Photovoltaic power prediction method

Abstract

The invention relates to the field of photovoltaic power prediction, in particular to a photovoltaic power prediction method, and the method comprises the steps: obtaining the spectral response function data of different photovoltaic modules in a photovoltaic power station, and measuring the spectral irradiance distribution of the surface of each photovoltaic module in real time; according to the spectral response function data and the spectral irradiance distribution, calculating the effective spectral irradiance of each component type in a grouping manner according to the component types; counting the temperature coefficient of the batch to which each photovoltaic module belongs according to historical data, and establishing a module-level temperature-power correction factor; according to the string topological relation, string-level temperature correction power is obtained, and the string-level temperature correction power, the effective spectral irradiance and the global irradiance data of the numerical weather forecast are output into a corrected power station-level photovoltaic power predicted value; the problem of low photovoltaic power prediction precision caused by inaccurate modeling due to parameter difference and temperature influence of photovoltaic modules in photovoltaic power station power prediction is solved.

Description

Technical Field

[0001] The present invention relates to the field of photovoltaic power prediction, and specifically to a photovoltaic power prediction method. Background Art

[0002] As an important part of renewable energy, the installed capacity of photovoltaic power generation has been continuously growing rapidly. However, the output power of photovoltaic power plants is significantly affected by environmental factors and the characteristics of components themselves, resulting in great uncertainty in power prediction. Accurate power prediction is the key technical basis for power grid dispatching, energy storage system optimization, and electricity market trading, and is crucial for improving the stability and economic benefits of the power grid.

[0003] Currently, photovoltaic power prediction does not fully consider the spectral response differences and temperature coefficient discreteness of different batches and types of photovoltaic components. Especially in large-scale photovoltaic power plants, the mixing of batches of components leads to the accumulation of systematic errors and a significant decrease in prediction accuracy. Also, the influence of dynamic factors such as the aging of photovoltaic components and the change of the thermal resistance of packaging materials is not fully considered, resulting in the gradual amplification of errors in photovoltaic power after long-term operation.

[0004] Therefore, there is an urgent need for a high-precision photovoltaic power prediction method that can deeply integrate component heterogeneity characteristics and dynamic thermodynamic behaviors to improve the accuracy of photovoltaic power prediction. Summary of the Invention

[0005] (1) Technical Problems to be Solved

[0006] The purpose of the present invention is to provide a photovoltaic power prediction method to solve the problem of low accuracy of photovoltaic power prediction in photovoltaic power plants due to inaccurate modeling caused by differences in photovoltaic component parameters and temperature effects.

[0007] (2) Technical Solutions

[0008] To achieve the above purpose, the present invention provides a photovoltaic power prediction method, which includes:

[0009] S1. Real-time obtain the spectral response function data of different photovoltaic components in a photovoltaic power plant, where the different photovoltaic components include at least two component types with spectral response characteristics differing by more than a preset difference threshold.

[0010] S2. Through a multi-spectral sensor network deployed above the photovoltaic array, measure the spectral irradiance distribution incident on the surface of each photovoltaic component in real time; according to the spectral response function data and the spectral irradiance distribution, calculate the effective spectral irradiance of each component type by grouping according to component type.

[0011] S3. According to historical data, statistically analyze the temperature coefficients of each batch to which the photovoltaic modules belong, and establish a component-level temperature-power correction factor; according to the string topology relationship, superimpose the temperature-power correction factors of all photovoltaic modules within the same string to obtain the string-level temperature-corrected power.

[0012] S4. Input the string-level temperature-corrected power and the effective spectral irradiance into a pre-trained photovoltaic power prediction model, and combine the global irradiance data of numerical weather forecasting to dynamically output the corrected power prediction value at the power station level.

[0013] Further, the method for calculating the effective spectral irradiance of each component type by grouping according to the component type based on the spectral response function data and the spectral irradiance distribution includes:

[0014] Discretize the spectral irradiance distribution according to a preset wavelength interval to obtain the corresponding irradiance value and wavelength range within each wavelength interval.

[0015] Calculate the spectral response contribution value of each wavelength interval according to the response value of each wavelength interval in the spectral response function data of the component type and the measured irradiance value of the corresponding wavelength interval, and sum the weighted spectral response contribution values of all wavelength intervals to obtain the initial effective spectral irradiance of the component type.

[0016] Normalize the initial effective spectral irradiance according to the standard solar spectral distribution data under standard test conditions to obtain the effective spectral irradiance.

[0017] Further, the method for statistically analyzing the temperature coefficients of each batch to which the photovoltaic modules belong according to historical data and establishing a component-level temperature-power correction factor includes:

[0018] Associate the production batch information according to the unique identification code of the photovoltaic module, extract the temperature coefficient data set of each production batch of photovoltaic modules in the historical database, and statistically analyze the mean value and dispersion parameter of the temperature coefficients of each production batch of photovoltaic modules.

[0019] Real-time collect the backplane temperature data of the photovoltaic module, and dynamically calculate the component-level temperature-power correction factor according to the mean value and dispersion parameter of the temperature coefficient of its production batch; among them, for the temperature-power correction factor of the photovoltaic module in the production batch with a dispersion exceeding the preset dispersion threshold, a dispersion compensation weight is introduced, and the dispersion compensation weight has a negative correlation with the dispersion parameter.

[0020] If the production batch to which the photovoltaic module belongs is a mixed batch and the temperature coefficient difference exceeds the preset temperature coefficient difference threshold, then adjust the temperature-power correction factor weight according to the electrical connection relationship of the photovoltaic modules in different production batches within the string.

[0021] When it is detected that the temperature coefficient dispersion of the PV modules in a string continuously exceeds the historical statistical range, a stepped temperature control excitation signal is applied to the PV modules of this production batch, and the transient change slope of the temperature response curve is collected to invert the actual temperature coefficient, update the mean value and dispersion parameter of the temperature coefficient of the PV modules of this production batch, and regenerate the component-level temperature-power correction factor.

[0022] Further, the method of applying a stepped temperature control excitation signal to the PV modules of this production batch and collecting the transient change slope of the temperature response curve to invert the actual temperature coefficient when it is detected that the temperature coefficient dispersion of the PV modules in a string continuously exceeds the historical statistical range includes:

[0023] Apply a stepped temperature control excitation signal through the temperature control actuator integrated on the backplane of the PV module, and synchronously collect the transient change data of the backplane temperature of the PV module according to the preset sampling frequency.

[0024] Perform filtering and noise reduction processing on the transient change data, extract the response curves of each temperature step interval, calculate the instantaneous change slope and steady-state offset, and construct a temperature-time transfer function in combination with the dynamic heat capacity model of the PV module.

[0025] Match the temperature-time transfer function with the preset standard temperature coefficient response database, and invert the actual temperature coefficient through least squares iterative optimization and calculate the confidence interval; if the half-width of the confidence interval is less than the preset precision threshold, output the actual temperature coefficient.

[0026] Further, the method of performing filtering and noise reduction processing on the transient change data, extracting the response curves of each temperature step interval, calculating the instantaneous change slope and steady-state offset, and constructing a temperature-time transfer function in combination with the dynamic heat capacity model of the PV module includes:

[0027] Identify the starting point and ending point of the temperature step interval through the dynamic threshold segmentation algorithm, fit the response curve by cubic spline interpolation within each segmented temperature step interval, calculate the instantaneous change slope of each sampling point on the response curve, and extract the maximum instantaneous change slope as the step response eigenvalue.

[0028] Establish a dynamic heat capacity model according to the material parameters and structural parameters of the PV module; input the step response eigenvalue and steady-state offset into the dynamic heat capacity model to obtain the parameter matrix of the temperature-time transfer function, and perform real-time calibration on the parameter matrix through the Kalman filtering algorithm; if the calibration residual of the parameter matrix exceeds the preset residual tolerance threshold, trigger the adaptive update of the dynamic heat capacity model.

[0029] Further, the method of establishing a dynamic heat capacity model according to the material parameters and structural parameters of the PV module includes:

[0030] Construct a three-dimensional non-uniform grid model based on the material parameters and structural parameters of the photovoltaic module; the material parameters are the physical property parameters of the constituent materials of the photovoltaic module, and the structural parameters are the geometric characteristics and physical layout parameters of the photovoltaic module.

[0031] Correct the degradation coefficients of the material parameters according to the historical aging data of the photovoltaic module. The degradation coefficients include the annual growth rate of the thermal resistance of the encapsulation material and the attenuation curve of the heat capacity of the cell, and generate an initial thermal resistance-heat capacity correlation matrix according to the corrected material parameters.

[0032] Iteratively optimize the initial thermal resistance-heat capacity correlation matrix and the backplane temperature gradient distribution data collected in real time by the three-dimensional non-uniform grid model to obtain dynamic heat capacity parameters; generate a dynamic heat capacity model according to the optimized dynamic heat capacity parameters.

[0033] Further, the method for generating a dynamic heat capacity model according to the optimized dynamic heat capacity parameters includes:

[0034] Map the optimized dynamic heat capacity parameters to the three-dimensional non-uniform grid model according to the grid nodes, extract the thermal resistance and heat capacity values of each grid node and convert them into equivalent lumped parameters, and construct a dynamic heat capacity model according to the equivalent lumped parameters.

[0035] Perform double verification of the dynamic heat capacity model in the frequency domain and time domain. In the frequency domain, verify the amplitude-frequency characteristics and phase characteristics of the transfer function through swept-frequency excitation testing. In the time domain, verify the root mean square value of the residual between the predicted temperature and the actual temperature through step response temperature control experiments; if the verification results of the double verification in the frequency domain and time domain meet the preset verification accuracy, bind the dynamic heat capacity model to the unique identification code of the photovoltaic module.

[0036] Further, the method for dynamically outputting the corrected power prediction value of the power station-level photovoltaic by inputting the series-connected string-level temperature-corrected power and the effective spectral irradiance into the pre-trained photovoltaic power prediction model and combining the global irradiance data of numerical weather forecasting includes:

[0037] Summarize the series-connected string-level temperature-corrected power into a power station-level temperature-corrected power matrix according to the string topology relationship and the power station electrical connection diagram. Perform time series alignment processing on the power station-level temperature-corrected power matrix, the effective spectral irradiance, and the global irradiance data of numerical weather forecasting and then fuse them into a multi-dimensional input feature vector. Input the multi-dimensional input feature vector into the pre-trained photovoltaic power prediction model, and the photovoltaic power prediction model is trained by a hybrid neural network.

[0038] If an extreme weather event warning is detected, trigger multi-scale time window prediction to obtain the final power station-level photovoltaic power prediction value.

[0039] Compare the final predicted photovoltaic power value at the power station level with the real-time power monitoring data to obtain the power prediction error value. If the power prediction errors in N consecutive time windows exceed the preset power error threshold, adaptively adjust the pre-trained photovoltaic power prediction model, where N≥3.

[0040] Further, the method for triggering multi-scale time window prediction to obtain the final predicted photovoltaic power value at the power station level when detecting an extreme weather event warning includes:

[0041] The multi-scale time window prediction includes short-term prediction and medium- and long-term prediction; the short-term prediction generates high-frequency prediction results with a minute-level time window; the medium- and long-term prediction generates trend prediction results with an hour-level time window; according to the extreme weather event warning level and the duration of the weather event, dynamically weight and fuse the prediction results of the short-term prediction and the medium- and long-term prediction to obtain the final predicted photovoltaic power value at the power station level.

[0042] (3) Beneficial effects

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] 1. By considering the spectral response differences of different photovoltaic modules, measuring the spectral irradiance distribution in real time, and combining component-level temperature monitoring and temperature coefficient correction, the prediction accuracy is greatly improved. Especially in complex power station systems with mixed use of different photovoltaic modules, high-precision photovoltaic power prediction is achieved.

[0045] 2. A dynamic adaptive temperature compensation mechanism is established. The actual temperature coefficient is inverted through a stepped temperature control excitation signal, and a dynamic heat capacity model is constructed based on the component material properties and structural parameters, realizing real-time tracking and correction of the component temperature coefficient changes with aging, and solving the prediction deviation problem caused by the component temperature coefficient due to aging or batch differences that cannot be addressed by traditional methods.

[0046] 3. The ability to respond to extreme weather is provided. Through the dynamic weighted fusion of multi-scale time window prediction and combined with the adaptive adjustment mechanism of the model, the power prediction stability and reliability under extreme weather conditions are effectively improved, providing a more accurate decision-making basis for power grid dispatching and photovoltaic power station management. Description of the drawings

[0047] Figure 1 It is a flowchart of a photovoltaic power prediction method according to Embodiment 1 of the present invention. Detailed implementation manners

[0048] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0049] Before giving examples, it is necessary to elaborate on the application scenarios of the inventive concept of the present invention. The present invention is a photovoltaic power prediction method. The impact of the heterogeneity of photovoltaic module characteristics on power prediction is an underlying problem that has been underestimated for a long time. The spectral response functions of different photovoltaic modules (such as monocrystalline silicon and thin-film batteries) are significantly different; moreover, the discreteness of the temperature coefficients among mixed-batch photovoltaic modules is greater than that among photovoltaic modules of the same batch, resulting in low accuracy of photovoltaic power prediction.

[0050] Embodiment 1: As Figure 1 shown, this embodiment provides a photovoltaic power prediction method, and the method includes:

[0051] S1. Real-time obtain the spectral response function data of different photovoltaic modules in a photovoltaic power station, where the different photovoltaic modules include at least two component types with spectral response characteristic differences greater than a preset difference threshold; photovoltaic power stations usually mix different types of photovoltaic modules, including various types such as monocrystalline silicon, polycrystalline silicon, amorphous silicon, and CIGS thin films. There are obvious differences in the spectral response characteristics of different photovoltaic modules. The "preset difference threshold" is defined as the average quantum efficiency difference in the main working band (300 - 1100 nm) exceeding 5% or the peak response wavelength difference exceeding 30 nm. For example, the quantum efficiency of monocrystalline silicon modules in the near-infrared region (780 - 1100 nm) usually reaches 80 - 90%, while the quantum efficiency of some thin-film modules in this region may be only 50 - 60%; the peak response wavelength of monocrystalline silicon is about 950 nm, while the peak response wavelength of some thin-film modules may be in the range of 500 - 600 nm. This difference in spectral response characteristics leads to significant differences in the power generation efficiency of different modules under the same irradiation conditions.

[0052] S2. Through the multi-spectral sensor network deployed above the photovoltaic array, the spectral irradiance distribution incident on the surfaces of each photovoltaic module is measured in real time; according to the spectral response function data and the spectral irradiance distribution, the effective spectral irradiance of each module type is calculated by grouping according to module type; one sensor is deployed per 2 - 4 hectares of the power station area, and a triangular network structure is formed among the sensors to achieve optimal coverage. Each sensor can simultaneously monitor the spectral irradiance of 12 - 20 bands, and the band division is determined according to the key points of the characteristic curve of the photovoltaic module, covering the main working band of 300 - 1100 nm. For example, in a typical 10 MW photovoltaic power station, about 15 - 20 multi-spectral sensors can be deployed, distributed and installed in different areas of the power station to ensure comprehensive monitoring of the spectral distribution in the power station. The sensor acquisition frequency is 5 - 15 minutes per time in stable weather and can be increased to 1 - 5 minutes per time under rapidly changing weather conditions.

[0053] S3. According to the historical data, the temperature coefficients of each batch to which the photovoltaic modules belong are statistically analyzed, and a component-level temperature-power correction factor is established; according to the string topology relationship, the temperature-power correction factors of all photovoltaic modules within the same string are superimposed to obtain the string-level temperature-corrected power.

[0054] S4. The string-level temperature-corrected power and the effective spectral irradiance are input into the pre-trained photovoltaic power prediction model, and combined with the global irradiance data of numerical weather prediction, the corrected power prediction value at the power station level is dynamically output.

[0055] The method for calculating the effective spectral irradiance of each module type by grouping according to the module type according to the spectral response function data and the spectral irradiance distribution includes:

[0056] The spectral irradiance distribution is discretized according to a preset wavelength interval to obtain the irradiance value and wavelength range corresponding to each wavelength interval; the discretization process divides the continuous spectral distribution into multiple discrete wavelength intervals, which is convenient for calculation and data processing. The division of the preset wavelength interval takes into account the spectral response characteristics of the photovoltaic module, and the main working band of 300 - 1100 nm is divided into 20 wavelength intervals. Among them, in the region where the spectral response changes greatly (such as the response peak region of monocrystalline silicon at 800 - 1000 nm), a finer division (about 10 - 15 nm / interval) is adopted, and in the region where the response changes less, a coarser division (about 30 - 50 nm / interval) is adopted.

[0057] The spectral response contribution value of each wavelength interval is calculated based on the response value of each wavelength interval in the spectral response function data of the component type and the measured irradiance value of the corresponding wavelength interval, and the weighted sum of the spectral response contribution values of all wavelength intervals is used to obtain the initial effective spectral irradiance of the component type; for each wavelength interval, the response value of the wavelength interval in the spectral response function data of the photovoltaic component type is multiplied by the measured irradiance value of the corresponding wavelength interval to obtain the spectral response contribution value of each wavelength interval. For example, if the average spectral response value of a monocrystalline silicon component in the wavelength interval of 950 - 965 nm is 0.72 A / W and the measured average irradiance in this wavelength interval is 0.65 W / m 2 , then the spectral response contribution value of this wavelength interval to the monocrystalline silicon component is 0.468 A / m 2 . The weighted sum of the spectral response contribution values of all wavelength intervals is calculated according to the energy distribution ratio of each wavelength interval in the entire working spectrum to obtain the initial effective spectral irradiance of the component type.

[0058] Based on the standard solar spectrum distribution data under standard test conditions, the initial effective spectral irradiance is normalized to obtain the effective spectral irradiance. According to the standard solar spectrum distribution data (such as the AM1.5G standard spectrum) under standard test conditions (STC), the initial effective spectral irradiance is normalized to obtain the final effective spectral irradiance. The normalization formula is: effective spectral irradiance = initial effective spectral irradiance × (1000 / initial effective spectral irradiance under standard conditions), where 1000 represents the irradiance value (W / m 2 ) under standard test conditions. The normalization process ensures that the effective spectral irradiance measured under different weather conditions and at different times can be directly compared with the performance under standard test conditions, providing a unified benchmark for power prediction.

[0059] The method for statistically calculating the temperature coefficient of each batch of photovoltaic components based on historical data and establishing a component-level temperature-power correction factor includes:

[0060] According to the unique identification code of the photovoltaic module, the production batch information is associated, and the temperature coefficient data set of each production batch of photovoltaic modules in the historical database is extracted to calculate the mean value and dispersion parameter of the temperature coefficient of each production batch of photovoltaic modules; the unique identification code of the photovoltaic module is usually printed on the back panel or frame of the module in the form of a serial number or a QR code. The power station management system records the correspondence between the module ID and the installation location through the module installation registration process. The temperature coefficient test values of each production batch of photovoltaic modules at different operating times are stored in the historical database, including the initial test value and the regular test value. The mean value of the temperature coefficient is calculated using the weighted average method, and the weight of the recent test value is higher than that of the early test value to reflect the trend of the performance of the photovoltaic module over time. The dispersion parameter is calculated using the weighted standard deviation to reflect the dispersion of the temperature coefficient of photovoltaic modules in the same batch.

[0061] The backplane temperature data of the photovoltaic module is collected in real time, and the temperature-power correction factor of the photovoltaic module level is dynamically calculated according to the mean value of the temperature coefficient and the dispersion parameter of the production batch to which it belongs; wherein, the temperature-power correction factor of the photovoltaic module of the production batch whose dispersion exceeds the preset dispersion threshold introduces the dispersion compensation weight, and the dispersion compensation weight is negatively correlated with the dispersion parameter; the backplane temperature data is collected using a distributed temperature sensor network, and the sensor is installed at the center of the module backplane, with an accuracy of ±0.5℃, and the sampling frequency is synchronized with the irradiance data collection (5-15 minutes / time). The output power of the photovoltaic module is negatively correlated with the temperature, and the temperature coefficient represents the power attenuation ratio when the temperature rises by 1℃. The basic calculation formula of the module-level temperature-power correction factor is: correction factor = 1 + mean value of temperature coefficient × (measured backplane temperature - 25℃), where 25℃ is the temperature of the standard test conditions. For the production batch of photovoltaic modules whose dispersion exceeds the preset dispersion threshold, the temperature-power correction factor needs to introduce the dispersion compensation weight. The preset discreteness threshold is usually set at 0.05% / ℃ of the standard deviation of the temperature coefficient. The calculation formula for the discreteness compensation weight is: discreteness compensation weight = 1-(discreteness parameter-preset discreteness threshold) × k, where k is the empirical adjustment coefficient. When the discreteness parameter is close to the preset discreteness threshold, the compensation weight is close to 1, and the effect on the correction factor is small; when the discreteness parameter significantly exceeds the preset discreteness threshold, the compensation weight is reduced, reducing the reliance on the estimated temperature coefficient of the photovoltaic modules of the production batch.

[0062] If the production batches of the photovoltaic modules are mixed batches and the temperature coefficient difference exceeds the preset temperature coefficient difference threshold, the temperature-power correction factor weights are adjusted according to the electrical connection relationships of the photovoltaic modules of different production batches within the string; if the production batches of the photovoltaic modules are mixed batches and the temperature coefficient difference exceeds the preset temperature coefficient difference threshold, usually set to 0.08% / °C, the temperature-power correction factor weights are adjusted according to the electrical connection relationships of the photovoltaic modules of different production batches within the string. For the photovoltaic modules of different production batches connected in series, the minimum value method is used to determine the current-limiting factor of the string; for the photovoltaic modules of different production batches connected in parallel, a weighted average is taken according to the capacity ratios of the production batches in the parallel connection. Ensure that under the condition of mixing photovoltaic modules of different production batches, the temperature-power correction can accurately reflect the limiting effect of the actual electrical connection on the power output.

[0063] When it is detected that the temperature coefficient dispersion of the photovoltaic modules of a production batch within the string continuously exceeds the historical statistical range, a stepped temperature control excitation signal is applied to the photovoltaic modules of this production batch, and the transient change slope of the temperature response curve is collected to invert the actual temperature coefficient, update the temperature coefficient mean and dispersion parameters of the photovoltaic modules of this production batch, and regenerate the component-level temperature-power correction factor. By analyzing the corresponding relationship between the temperature change and the power output, calculate the current actual temperature coefficient, update the temperature coefficient database of this production batch, and regenerate a more accurate component-level temperature-power correction factor.

[0064] The method of applying a stepped temperature control excitation signal to the photovoltaic modules of a production batch within the string when it is detected that the temperature coefficient dispersion of the photovoltaic modules of a production batch within the string continuously exceeds the historical statistical range, and collecting the transient change slope of the temperature response curve to invert the actual temperature coefficient includes:

[0065] Apply a stepped temperature control excitation signal through the temperature control actuator integrated on the backplane of the photovoltaic module, and synchronously collect the transient change data of the backplane temperature of the photovoltaic module according to the preset sampling frequency; apply a stepped temperature control excitation signal through the temperature control actuator integrated on the backplane of the photovoltaic module. The temperature control actuator can be a thin-film electrothermal element attached to the backplane of the module, which can provide precisely controlled heating power. The stepped temperature control excitation signal is designed as multiple consecutive temperature steps. The initial step amplitude is 2 - 3°C, and the subsequent steps are adjusted according to the previous response, with a maximum not exceeding 5°C to avoid thermal shock. Each step lasts for 5 - 10 minutes to ensure that the module temperature reaches a stable state. Synchronously collect the transient change data of the backplane temperature of the photovoltaic module according to the preset sampling frequency. The sampling frequency is set to 1 - 5 Hz, which is sufficient to capture the dynamic process of the temperature change. At the same time, record the change data of the electrical parameters (open-circuit voltage, short-circuit current, maximum power point voltage and current) of the module to establish the corresponding relationship between the temperature change and the power change.

[0066] Filter and denoise the transient change data, extract the response curves of each temperature step interval, calculate the instantaneous change slope and steady-state offset, and construct a temperature-time transfer function in combination with the dynamic heat capacity model of the photovoltaic module; wavelet transform or moving average filter is used for filtering and denoising to remove measurement noise and environmental interference.

[0067] Match the temperature-time transfer function with a preset standard temperature coefficient response database, and inversely calculate the actual temperature coefficient through least squares iterative optimization and calculate the confidence interval; if the half-width of the confidence interval is less than the preset precision threshold, output the actual temperature coefficient. The standard temperature coefficient response database contains typical temperature response characteristic parameters of components of different types and different aging degrees. By comparing the differences between the measured response and the standard response, the most matching temperature coefficient value is found in combination with the least squares method. At the same time, calculate the confidence interval of the temperature coefficient to reflect the reliability of the estimation result. If the half-width of the confidence interval is less than the preset precision threshold, such as set to 0.01% / °C, it is considered that the estimation result is reliable, and the actual temperature coefficient is output for updating the database; otherwise, increase the test time or adjust the test parameters and then re-measure.

[0068] The method of filtering and denoising the transient change data, extracting the response curves of each temperature step interval, calculating the instantaneous change slope and steady-state offset, and constructing a temperature-time transfer function in combination with the dynamic heat capacity model of the photovoltaic module includes:

[0069] Identify the starting point and ending point of the temperature step interval through the dynamic threshold segmentation algorithm. In each segmented temperature step interval, obtain the response curve through cubic spline interpolation fitting, calculate the instantaneous change slope of each sampling point on the response curve, and extract the maximum instantaneous change slope as the step response characteristic value; the dynamic threshold segmentation algorithm automatically detects the boundary of the step signal according to the statistical characteristics of the temperature change rate. When the temperature change rate exceeds the threshold, it is marked as the step starting point, and after a certain time (usually 3-5 sampling periods) when it is lower than the threshold, it is marked as the ending point. In each segmented temperature step interval, obtain the response curve through cubic spline interpolation fitting. Cubic spline interpolation ensures the continuity and smoothness of the curve, reducing the influence of noise on subsequent analysis. Calculate the instantaneous change slope of each sampling point on the fitted curve, that is, the ratio of the temperature change rate (dT / dt) to the power change rate (dP / dt). Extract the maximum instantaneous change slope from all sampling points as the step response characteristic value, which reflects the maximum sensitivity of the photovoltaic module to temperature changes.

[0070] A dynamic heat capacity model is established based on the material parameters and structural parameters of the photovoltaic module; the step response eigenvalue and the steady-state offset are input into the dynamic heat capacity model to obtain the parameter matrix of the temperature-time transfer function, and the parameter matrix is calibrated in real time through the Kalman filter algorithm; if the calibration residual of the parameter matrix exceeds the preset residual tolerance threshold, an adaptive update of the dynamic heat capacity model is triggered. The dynamic heat capacity model is represented in the form of a thermoelectric equivalent network, which includes multiple thermal resistance and heat capacity elements to simulate the heat conduction characteristics of each layer (such as glass, EVA, solar cell, backplane, etc.) of the photovoltaic module. The step response eigenvalue and the steady-state offset are input into the dynamic heat capacity model to obtain the parameter matrix of the temperature-time transfer function. The parameter matrix contains all the coefficients describing the dynamic characteristics of the system, such as time constants, gain coefficients, etc. A second-order Kalman filter algorithm is used to calibrate the parameter matrix in real time. The initial process noise covariance matrix is determined according to the standard deviation of historical temperature fluctuations, the measurement noise covariance matrix is set according to the accuracy of the temperature sensor, and the state transition matrix is determined according to the thermal time constant of the module (usually 5-15 minutes). The Kalman filter can effectively handle measurement noise and model uncertainty and continuously optimize parameter estimation. The adaptive update process re-evaluates the structure and parameters of the dynamic heat capacity model, such as increasing the order of the dynamic heat capacity model or adjusting the parameter range, to ensure that it can accurately describe the current thermal dynamic characteristics of the photovoltaic module.

[0071] The method for establishing a dynamic heat capacity model according to the material parameters and structural parameters of the photovoltaic module includes:

[0072] A three-dimensional non-uniform grid model is constructed according to the material parameters and structural parameters of the photovoltaic module; the material parameters are the physical property parameters of the constituent materials of the photovoltaic module, and the structural parameters are the geometric characteristics and physical layout parameters of the photovoltaic module; the material parameters include physical properties such as the thermal conductivity, specific heat capacity, and density of each layer of material, and the structural parameters include geometric characteristics such as the thickness and area of each layer and the internal electrical connection method. For example, the three-dimensional non-uniform grid model is divided into 10×10 to 20×20 cells in the plane direction of the module and 4-6 layers in the thickness direction, corresponding to the glass layer, EVA layer, solar cell layer, and backplane layer of the module. The grid density of each layer is inversely proportional to the thermal conductivity of the material, and the grid density of the material layer with lower thermal conductivity is higher to improve the calculation accuracy of the temperature gradient.

[0073] Modify the degradation coefficients of material parameters according to the historical aging data of photovoltaic modules. The degradation coefficients include the annual growth rate of the thermal resistance of the encapsulation material and the attenuation curve of the heat capacity of the cell, and generate an initial thermal resistance-capacitance correlation matrix according to the modified material parameters; Modify the degradation coefficients of material parameters according to the historical aging data of photovoltaic modules. The degradation coefficients mainly include the annual growth rate of the thermal resistance of the encapsulation material and the attenuation curve of the heat capacity of the cell. The thermal resistance of the encapsulation material (such as EVA) increases with time. According to industry experience data, its annual growth rate is about 2-4%, and the total growth is about 15-25% after 5 years of use; The heat capacity of the cell decays with aging, and the decay rate is about 5-10% within 5 years. The degradation coefficients are obtained by analyzing the temperature response characteristics of the module at different operating durations and are updated regularly. Generate an initial thermal resistance-capacitance correlation matrix according to the modified material parameters to describe the heat transfer path and efficiency between different parts of the module.

[0074] Iteratively optimize the initial thermal resistance-capacitance correlation matrix and the backplane temperature gradient distribution data collected in real time by the three-dimensional non-uniform grid model to obtain dynamic heat capacity parameters; Generate a dynamic heat capacity model according to the optimized dynamic heat capacity parameters. The iterative optimization uses the gradient descent algorithm, and the optimization goal is to minimize the mean square error between the predicted temperature and the measured temperature of the model. The optimization process usually requires 10-20 iterations, and the convergence criterion is that the parameter change in two consecutive iterations is less than 1%. The optimized dynamic heat capacity parameters can accurately reflect the current thermal dynamic characteristics of the photovoltaic module.

[0075] The method of generating a dynamic heat capacity model according to the optimized dynamic heat capacity parameters includes:

[0076] Map the optimized dynamic heat capacity parameters to the three-dimensional non-uniform grid model according to the grid nodes, extract the thermal resistance and heat capacity values of each grid node and convert them into equivalent lumped parameters, and construct a dynamic heat capacity model according to the equivalent lumped parameters; Map the optimized dynamic heat capacity parameters to the three-dimensional non-uniform grid model according to the grid nodes, and extract the thermal resistance and heat capacity values of each grid node. The mapping process preserves the spatial distribution information to ensure that the model can accurately reflect the temperature gradient at different positions of the module. The extracted distributed thermal resistance and heat capacity values are converted into equivalent lumped parameters through the electrothermal equivalent circuit principle, which simplifies the calculation complexity while maintaining the dynamic characteristics. The equivalent lumped parameters include the thermal inertia factor (reflecting the temperature change rate) and the order and coefficients of the temperature-time transfer function. The thermal inertia factor is usually expressed as a time constant, which reflects the time required for the module to reach the temperature steady state; the order of the transfer function determines the complexity of the system response, usually 2-3 orders; the coefficients determine the weight of each order term and reflect the influence degree of thermal processes at different time scales.

[0077] Perform double verification of the dynamic heat capacity model in the frequency domain and the time domain. In the frequency domain, verify the amplitude-frequency characteristic and phase characteristic of the transfer function through swept-frequency excitation testing. In the time domain, verify the root mean square value of the residual between the predicted temperature and the actual temperature through a step temperature control experiment. If the verification results of the double verification in the frequency domain and the time domain meet the preset verification accuracy, bind the dynamic heat capacity model to the unique identification code of the photovoltaic module. The swept-frequency excitation testing is carried out in the frequency range of 0.001 - 0.1 Hz to measure the influence amplitude and phase difference of temperature changes at different frequencies on the power output. The verification standard requires that the relative error between the amplitude-frequency characteristic of the model and the measured data does not exceed 3% within the test frequency range, and the phase error does not exceed 10°. The step temperature control experiment applies a temperature step of 2 - 5 °C for 30 - 60 minutes and records the change process of temperature and power. The verification standard requires that the root mean square error of temperature prediction is less than 0.5 °C, and the relative root mean square error of power prediction is less than 2%. If the verification results of the double verification in the frequency domain and the time domain meet the preset verification accuracy, bind the dynamic heat capacity model to the unique identification code of the photovoltaic module and save it in the power station database as the dedicated model of this module.

[0078] The method of inputting the series-connected temperature-corrected power and the effective spectral irradiance into a pre-trained photovoltaic power prediction model and dynamically outputting the corrected power prediction value at the power station level in combination with the global irradiance data of numerical weather forecasting includes:

[0079] The cascade temperature correction power of each group is summarized into a power station-level temperature correction power matrix according to the string topology relationship and the power station electrical connection diagram. After performing time series alignment processing on the power station-level temperature correction power matrix, the effective spectral irradiance, and the global irradiance data of numerical weather forecasting, they are fused into a multi-dimensional input feature vector. The multi-dimensional input feature vector is input into a pre-trained photovoltaic power prediction model, which is trained by a hybrid neural network; the strings and inverters in a photovoltaic power station form a complex topology network. Integrating the string-level temperature correction power into the power station-level power requires considering the actual electrical connection relationship. The power station electrical connection diagram describes the parallel relationship between strings, the configuration and connection method of grid-connected inverters. Through the power station electrical connection diagram, the impact of the power change of any string on the total output of the power station can be calculated. The hierarchical aggregation method is adopted during the summarization process. The first layer is the component layer, where each photovoltaic component is processed through a temperature-power correction factor; the second layer is the string layer. Considering the series characteristics of the components within the string, the lowest temperature correction power within the string is used as a limiting factor; the third layer is the parallel group layer, where the strings connected in parallel are weighted and summed according to their respective rated capacity ratios; the last layer is the inverter layer, considering the efficiency curve of the inverter. The power station-level temperature correction power matrix is a multi-dimensional data structure, including a time dimension (prediction time point) and a spatial dimension (string position and topology relationship). Time series alignment processing is a key step to ensure the time consistency of multi-source data. For example, the time granularity of the power station-level temperature correction power matrix is 15 minutes, the numerical weather forecasting data is usually 1-hour granularity, and the spectral irradiance is 5-15 minutes granularity. A unified time benchmark converts data with different granularities into a unified 15-minute interval. Cubic spline interpolation is used for time refinement of coarser-grained data; time window averaging is used for aggregation of finer-grained data. The data fusion process uses feature engineering methods to construct a multi-dimensional input feature vector. The feature vector includes: the power prediction value (main feature) in the power station-level temperature correction power matrix; the effective spectral irradiance of each band and its time gradient (reflecting the spectral change trend); meteorological parameters such as the global irradiance, cloud cover, wind speed, and air temperature of numerical weather forecasting. All features are standardized to eliminate the dimension difference. The pre-trained photovoltaic power prediction model is trained by a hybrid neural network, including a hybrid structure of a temporal convolutional network (TCN) and a long short-term memory network (LSTM). The temporal convolutional network (TCN) is responsible for capturing short-term time patterns and is used to process the rapid changes in irradiance and temperature; the long short-term memory network (LSTM) captures long-term dependencies and is used for periodic patterns. The output of the photovoltaic power prediction model is the predicted values of the power station-level photovoltaic power at different future time points and their 95% confidence intervals.

[0080] If an extreme weather event warning is detected, multi-scale time window prediction is triggered to obtain the final power prediction value of the power station-level photovoltaic power; extreme weather events include weather conditions such as heavy rainfall, thunderstorms, sandstorms, and strong convections that may cause drastic fluctuations in photovoltaic output. The extreme weather event warning is derived from the real-time warning information issued by the meteorological department and is divided into four levels (Levels I-IV). The multi-scale time window prediction combines the advantages of short-term prediction and medium- and long-term prediction to improve the prediction accuracy under extreme conditions.

[0081] The power prediction error value is obtained by comparing the final power prediction value of the power station-level photovoltaic power with the real-time power monitoring data. If the power prediction errors in N consecutive time windows exceed the preset power error threshold, the pre-trained photovoltaic power prediction model is adaptively adjusted, where N≥3. The real-time power monitoring data comes from the power station SCADA system, which records the real-time output power of each inverter and busbar box in the power station. The power prediction error is calculated in two ways: relative error and absolute error. If the power prediction errors in N consecutive time windows exceed the preset power error threshold, such as the relative error exceeds ±5% or the absolute error exceeds 100 kW, the pre-trained photovoltaic power prediction model is triggered for adaptive adjustment. The adaptive adjustment uses the online incremental learning method, with the measured data in the recent 30 days as the training set, and incremental updates are performed every 7 days. The update learning rate is 10-30% of the basic learning rate, and at the same time, a forgetting factor (0.9-0.95) is introduced to reduce the weight impact of long-term data.

[0082] The method of triggering multi-scale time window prediction to obtain the final power prediction value of the power station-level photovoltaic power if an extreme weather event warning is detected includes:

[0083] The multi-scale time window prediction includes short-term prediction and medium- and long-term prediction; the short-term prediction generates high-frequency prediction results with a minute-level time window; the medium- and long-term prediction generates trend prediction results with an hour-level time window; according to the extreme weather event warning level and the duration of the weather event, the prediction results of the short-term prediction and the medium- and long-term prediction are dynamically weighted and fused to obtain the final power prediction value of the power station-level photovoltaic power. The short-term prediction generates high-frequency prediction results with a minute-level time window (5-15 minutes), mainly based on real-time monitoring data and short-term weather change trends, and is suitable for capturing the rapid fluctuations of photovoltaic output. The medium- and long-term prediction generates trend prediction results with an hour-level time window (1-24 hours), mainly based on numerical weather forecasts and historical similar day patterns, and is suitable for grasping the overall trend of photovoltaic output. The weighted fusion uses a time-varying weight method, with a higher weight for the short-term prediction at the recent time point and a higher weight for the medium- and long-term prediction at the far time point, to achieve a smooth transition from short-term accurate prediction to long-term trend prediction.

[0084] Finally, it should be noted that: Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A photovoltaic power prediction method, characterized in that The method includes: Obtaining in real time spectral response function data of different photovoltaic modules in a photovoltaic power station, where the different photovoltaic modules include at least two module types with spectral response characteristic differences greater than a preset difference threshold; Measuring in real time the spectral irradiance distribution incident on the surfaces of each photovoltaic module through a multi-spectral sensor network deployed above the photovoltaic array; calculating the effective spectral irradiance of each module type by grouping according to the module type based on the spectral response function data and the spectral irradiance distribution; Statistically analyzing the temperature coefficients of each batch to which the photovoltaic modules belong based on historical data, and establishing a module-level temperature-power correction factor; according to the string topology relationship, superimposing the temperature-power correction factors of all photovoltaic modules within the same string to obtain the string-level temperature-corrected power; Inputting the string-level temperature-corrected power and the effective spectral irradiance into a pre-trained photovoltaic power prediction model, and dynamically outputting a corrected power prediction value at the power station level in combination with the global irradiance data of numerical weather forecasting.

2. The photovoltaic power prediction method according to claim 1, wherein The method for calculating the effective spectral irradiance of each module type by grouping according to the module type based on the spectral response function data and the spectral irradiance distribution includes: Performing discretization processing on the spectral irradiance distribution according to a preset wavelength interval to obtain the irradiance value and wavelength range corresponding to each wavelength interval; Calculating the spectral response contribution value of each wavelength interval based on the response value of each wavelength interval in the spectral response function data of the module type and the measured irradiance value of the corresponding wavelength interval, and weighted summing the spectral response contribution values of all wavelength intervals to obtain the initial effective spectral irradiance of the module type; Performing normalization processing on the initial effective spectral irradiance according to the standard solar spectral distribution data under standard test conditions to obtain the effective spectral irradiance.

3. The photovoltaic power prediction method according to claim 1, wherein The method for statistically analyzing the temperature coefficients of each batch to which the photovoltaic modules belong based on historical data and establishing a module-level temperature-power correction factor includes: Associating the production batch information according to the unique identification code of the photovoltaic module, extracting the temperature coefficient data set of each production batch of photovoltaic modules in the historical database, and statistically analyzing the average value and dispersion parameter of the temperature coefficients of each production batch of photovoltaic modules; Collecting in real time the backplane temperature data of the photovoltaic module, and dynamically calculating the module-level temperature-power correction factor according to the average value and dispersion parameter of the temperature coefficients of its production batch; among them, for the temperature-power correction factor of the production batch of photovoltaic modules with a dispersion exceeding the preset dispersion threshold, a dispersion compensation weight is introduced, and the dispersion compensation weight has a negative correlation with the dispersion parameter; If the production batch to which the photovoltaic module belongs is a mixed batch and the temperature coefficient difference exceeds the preset temperature coefficient difference threshold, then adjust the weight of the temperature-power correction factor according to the electrical connection relationship of the photovoltaic modules of different production batches within the string. When it is detected that the temperature coefficient dispersion of the PV modules in a string continuously exceeds the historical statistical range, a stepped temperature control excitation signal is applied to the PV modules of this production batch, and the transient change slope of the temperature response curve is collected to invert the actual temperature coefficient, update the mean value and dispersion parameter of the temperature coefficient of the PV modules of this production batch, and regenerate the component-level temperature-power correction factor.

4. A photovoltaic power prediction method according to claim 3, characterized in that, The method of applying a stepped temperature control excitation signal to the PV modules of this production batch and collecting the transient change slope of the temperature response curve to invert the actual temperature coefficient when it is detected that the temperature coefficient dispersion of the PV modules in a string continuously exceeds the historical statistical range includes: Applying a stepped temperature control excitation signal through a temperature control actuator integrated on the backplane of the PV module, and synchronously collecting the transient change data of the backplane temperature of the PV module according to a preset sampling frequency; Performing filtering and noise reduction processing on the transient change data, extracting the response curves of each temperature step interval, calculating the instantaneous change slope and steady-state offset, and constructing a temperature-time transfer function in combination with the dynamic heat capacity model of the PV module; Matching the temperature-time transfer function with a preset standard temperature coefficient response database, and inversely calculating the actual temperature coefficient and calculating the confidence interval through least squares iterative optimization; if the half-width of the confidence interval is less than the preset precision threshold, output the actual temperature coefficient.

5. A photovoltaic power prediction method according to claim 4, characterized in that, The method of performing filtering and noise reduction processing on the transient change data, extracting the response curves of each temperature step interval, calculating the instantaneous change slope and steady-state offset, and constructing a temperature-time transfer function in combination with the dynamic heat capacity model of the PV module includes: Identifying the start point and end point of the temperature step interval through a dynamic threshold segmentation algorithm, fitting the response curve through cubic spline interpolation within each segmented temperature step interval, calculating the instantaneous change slope of each sampling point on the response curve, and extracting the maximum instantaneous change slope as the step response eigenvalue; Establishing a dynamic heat capacity model according to the material parameters and structural parameters of the PV module; inputting the step response eigenvalue and steady-state offset into the dynamic heat capacity model to obtain the parameter matrix of the temperature-time transfer function, and performing real-time calibration on the parameter matrix through the Kalman filter algorithm; if the calibration residual of the parameter matrix exceeds the preset residual tolerance threshold, trigger the adaptive update of the dynamic heat capacity model.

6. A photovoltaic power prediction method according to claim 5, characterized in that The method of establishing a dynamic heat capacity model according to the material parameters and structural parameters of the PV module includes: Constructing a three-dimensional non-uniform grid model according to the material parameters and structural parameters of the PV module; the material parameters are the physical property parameters of each constituent material of the PV module, and the structural parameters are the geometric characteristics and physical layout parameters of the PV module; Correcting the degradation coefficient of the material parameters according to the historical aging data of the PV module, the degradation coefficient includes the annual growth rate of the thermal resistance of the encapsulation material and the attenuation curve of the heat capacity of the battery chip, and generating an initial thermal resistance-heat capacity correlation matrix according to the corrected material parameters; Iteratively optimize the initial thermal resistance-capacitance correlation matrix with the backplane temperature gradient distribution data collected in real time by the three-dimensional non-uniform grid model to obtain dynamic capacitance parameters; map and generate a dynamic capacitance model according to the optimized dynamic capacitance parameters.

7. A photovoltaic power prediction method according to claim 5, characterized in that, The method for mapping and generating a dynamic capacitance model according to the optimized dynamic capacitance parameters includes: Map the optimized dynamic capacitance parameters to the three-dimensional non-uniform grid model according to grid nodes, extract the thermal resistance and capacitance values of each grid node and convert them into equivalent lumped parameters, and construct a dynamic capacitance model according to the equivalent lumped parameters; Perform double verification on the dynamic capacitance model in the frequency domain and the time domain. In the frequency domain, verify the amplitude-frequency characteristics and phase characteristics of the transfer function through a swept-frequency excitation test. In the time domain, verify the root mean square value of the residuals between the predicted temperature and the actual temperature through a step temperature control experiment. If the verification results of the double verification in the frequency domain and the time domain meet the preset verification accuracy, bind the dynamic capacitance model to the unique identification code of the photovoltaic module.

8. A photovoltaic power prediction method according to claim 1, characterized in that The method for inputting the string-level temperature-corrected power and the effective spectral irradiance into a pre-trained photovoltaic power prediction model and dynamically outputting a corrected plant-level photovoltaic power prediction value in combination with the global irradiance data of numerical weather forecasting includes: Summarize the string-level temperature-corrected powers into a plant-level temperature-corrected power matrix according to the string topology relationship and the plant electrical connection diagram. Align the time series of the plant-level temperature-corrected power matrix, the effective spectral irradiance, and the global irradiance data of numerical weather forecasting and then fuse them into a multi-dimensional input feature vector. Input the multi-dimensional input feature vector into the pre-trained photovoltaic power prediction model, and the photovoltaic power prediction model is obtained by training a hybrid neural network; If an extreme weather event warning is detected, trigger a multi-scale time window prediction to obtain a final plant-level photovoltaic power prediction value; Compare the final plant-level photovoltaic power prediction value with the real-time power monitoring data to obtain a power prediction error value. If the power prediction errors in N consecutive time windows exceed the preset power error threshold, adaptively adjust the pre-trained photovoltaic power prediction model, where N≥3.

9. A photovoltaic power prediction method according to claim 8, characterized in that, The method for triggering a multi-scale time window prediction to obtain a final plant-level photovoltaic power prediction value if an extreme weather event warning is detected includes: The multi-scale time window prediction includes short-term prediction and medium- and long-term prediction; the short-term prediction generates high-frequency prediction results with a minute-level time window; the medium- and long-term prediction generates trend prediction results with an hour-level time window; according to the extreme weather event warning level and the duration of the weather event, dynamically weight and fuse the prediction results of the short-term prediction and the medium- and long-term prediction to obtain a final plant-level photovoltaic power prediction value.

Images