detecting the positions of main grid lines and fine grid lines using Hough transform, analyzing the continuity of grid lines to identify broken grid locations and count the number of broken grids. Temperature features are extracted from infrared thermal imaging images using statistical analysis methods. The average temperature of each pixel on the module surface is calculated as the module's average temperature. Pixel regions with temperatures exceeding the module's average temperature by more than 10°C are identified as local hot spots, and their number is counted. The difference between the highest temperature in the hot spot and the module's average temperature is calculated to obtain the hot spot temperature difference. The standard deviation of the module surface temperature is calculated, and its reciprocal is used as a quantitative indicator of temperature distribution uniformity. The extracted microcrack distribution density, the proportion of latent failure areas, the number of broken grids in the cells, the hot spot temperature difference, the temperature distribution uniformity, and the number of local hot spots are used to construct a six-dimensional module health status feature vector. The purpose of this step is to convert image information into quantifiable health indicators, providing structured input features for the module degradation model.

[0062] The specific implementation of step S05 is as follows: The standard deviation of historical prediction error is calculated based on the difference between actual power generation and historical predicted power generation in historical power generation data. This standard deviation reflects the prediction stability and accuracy level of the prediction model over the past 90 days. Short-term cloud change characteristics, medium-term meteorological fluctuation characteristics, long-term climate trend characteristics, and the standard deviation of historical prediction error are input into a multi-scale meteorological fusion prediction model. This model uses three parallel bidirectional long short-term memory (LSTM) network branches to process meteorological characteristics at different time scales. The bidirectional LSTM network can simultaneously capture the forward and backward dependencies of the time series. Each bidirectional LSTM network branch is connected to a multi-head self-attention module containing eight attention heads. The relationship between each time step and other time steps in the sequence is calculated through a scaled dot product attention mechanism. The output features of the three branches are weighted and fused through a cross-scale attention fusion layer. The fusion layer uses a gating mechanism to dynamically adjust the contribution weights of each time scale. The gating weights are calculated using the Sigmoid activation function based on the standard deviation of historical prediction errors. Time scales with smaller historical errors will receive larger fusion weights. The fused feature vector is then processed through two fully connected network layers to extract higher-order features. Finally, a conditional random field layer is used to output the solar irradiance prediction sequence. The conditional random field layer uses the Viterbi algorithm to solve for the optimal annotation path and applies a temporal continuity constraint to the prediction sequence to ensure that the prediction values ​​at adjacent times conform to physical laws. The purpose of this step is to comprehensively utilize meteorological information at multiple time scales and adaptively adjust the weights of each scale based on historical prediction performance, thereby improving the prediction accuracy for changes at different time scales.

[0063] The specific implementation of step S06 is as follows: The component health status feature vector is input into a convolutional neural network (CNN) for spatial feature extraction. The CNN uses a 3×3 kernel and 32 kernels. Through multi-layer convolution and pooling operations, it extracts spatial correlation patterns between health features. The monthly power decay rate records from the past 12 months of historical power generation data are input into a recurrent neural network (RNN) to extract temporal evolution patterns. The RNN can capture the long-term dependence and periodic change patterns of the decay process. In the feature fusion layer, the spatial features extracted by the CNN and the temporal features extracted by the RNN are concatenated. The concatenated feature vector is then input into a noise injection module based on stochastic differential equations. This module models the feature transformation process as a stochastic dynamic system containing deterministic drift terms and random diffusion terms. To characterize the average trend of module degradation, the random diffusion term introduces structured noise through Brownian motion to simulate the random fluctuation component in the degradation process. The intensity coefficient of Brownian motion is determined by the product of the module's operating years and the uniformity of temperature distribution. The longer the operating years or the more uneven the temperature distribution, the greater the noise intensity. The Euler-Maruyama numerical method is used to discretize and solve the stochastic differential equation. The parameter distribution of the noise injection module is optimized through stochastic variational inference. The variational distribution adopts a Gaussian distribution with diagonal covariance. Gradient backpropagation is achieved through reparameterization techniques. The feature vector processed by the noise injection module outputs the power degradation coefficient and the power degradation trend curve for the next 30 days through a three-layer fully connected network. The purpose of this step is to accurately model the nonlinear degradation characteristics and random fluctuation law of photovoltaic modules and improve the detection sensitivity of sudden performance degradation.

[0064] The specific implementation of step S07 is as follows: A preliminary power generation prediction value is calculated based on the solar irradiance prediction sequence and the power attenuation coefficient. During calculation, the predicted solar irradiance is multiplied by the nominal power of the photovoltaic module and then by the power attenuation coefficient to obtain the preliminary power generation at each time point. The reciprocal of the historical prediction error standard deviation is normalized to obtain the prediction confidence level for each time scale. The confidence level reflects the reliability level of predictions at different time scales. The preliminary power generation prediction values ​​corresponding to short-term, medium-term, and long-term time scales are weighted and fused with their respective prediction confidence levels. During fusion, the prediction values ​​at each time scale are multiplied by their corresponding confidence weights and then summed to obtain the fused power generation prediction value. The purpose of this step is to comprehensively consider the two key factors of solar irradiance variation and module performance degradation, and to weight the prediction results at different time scales based on historical prediction performance, thereby obtaining a more accurate and reliable power generation prediction result.

[0065] The specific implementation of step S08 is as follows: The prediction deviation sequence between the fused power generation prediction value and historical power generation data is calculated. A sliding window with a length of 10 prediction time points is used to calculate the standard deviation of the prediction deviation sequence within the sliding window. The sliding window is moved forward by one time point each time point for iterative calculation. When the standard deviation of three consecutive sliding windows is greater than 15%, it is determined that the prediction deviation fluctuation is too large, and the Kalman filter smoothing mechanism is activated. The Kalman filter describes the dynamic evolution of the prediction deviation sequence through a state-space model. The state update equation includes a system noise term, and the observation equation includes a measurement noise term. The true state of the prediction deviation is estimated through two recursive steps: observation update and time update. The Kalman gain is calculated based on the system noise covariance and the measurement noise covariance. The Kalman gain is optimized by minimizing the mean square error of the state estimation. After Kalman filtering, a smoothed power generation prediction value is obtained. At the same time, the historical error of each time scale prediction is recalculated based on the prediction deviation sequence after smoothing. The inverse of the historical error is normalized to obtain a new fusion weight coefficient. The new fusion weight coefficient replaces the original weight and is used to update the fusion weight of each time scale feature in the multi-scale meteorological fusion prediction model. The purpose of this step is to suppress the drastic fluctuations of the prediction results to ensure the stability of the prediction curve, and to improve the model's response to prediction errors by adaptively adjusting the fusion weight.

[0066] The specific implementation of step S09 is as follows: The distribution skewness coefficient of the smoothed power generation prediction values ​​over different time periods is statistically analyzed. The skewness coefficient is calculated by dividing the third central moment of the prediction value by the cube of the standard deviation. When the absolute value of the skewness coefficient is greater than 0.8, it is determined that the distribution of the prediction values ​​is significantly skewed. The smoothed power generation prediction values ​​are then subjected to Box-Cox transformation for equalization. The Box-Cox transformation uses a family of power transforms to convert non-normally distributed data into an approximately normal distribution. The transformation parameters are determined by the maximum likelihood estimation method. The maximum likelihood estimation solves for the optimal transformation parameters by maximizing the log-likelihood function of the transformed data. When the transformation parameters are close to 0, the transformation degenerates into a logarithmic transformation; when the transformation parameters are close to 1, the original data remains basically unchanged. After processing by the Box-Cox transformation, the equalized power generation prediction value is obtained and output as the final prediction result. The purpose of this step is to improve the symmetry of the predicted value distribution, making the data more consistent with the normal distribution assumption, thereby improving the accuracy of subsequent statistical analysis and error assessment.

[0067] Step S10 is an optional step. Its specific implementation involves decomposing the real-time observation data sequence from the ground meteorological station into trend components and fluctuation components. Low-frequency, mid-frequency, and high-frequency wavelet coefficients corresponding to the trend and fluctuation components are extracted using wavelet transform. The signal component power reconstructed from the low-frequency and mid-frequency wavelet coefficients is calculated as the effective meteorological signal power, and the signal component power corresponding to the high-frequency wavelet coefficients is calculated as the noise power. The signal-to-noise ratio (SNR) is equal to the ratio of the effective meteorological signal power to the noise power and converted to decibels. When the SNR is less than 20 dB, the data is considered to be severely polluted by noise. Multi-scale wavelet transform is then employed. The noise reduction method denoises the real-time observation data of ground weather stations. During noise reduction, a soft threshold shrinkage method is used to suppress noise components for high-frequency wavelet coefficients. The soft threshold is calculated based on the noise standard deviation and the logarithm of the data length. Low-frequency wavelet coefficients are kept unchanged to preserve the main components of the signal. The denoised real-time observation data of ground weather stations is reconstructed through inverse wavelet transform. After updating the real-time observation data of ground weather stations, the process returns to step S02 to perform multi-time-scale decomposition again. The purpose of this step is to improve the quality of input data, reduce the impact of measurement noise on prediction accuracy, and ensure that the prediction model can still maintain good prediction performance under low-quality data conditions.

[0068] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the technical problem lies in breaking through the technical limitations of traditional single-scale modeling and static attenuation correction, and establishing a two-dimensional dynamic collaborative mechanism for meteorological forecasting and component attenuation. In the meteorological forecasting dimension, by performing multi-timescale decomposition on satellite cloud image sequences, ground meteorological station observation data, and numerical weather prediction model output data, meteorological changes are deconstructed into short-term cloud layer change characteristics, medium-term meteorological fluctuation characteristics, and long-term climate trend characteristics. The temporal dependencies of each scale are independently extracted using three parallel bidirectional long short-term memory network branches. Then, the contribution weights of each scale are dynamically adjusted according to the standard deviation of historical prediction errors through a cross-scale attention fusion layer, enabling the model to adaptively emphasize prediction information at different time scales according to different meteorological conditions. In the component degradation dimension, a six-dimensional health status feature vector, including microcrack distribution density and hot spot temperature difference, is extracted from electroluminescence detection images and infrared thermal imaging images. This vector is then input into the component degradation dynamic tracking model. A noise injection module based on stochastic differential equations is used to model the degradation process as a stochastic dynamic system containing deterministic drift and stochastic diffusion terms. The intensity coefficient of the Brownian motion term is determined by the product of the component's operating years and the temperature distribution uniformity, accurately describing the nonlinear characteristics and stochastic fluctuations of the degradation process. The logical consistency of this technical solution lies in its multi-scale decomposition of complex meteorological changes into independently modelable time-scale components, achieving dynamic weighted fusion between scales through an attention mechanism. Simultaneously, modeling component degradation as a stochastic process captures uncertainties in actual operation. Finally, the synergistic effect of a two-layer prediction architecture enables coupled prediction of meteorological conditions and component status.

[0069] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0070] In this embodiment, the specific implementation of steps S01-S03 is the same as described above, and will not be repeated in detail here.

[0071] The specific implementation of step S04 involves performing feature extraction processing on the electroluminescence detection image and the infrared thermal imaging image. The formula for calculating the microcrack distribution density is as follows:

[0072] ;

[0073] In the formula, The microcrack distribution density is dimensionless. The total length of the detected cracks, in units of ; For reference length, the value is 1. ; The total area of ​​the components, in units of ; For reference area, the value is 1. .

[0074] The formula for calculating the proportion of the area of ​​latent failure zone is as follows:

[0075] ;

[0076] In the formula, The percentage of the area of ​​the latent failure region is dimensionless. The dark area in the electroluminescence detection image is expressed in units of 1. ; The total area of ​​the components, in units of .

[0077] The formula for calculating the hot spot temperature difference is as follows:

[0078] ;

[0079] In the formula, This represents the temperature difference of the hot spot, which is dimensionless. This represents the highest temperature in the hotspot region, expressed in °C. The average temperature of the components is expressed in °C. The reference temperature is 1℃.

[0080] The formula for calculating temperature distribution uniformity is as follows:

[0081] ;

[0082] In the formula, The uniformity of temperature distribution is dimensionless. This represents the standard deviation of the component surface temperature, in °C. The reference temperature standard deviation is set to 1℃.

[0083] The expression for the component health status feature vector is:

[0084] ;

[0085] In the formula, The component health status feature vector is a six-dimensional column vector. This refers to the number of broken grids in the battery cell. This refers to the number of localized hotspots. For reference quantity, the value is 100.

[0086] The specific implementation of step S05 is as follows: the formula for calculating the standard deviation of historical prediction error is expressed as follows:

[0087] ;

[0088] In the formula, The standard deviation of historical prediction errors is dimensionless. The historical number of days is 90. For the first The actual daily power generation, in units of ; For the first Historical projected power generation for the day, in units of ; The nominal daily power generation of the module, in units of .

[0089] The formula for calculating the attention weight coefficient of the multi-scale meteorological fusion prediction model is as follows:

[0090] ;

[0091] In the formula, For the first Attention weight coefficients at each time scale, dimensionless; For the first Adjustment factors at each time scale, short time scale The value is 1.5, in the medium timescale. The value is 1.0, for long-term scales. The value is 0.8; For the first The historical forecast error standard deviation at each time scale, in units of ; The reference error standard deviation is set to 1. .

[0092] The core formulas of the sequence labeling optimization algorithm based on the attention mechanism in the multi-scale meteorological fusion prediction model include the self-attention mechanism and the conditional random field layer. The self-attention mechanism is expressed by the scaling dot product attention formula as follows:

[0093] ;

[0094] In the formula, For query matrix, the dimension is ; The key matrix has dimensions of . ; It is a value matrix with dimension . ; The sequence length; The dimension of the key vector is 64. The dimension of the value vector is 64; It is a normalized exponential function.

[0095] The expression for the multi-head attention mechanism is:

[0096] ;

[0097] In the formula, For the first The output of each attention head; The number of attention heads is set to 8. , , For the first The projection matrix of the head; To output the projection matrix; This indicates a splicing operation.

[0098] The sequence scoring function of the conditional random field layer is expressed as follows:

[0099] ;

[0100] In the formula, For label sequence In the input sequence The scoring under the given conditions is dimensionless; The sequence length; For from the tag Transfer to label The transfer fraction; The normalization factor for the transfer scores is 10; For a moment Launch Tag The launch fraction; This is the emission fraction normalization factor, with a value of 10.

[0101] The dynamic programming recurrence relation for solving the optimal labeled path using the Viterbi algorithm is expressed as follows:

[0102] ;

[0103] In the formula, For a moment In state The maximum path fraction, dimensionless; It is the set of all possible states; For a moment State index; For a moment State index; The state transition score; This is the normalization factor for the transfer score; For the number of launches; This is the emission fraction normalization factor.

[0104] The formula for generating adversarial examples in adversarial training is expressed as follows:

[0105] ;

[0106] In the formula, For adversarial examples, with input Same dimensions; This is the original input sample; The perturbation amplitude is taken as 5% of the standard deviation of the input feature sequence; It is a symbolic function; For loss function For input The gradient; This is a real label.

[0107] The specific implementation of step S06 is as follows: the formula for calculating the Brownian motion intensity coefficient is expressed as follows:

[0108] ;

[0109] In the formula, is the Brownian motion intensity coefficient, which is dimensionless; This is the baseline strength coefficient, with a default value of 0.05. The component's lifespan, in years; For reference operating years, the value is taken as 1 year; For reference temperature distribution uniformity, a value of 1 is used; The uniformity of temperature distribution is dimensionless.

[0110] The expression for the power attenuation coefficient is:

[0111] ;

[0112] In the formula, The power attenuation coefficient is dimensionless. The initial power factor is set to 1. This is the age-related decay factor, with an empirical value of 0.007; The microcrack attenuation coefficient has an empirical value of 0.15. This is the attenuation coefficient for the failure zone, with an empirical value of 0.25.

[0113] The expression for the power decay trend curve is:

[0114] ;

[0115] In the formula, For the future The predicted power attenuation trend for the day, in units of ; The nominal instantaneous power of the component, in units of ; The power attenuation coefficient is dimensionless. The predicted number of days ranges from 1 to 30 days. For reference days, the value is 1 day; The linear decay rate coefficient has an empirical value of 0.0002. The defect acceleration attenuation coefficient has an empirical value of 0.05. The density of microcrack distribution is dimensionless.

[0116] In the component decay dynamic tracking model, the noise injection learning framework based on stochastic differential equations models the feature evolution process as a stochastic dynamic system. The general form of the stochastic differential equations is as follows:

[0117] ;

[0118] In the formula, For a moment The eigenvector state; The drift term function represents the deterministic evolution trend; Let be the diffusion term function, representing the intensity of random fluctuations; For the differential increment of Brownian motion, obeys distributed; For time derivative.

[0119] The Euler-Maruyama method is used to numerically discretize stochastic differential equations. The discretization formula is expressed as follows:

[0120] ;

[0121] In the formula, For a moment The eigenvector state; For a moment The eigenvector state; For the drift term function; For time step; This is a diffusion term function, the value of which is related to the Brownian motion intensity coefficient. Proportional; Let be a standard normally distributed random vector, following... distributed.

[0122] The Fokker-Planck equation describing the evolution of the probability density of eigenvectors is as follows:

[0123] ;

[0124] In the formula, For feature vectors At any moment The probability density function; The dimension of the feature vector; The eigenvector of the eigenvector One component; The first term of the drift term function One component; The diffusion matrix of the th Line 1 Column elements.

[0125] Stochastic variational inference optimizes variational parameters by maximizing the lower bound of evidence. The expression for the lower bound of evidence is:

[0126] ;

[0127] In the formula, The lower bound of the evidence is dimensionless; For variational posterior distribution, given by parameters control; Indicates distribution The following expectations; Let be the likelihood function, given by the parameters control; For observation output; Let KL divergence be a metric. This is the prior distribution.

[0128] The expression for the reparameterization technique is:

[0129] ;

[0130] In the formula, These are latent variables obtained through sampling; The mean vector of the variational distribution is generated by the neural network based on the input. and parameters Output; Let be the standard deviation vector of the variational distribution; Let be a standard normally distributed random vector, following... distributed; This represents element-wise product.

[0131] The specific implementation of step S07 is as follows: the calculation formula for the preliminary power generation forecast is expressed as follows:

[0132] ;

[0133] In the formula, For the first A time scale at time Preliminary power generation forecast, in units of ; For the first A time scale at time The predicted solar irradiance values, in units of ; For reference irradiance, the value is taken as 1000. ; The nominal instantaneous power of the component, in units of ; The power attenuation coefficient is dimensionless.

[0134] The formula for calculating prediction confidence is as follows:

[0135] ;

[0136] In the formula, For the first Prediction confidence at each time scale, dimensionless; For the first The historical forecast error standard deviation at each time scale, in units of ; The reference error standard deviation is set to 1. .

[0137] The formula for calculating the predicted value of integrated power generation is expressed as follows:

[0138] ;

[0139] In the formula, For a moment The predicted value of the combined power generation, in units of ; For the first Prediction confidence at each time scale, dimensionless; For the first A time scale at time Preliminary power generation forecast, in units of .

[0140] The specific implementation of step S08 is as follows: the calculation formula for the prediction deviation sequence is expressed as follows:

[0141] ;

[0142] In the formula, For a moment The prediction bias is dimensionless. For a moment The predicted value of the combined power generation, in units of ; For a moment Historical power generation data, in units of ; This is a normalized power reference, in units of The value is the nominal instantaneous power of the component.

[0143] The formula for calculating the sliding standard deviation is as follows:

[0144] ;

[0145] In the formula, For a moment The sliding standard deviation is dimensionless. This is the length of the sliding window, with a value of 10. For a moment The prediction bias is dimensionless.

[0146] The state-space model of the Kalman filter includes a state update equation and an observation equation. The state update equation is expressed as follows:

[0147] ;

[0148] In the formula, For a moment The state vector contains the true state of the prediction bias and its rate of change; The state transition matrix is ​​expressed as follows:

[0149] ;

[0150] In the formula, For time step, the unit is ; Let be the system noise vector, which follows a zero-mean covariance matrix. The Gaussian distribution.

[0151] The observation equation is expressed as follows:

[0152] ;

[0153] In the formula, For a moment The observed values, i.e., the observed values ​​of the prediction bias sequence; The observation matrix is ​​expressed as follows:

[0154] ;

[0155] In the formula, To measure noise, it follows a mean of zero and a variance of... The Gaussian distribution.

[0156] The formula for calculating Kalman gain is as follows:

[0157] ;

[0158] In the formula, For a moment The Kalman gain matrix; For a moment The prior error covariance matrix; The observation matrix; To measure the noise variance.

[0159] The formula for calculating the smoothed power generation forecast is as follows:

[0160] ;

[0161] In the formula, For a moment The smoothed power generation forecast, in units of ; For a moment The predicted value of the combined power generation, in units of ; This is a normalized power reference, in units of ; For a moment The state estimation vector; This represents the first component of the state vector, which is the estimated prediction bias.

[0162] The new formula for calculating the fusion weight coefficient is as follows:

[0163] ;

[0164] In the formula, For the first The new fusion weight coefficients at each time scale are dimensionless. For the first Each time scale is a historical error recalculated after smoothing, with units of [missing information]. ; The reference error standard deviation is set to 1. .

[0165] The specific implementation of step S09 is as follows: the formula for calculating the distribution skewness coefficient is expressed as follows:

[0166] ;

[0167] In the formula, The skewness coefficient is dimensionless. This represents the total number of predicted time points. For a moment The smoothed power generation forecast, in units of ; This is a normalized power reference, in units of ; The mean of the normalized predicted values ​​is dimensionless and is calculated using the following formula: ; The standard deviation of the normalized predicted values ​​is dimensionless and is calculated using the following formula: .

[0168] The formula for the Box-Cox transform is as follows:

[0169] ;

[0170] In the formula, For a moment The normalized transformation form of the equilibrium power generation forecast is dimensionless. For a moment The smoothed power generation forecast, in units of ; This is a normalized power reference, in units of ; The transformation parameters are determined through maximum likelihood estimation.

[0171] The final output of the balanced power generation prediction needs to be converted back to the actual power, and the expression is:

[0172] ;

[0173] In the formula, For a moment The estimated balanced power generation, in units of ; This is a normalized power reference, in units of ; For normalization transformation, it is dimensionless; These are the transformation parameters.

[0174] The specific implementation of step S10 is as follows: the formula for calculating the effective meteorological signal power is expressed as follows:

[0175] ;

[0176] In the formula, Effective meteorological signal power, unit: ; These are wavelet coefficients, and their units are the same as those used for meteorological data. This is the set of indices for low-frequency and mid-frequency wavelet coefficients.

[0177] The formula for calculating noise power is as follows:

[0178] ;

[0179] In the formula, Noise power, unit: ; These are wavelet coefficients, and their units are the same as those used for meteorological data. This is the set of indices for high-frequency wavelet coefficients.

[0180] The formula for calculating the signal-to-noise ratio is as follows:

[0181] ;

[0182] In the formula, Signal-to-noise ratio, unit: ; Effective meteorological signal power, unit: ; Noise power, unit: .

[0183] The formula for calculating the soft threshold of wavelet transform is as follows:

[0184] ;

[0185] In the formula, This is a soft threshold, and it has the same unit as the data. This represents the noise standard deviation, and is in the same unit as the data. The length of the data sequence.

[0186] The processing formula for the soft threshold shrinkage method is expressed as follows:

[0187] ;

[0188] In the formula, For the processed first One high-frequency wavelet coefficient; For the original number One high-frequency wavelet coefficient; For symbolic functions, defined as when , when , when ; This is a soft threshold.

[0189] To better understand and implement this invention, the following is a specific application scenario of embodiment 2: A technical team is responsible for the installation of a 50... The project aimed to optimize the power generation forecast of a photovoltaic power station. This station, commissioned in 2020, has been operating for five years. Historical operational data shows significant fluctuations in forecast accuracy, particularly under conditions of rapid cloud cover changes and accelerated component performance degradation. Traditional forecasting methods struggle to accurately capture actual power generation trends, impacting the station's dispatch decisions and economic efficiency. The technical team decided to employ the photovoltaic power station power generation forecasting method of this invention for systematic forecasting optimization of the station.

[0190] The technical team first collected multi-source data on the area where the photovoltaic power station is located. Satellite cloud image sequence data was acquired through the visible light and infrared channels of meteorological satellites, with a time resolution set to 5. Each image covers 100 square kilometers around the power station. ×100 The area covered. Real-time observation data from ground weather stations includes solar irradiance, ambient temperature, wind speed, and humidity, with a sampling frequency of 1. Numerical weather prediction models output data including data for the next 72 days. The meteorological element forecast values ​​have a time resolution of 3. Historical power generation data covers the past 90 years. The data includes daily power generation records and corresponding meteorological condition records. The data also shows the operational lifespan of the modules in the power station, which consists of eight module arrays with operational lifespans of 5 years, 5 years, 4.5 years, 4.5 years, 4 years, 4 years, 3.5 years, and 3 years, respectively.

[0191] The technical team performed multi-timescale decomposition processing on the collected data. Satellite cloud image sequence data was decomposed into short-timescale cloud layer change characteristics, with a time window length set to 15. The extracted cloud movement velocity vector is 12. The gradient of cloud cover change is 0.08. The cloud optical thickness fluctuation amplitude was 0.35. Real-time observation data from ground meteorological stations were decomposed into mesoscale meteorological fluctuation characteristics, with a time window length set to 3. The hourly rate of change of solar irradiance was extracted to be 85%. The daily temperature range is 13.5℃, and the main peak frequency of the wind speed fluctuation spectrum is 0.12. The numerical weather prediction model output data is decomposed into long-term climate trend features, with the time window length set to 48. The slope of the multi-day average solar irradiance trend was extracted to be 18%. The evolution path of the pressure system points northeast, and the amplitude of the seasonal periodic component is 120. .

[0192] The technical team conducts electroluminescence (EML) and infrared thermal imaging inspections on each module array within the photovoltaic power plant quarterly. The wavelength range for acquiring EML images is 1100–1200 nm. The infrared thermal imaging image acquisition temperature resolution is 0.1℃. The microcrack distribution density of the eight component arrays extracted from the electroluminescence detection image was 2.8. 3.1 2.5 2.9 1.9 2.2 1.5 and 1.2 The percentages of areas with latent failures were 0.048, 0.056, 0.042, 0.051, 0.033, 0.038, 0.025, and 0.019, respectively, with the number of broken grids in the battery cells being 18, 21, 15, 19, 11, 13, 8, and 6, respectively. The temperature differences of hot spots extracted from the infrared thermal imaging images of the eight module arrays were 8.5℃, 9.8℃, 7.2℃, 8.9℃, 5.6℃, 6.4℃, 4.1℃, and 3.3℃, ​​respectively; the temperature distribution uniformity was 0.42, 0.38, 0.46, 0.40, 0.54, 0.50, 0.62, and 0.68, respectively; and the number of local hot spots was 7, 9, 6, 8, 4, 5, 3, and 2, respectively. The technical team established health status feature vectors for the eight module arrays, with each feature vector containing data from the aforementioned six dimensions.

[0193] The technical team calculated the historical prediction error standard deviation to be 0.082 based on historical power generation data. Short-term cloud change characteristics, medium-term meteorological fluctuation characteristics, long-term climate trend characteristics, and the historical prediction error standard deviation were input into a multi-scale meteorological fusion prediction model. This model employs three parallel bidirectional long short-term memory network branches to process features at different time scales, with each branch having a hidden layer dimension of 128 and a multi-head self-attention module with 8 heads. The cross-scale attention fusion layer calculated attention weight coefficients of 0.35, 0.42, and 0.23 for short-term, medium-term, and long-term scales based on the historical prediction error standard deviation. The model outputs the future 72... Solar irradiance prediction sequences at different time scales, such as Figure 2 As shown, the predicted sequence indicates that solar irradiance will first decrease and then increase over the next three days.

[0194] The technical team input the component health status feature vector, power degradation records from historical power generation data, component operating years data, and temperature distribution uniformity into the component degradation dynamic tracking model. Historical power degradation records show that the monthly power degradation rates for the past 12 months were 0.0042, 0.0045, 0.0048, 0.0051, 0.0053, 0.0056, 0.0059, 0.0062, 0.0065, 0.0068, 0.0071, and 0.0074, respectively. The convolutional neural network in the component degradation dynamic tracking model has a 3×3 kernel size, 32 kernels, and a recurrent neural network with a hidden layer dimension of 64. The intensity coefficients of the Brownian motion term in the noise injection module based on stochastic differential equations are determined by the product of the component's operational years and the temperature distribution uniformity. The intensity coefficients for the eight component arrays are 2.10, 1.90, 2.07, 1.80, 2.16, 2.00, 2.17, and 2.04, respectively. The model outputs power attenuation coefficients for the eight component arrays as 0.928, 0.915, 0.936, 0.922, 0.951, 0.942, 0.963, and 0.972, respectively, and for the next 30 years... The power attenuation trend curve.

[0195] The technical team calculated preliminary power generation forecasts based on the solar irradiance prediction sequence and power attenuation coefficient. These preliminary power generation forecasts were then weighted and fused with the prediction confidence levels for each time scale. The prediction confidence levels were calculated by normalizing the inverse of the standard deviation of historical prediction errors. The prediction confidence levels for short-term, medium-term, and long-term scales were 0.38, 0.35, and 0.27, respectively. The fused power generation forecasts cover the next 72 years. The hourly power generation.

[0196] The technical team calculated the prediction deviation sequence between the fused power generation forecast and historical power generation data. Using a sliding window of 10 prediction time points, they calculated the moving standard deviation and found that the moving standard deviations for three consecutive sliding windows from prediction time points 18 to 27 were 0.162, 0.178, and 0.185, respectively, all exceeding the threshold of 0.15. The team then smoothed the prediction deviation sequence using Kalman filtering. The system noise covariance of the state-space model was set to 0.005, the measurement noise covariance to 0.012, and the calculated Kalman gain was 0.294. The smoothed prediction deviation sequence showed a significant reduction in fluctuation. Figure 3 As shown, the technical team recalculated the historical errors of the predictions at each time scale based on the smoothed prediction bias sequence, and the adjusted fusion weight coefficients were 0.32, 0.40 and 0.28, respectively.

[0197] The technical team analyzed the skewness coefficient of the smoothed power generation forecasts across different time periods and found that the skewness coefficient for a certain period was 0.87, exceeding the threshold of 0.8. The team then performed a Box-Cox transformation on the smoothed power generation forecasts to equalize them. Using maximum likelihood estimation, they determined the transformation parameter to be 0.42. After the transformation, the skewness coefficient of the data was reduced to 0.15, resulting in the obtained and output of the balanced power generation forecast.

[0198] During the forecasting process, the technical team detected a signal-to-noise ratio of 18 for the real-time observation data from ground weather stations. Below 20 The technical team performed wavelet transform multi-scale denoising on real-time observation data from ground weather stations, using the Daubechies wavelet basis function and setting the decomposition level to 5. High-frequency wavelet coefficients were denoised using a soft thresholding method to suppress noise components, with the threshold set to 0.35 based on the noise standard deviation and the logarithm of the data length. The signal-to-noise ratio was improved to 25 after denoising. The data quality has been significantly improved. After updating the real-time observation data from the ground weather stations, the technical team reprocessed the data by decomposing it across multiple time scales, resulting in more accurate meteorological feature inputs.

[0199] It should be noted that the variables involved in this invention are explained in detail in Tables 1, 2, 3, and 4 below.

[0200] Table 1. Variable Explanation Table (Part 1)

[0201]

[0202] Table 2. Variable Explanation Table (Part Two)

[0203]

[0204] Table 3. Variable Explanation Table (Part 3)

[0205]

[0206] Table 4. Variable Explanation Table (Part Four)

[0207]

[0208] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the power generation of a photovoltaic power plant, characterized in that, The satellite cloud image sequence data of the region where the photovoltaic power station is located, the real-time observation data of the ground meteorological station, the output data of the numerical weather prediction model, the historical power generation data and the component operation life data are collected, the satellite cloud image sequence data, the real-time observation data of the ground meteorological station and the output data of the numerical weather prediction model are processed by multi-time scale decomposition to obtain short-time scale cloud change characteristics, medium-time scale meteorological fluctuation characteristics and long-time scale climate trend characteristics, the electroluminescent detection image and the infrared thermal imaging image of each component array in the photovoltaic power station are collected, the electroluminescent detection image and the infrared thermal imaging image are processed by feature extraction to establish a component health state feature vector, the short-time scale cloud change characteristics, the medium-time scale meteorological fluctuation characteristics, the long-time scale climate trend characteristics and the historical prediction error standard deviation are input into a multi-scale meteorological fusion prediction model to output a solar irradiance prediction sequence, the component health state feature vector, the power attenuation record in the historical power generation data, the component operation life data and the temperature distribution uniformity are input into a component attenuation dynamic tracking model to output a power attenuation coefficient and a power attenuation trend curve, the preliminary power generation prediction value is calculated according to the solar irradiance prediction sequence and the power attenuation coefficient, and weighted fusion is performed to obtain a fused power generation prediction value, and the prediction deviation sequence is smoothed and balanced to obtain a final power generation prediction value.

2. The method of claim 1, wherein, The real-time observation data of the ground meteorological station includes solar irradiance, ambient temperature, wind speed and humidity.

3. The method of claim 2, wherein, The output data of the numerical weather prediction model includes the forecast values of meteorological elements in the future 72 hours.

4. The method of claim 3, wherein, The historical power generation data includes daily power generation records and corresponding meteorological condition records in the past 90 days.

5. The method of claim 4, wherein, The steps of feature extraction processing of the electroluminescent detection image and the infrared thermal imaging image are specifically extracting micro-crack distribution density, hidden failure area area ratio and battery piece broken grid number from the electroluminescent detection image, and extracting hot spot temperature difference, temperature distribution uniformity and local overheating point number from the infrared thermal imaging image to establish a component health state feature vector.

6. The method of claim 5, wherein, The component health state feature vector is a six-dimensional vector containing micro-crack distribution density, hidden failure area area ratio, battery piece broken grid number, hot spot temperature difference, temperature distribution uniformity and local overheating point number.

7. The method of claim 6, wherein, The structure of the multi-scale meteorological fusion prediction model is that the input layer receives short-time scale cloud change characteristics, medium-time scale meteorological fluctuation characteristics, long-time scale climate trend characteristics and historical prediction error standard deviation, three parallel bidirectional long short-term memory network branches process short-time scale cloud change characteristics, medium-time scale meteorological fluctuation characteristics and long-time scale climate trend characteristics respectively, a multi-head self-attention module is connected after each bidirectional long short-term memory network branch, the outputs of the three bidirectional long short-term memory network branches are weighted and fused through a cross-scale attention fusion layer, the cross-scale attention fusion layer adopts a gating mechanism to dynamically adjust the contribution weight of each time scale, and the fused feature vector is output through a fully connected network and a conditional random field layer to output a solar irradiance prediction sequence.

8. The method of claim 7, wherein, The structure of the component attenuation dynamic tracking model is that the input layer receives the component health state feature vector, the power attenuation record in the historical power generation data, the component operation time data and the temperature distribution uniformity, the spatial mode of the component health state feature vector is extracted through the convolutional neural network, the time evolution law of the power attenuation record in the historical power generation data is extracted through the recurrent neural network, the features of the convolutional neural network and the recurrent neural network are spliced in the feature fusion layer, the spliced feature vector is input into the noise injection module based on the stochastic differential equation, and the feature vector is processed through the noise injection module based on the stochastic differential equation, and then the power attenuation coefficient and the power attenuation trend curve are output through the full connection network.

9. The method of claim 8, wherein, The noise injection module based on the stochastic differential equation models the feature transformation process as a stochastic dynamic system, simulates the random fluctuation component in the component attenuation process by introducing a Brownian motion term, and determines the strength coefficient of the Brownian motion term according to the product of the component operation time data and the temperature distribution uniformity.

10. The method of claim 9, wherein, The steps of calculating the preliminary power generation prediction value based on the solar irradiance prediction sequence and the power attenuation coefficient and performing weighted fusion are specifically that the preliminary power generation prediction value is weighted and fused with the prediction confidence of each time scale to obtain a fused power generation prediction value, and the prediction confidence is obtained by calculating the reciprocal normalization of the historical prediction error standard deviation.

Images