A method for photovoltaic system power prediction

Abstract

The application relates to the field of photovoltaic technology and discloses a photovoltaic system power prediction method, which comprises the following steps: acquiring multi-source time sequence data of each photovoltaic component to determine first coupling coefficient time sequence data and second coupling coefficient time sequence data, combining the multi-source time sequence data to construct coupling characteristic vector time sequence data, inputting the coupling characteristic vector time sequence data into a preset power prediction model to obtain component predicted power time sequence data of each photovoltaic component within a preset future time length, and finally superimposing all the component predicted power time sequence data to obtain predicted total power time sequence data of the photovoltaic system within the preset future time length. The method considers multi-factor coupling and component individual differences, effectively solves the problems of low prediction accuracy, significant calculation error and power estimation under complex working conditions caused by insufficient consideration of multi-factor coupling and neglect of component individual differences in the prior art, improves the power prediction accuracy, reduces the calculation error and power estimation deviation under complex working conditions, and the like.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic technology, and in particular to a method for predicting the power of a photovoltaic system. Background Technology

[0002] Against the backdrop of accelerated global energy transition, photovoltaic (PV) power, as a core pillar of clean energy, has experienced explosive growth in installed capacity. However, the output power of PV systems is dynamically affected by various factors, exhibiting strong volatility and high uncertainty. This poses severe challenges to precise power scheduling, grid connection stability control, and maximizing power generation revenue. The core computational needs in the PV field focus on two main areas: dynamic power prediction and maximum power point (MPP) tracking. The accuracy and real-time performance of these calculations directly determine the system's power generation efficiency, grid connection security, and economic returns, making them a core focus of current PV technology research and development.

[0003] Existing photovoltaic (PV) system power prediction schemes suffer from two fundamental flaws: First, schemes based on statistical models (such as ARIMA and SVM) or basic machine learning models (such as LSTM) primarily rely on historical power generation data and average sunshine and temperature data provided by weather stations. Prediction is achieved by fitting data patterns (e.g., collecting average sunshine and temperature data from the past year and historical power generation data to train an SVM model for 24-hour power prediction the following day). Such schemes have low prediction accuracy and significant calculation errors under complex operating conditions. Second, schemes using simplified series-parallel models assume that all components within the same string channel have identical characteristics. These schemes exhibit significant power estimation biases. These flaws directly limit the achievement of goals such as improving PV system power generation efficiency, ensuring grid connection stability, and maximizing economic benefits. Summary of the Invention

[0004] Based on this, it is necessary to propose a photovoltaic system power prediction method to address the above problems. This method takes into account the coupling of multiple factors and the individual differences of components, effectively solving the problems of low prediction accuracy, significant calculation errors and power estimation under complex operating conditions caused by insufficient consideration of multi-factor coupling and neglect of individual component differences in the existing technology. This method improves the power prediction accuracy and reduces the calculation errors and power estimation deviations under complex operating conditions.

[0005] To achieve the above objectives, the present invention provides a photovoltaic system power prediction method in a first aspect, the method comprising: Obtain the multi-source time-series data of the nth photovoltaic module in the photovoltaic system, where the initial value of n is 1; Based on the irradiance time-series data, ambient temperature time-series data, and module attenuation coefficient time-series data in the multi-source time-series data of the nth photovoltaic module, determine the first coupling coefficient time-series data; The second coupling coefficient time series data is determined based on the component short-circuit current time series data and cloud shading ratio time series data in the multi-source time series data of the nth photovoltaic module. Based on the multi-source time-series data of the nth photovoltaic module, the time-series data of the first coupling coefficient, and the time-series data of the second coupling coefficient, a coupling feature vector time-series data is constructed. The time series data of the coupled feature vector is input into the preset power prediction model to obtain the time series data of the predicted power of the nth photovoltaic module within a preset future time period; Let n = n + 1, return to the step of obtaining the multi-source time series data of the nth photovoltaic module in the photovoltaic system, until n is greater than the total number of photovoltaic modules in the photovoltaic system. Based on the component predicted power time series data of all photovoltaic modules in the preset future time, determine the predicted total power time series data of the photovoltaic system in the preset future time.

[0006] Optionally, the timing data of the first coupling coefficient can be determined using the following formula: ; in, The first coupling coefficient is the first coupling coefficient at time step t in the time series data of the first coupling coefficient. The component attenuation coefficient is the component attenuation coefficient at time step t in the component attenuation coefficient time series data. The light intensity at time step t in the light intensity time series data is... The ambient temperature at time step t in the ambient temperature time series data is given.

[0007] Optionally, determining the second coupling coefficient time series data based on the module short-circuit current time series data and cloud shading ratio time series data from the multi-source time series data of the nth photovoltaic module includes: Determine the maximum, minimum, and average component short-circuit current in the component short-circuit current timing data; The component mismatch is determined based on the maximum component short-circuit current, the minimum component short-circuit current, and the average component short-circuit current. The second coupling coefficient time series data is determined based on the component mismatch degree and the cloud occlusion ratio time series data.

[0008] Optionally, the timing data of the component mismatch and the second coupling coefficient are determined using the following formula: ; in, The component mismatch degree, The maximum component short-circuit current, The short-circuit current of the minimum component. The average component short-circuit current. The second coupling coefficient is the second coupling coefficient at time step t in the time series data. The cloud occlusion percentage is the cloud occlusion percentage at time step t in the time series data of cloud occlusion percentage.

[0009] Optionally, the method further includes: The system acquires the real-time output power of all photovoltaic modules in the k-th string channel of the m-th photovoltaic inverter in the photovoltaic system, and acquires the module series-parallel topology information of the photovoltaic system, where the initial values ​​of m and k are both 1. Based on the real-time output power of all photovoltaic modules in the kth string channel of the mth photovoltaic inverter and the serial-parallel topology information of the modules, determine the real-time total output power of the kth string channel of the mth photovoltaic inverter. Based on the time-series data of the predicted power of all photovoltaic modules within a preset future time period, and the information on the series and parallel topology of the modules, the predicted total output power of the kth string channel of the mth photovoltaic inverter is determined, wherein the predicted total output power of the kth string channel of the mth photovoltaic inverter and the real-time total output power are the power at the same time step. Determine the real-time relative error between the real-time total output power and the predicted total output power of the k-th string channel of the m-th photovoltaic inverter; If the average value of the real-time relative error over a number of consecutive preset sampling periods is greater than a preset percentage, determine the real-time absolute error between the real-time total output power of the k-th string channel of the m-th photovoltaic inverter and the predicted total output power. Obtain the current maximum power point step size coefficient of the photovoltaic system; Based on the real-time absolute error, the current maximum power point step size coefficient of the photovoltaic system is corrected to obtain the corrected current maximum power point step size coefficient. The conductivity increment is determined based on the corrected current maximum power point step size coefficient, and the real-time total output power and real-time total output voltage of the kth string channel of the mth photovoltaic inverter. When the conductance increment is not equal to 0, the real-time total output voltage of the kth string channel of the mth photovoltaic inverter is adjusted according to the conductance increment and the corrected current maximum power point step size coefficient. Let k = k + 1, return to the step of obtaining the real-time output power of all photovoltaic modules in the kth string channel of the mth photovoltaic inverter in the photovoltaic system, until k is greater than the total number of string channels of the mth photovoltaic inverter, and adjust the real-time total output voltage of all string channels of the mth photovoltaic inverter. Let m = m + 1, then return to the step of obtaining the real-time output power of all photovoltaic modules in the k-th string channel of the m-th photovoltaic inverter in the photovoltaic system, until m is greater than the total number of photovoltaic inverters in the photovoltaic system, and adjust the real-time total output voltage of all string channels of all photovoltaic inverters.

[0010] Optionally, obtaining the current maximum power point step size coefficient of the photovoltaic system includes: Real-time acquisition of the real-time rate of change of light intensity and the real-time shadow movement speed of the photovoltaic system; The step size coefficient of the current maximum power point is determined based on the real-time light intensity change rate and the real-time shadow movement speed.

[0011] Optionally, adjusting the real-time total output voltage of the k-th string channel of the m-th photovoltaic inverter based on the conductance increment and the corrected current maximum power point step size coefficient includes: Obtain the reference values ​​of the open-circuit voltage of all photovoltaic modules in the kth string channel of the mth photovoltaic inverter; Based on the reference values ​​of the open-circuit voltage of all photovoltaic modules in the kth string channel of the mth photovoltaic inverter, and the information on the series-parallel topology of the modules, the reference value of the total open-circuit voltage of the modules in the kth string channel of the mth photovoltaic inverter is determined. The voltage change is determined based on the reference value of the total open-circuit voltage of the module in the kth string channel of the mth photovoltaic inverter and the corrected current maximum power point step size coefficient. The real-time total output voltage of the k-th string channel of the m-th photovoltaic inverter is adjusted based on the voltage change and the conductance increment.

[0012] Optionally, adjusting the real-time total output voltage of the k-th string channel of the m-th photovoltaic inverter based on the voltage change and the conductance increment includes: When the conductivity increment is greater than 0, the real-time total output voltage of the kth string channel of the mth photovoltaic inverter is increased according to the voltage change. When the conductivity increment is less than 0, the real-time total output voltage of the kth string channel of the mth photovoltaic inverter is reduced according to the voltage change.

[0013] Optionally, the method further includes: Based on the real-time absolute error, the attention layer parameters of the multi-head attention layer in the preset power prediction model are corrected to obtain the corrected preset power prediction model. The modified preset power prediction model is used as the preset power prediction model.

[0014] Optionally, the preset power prediction model includes an input layer, a hidden layer, a multi-head attention layer, and an output layer connected in sequence. The hidden layer includes two stacked LSTM layers, and the output layer is a fully connected layer. The input layer is used to receive the temporal data of the coupled feature vector; The hidden layer is used to perform time-series capture processing on the coupled feature vector time-series data to obtain first feature vector time-series data and second feature vector time-series data, and then concatenates the first feature vector time-series data and the second feature vector time-series data to obtain deep feature vector time-series data; The multi-head attention layer is used to determine the feature weight time-series data of each attention head based on the deep feature vector time-series data, and to perform weight allocation processing on the deep feature vector time-series data based on the feature weight time-series data of each attention head to obtain the target feature vector time-series data of each attention head, and to perform feature fusion processing on the target feature vector time-series data of each attention head to obtain fused feature vector time-series data. The output layer is used to perform linear transformation on the time series data of the fused feature vector to obtain the time series data of the predicted power of the nth photovoltaic module within the preset future time period.

[0015] To achieve the above objectives, the present invention provides a photovoltaic system power prediction device in a second aspect, the device comprising: The acquisition module is used to acquire the multi-source time-series data of the nth photovoltaic module in the photovoltaic system, where the initial value of n is 1; The first determining module is used to determine the first coupling coefficient time series data based on the irradiance time series data, ambient temperature time series data and module attenuation coefficient time series data in the multi-source time series data of the nth photovoltaic module; The second determining module is used to determine the second coupling coefficient time series data based on the component short-circuit current time series data and cloud shading ratio time series data in the multi-source time series data of the nth photovoltaic module. The construction module is used to construct coupling feature vector time series data based on the multi-source time series data of the nth photovoltaic module, the first coupling coefficient time series data, and the second coupling coefficient time series data; The model prediction module is used to input the coupled feature vector time series data into the preset power prediction model to obtain the module prediction power time series data of the nth photovoltaic module within a preset future time period; The iterative aggregation module is used to set n=n+1 and return to the step of obtaining the multi-source time series data of the nth photovoltaic module in the photovoltaic system until n is greater than the total number of photovoltaic modules in the photovoltaic system. Based on the component predicted power time series data of all photovoltaic modules in the preset future time, the predicted total power time series data of the photovoltaic system in the preset future time is determined.

[0016] To achieve the above objectives, the present invention provides, in a third aspect, a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the photovoltaic system power prediction method as described in any one of the first aspects.

[0017] To achieve the above objectives, the present invention provides a computer device in a fourth aspect, including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the photovoltaic system power prediction method as described in any one of the first aspects.

[0018] The present invention has the following beneficial effects: The above method obtains multi-source time-series data of the nth photovoltaic module in a photovoltaic system, where the initial value of n is 1. Then, based on the irradiance time-series data, ambient temperature time-series data, and module attenuation coefficient time-series data in the multi-source time-series data of the nth photovoltaic module, it determines the first coupling coefficient time-series data. Based on the module short-circuit current time-series data and cloud shading ratio time-series data in the multi-source time-series data of the nth photovoltaic module, it determines the second coupling coefficient time-series data. Then, based on the multi-source time-series data of the nth photovoltaic module, the first coupling coefficient time-series data, and the second coupling coefficient time-series data, it constructs coupling feature vector time-series data. The coupling feature vector time-series data is then input into a preset power prediction model to obtain the module predicted power time-series data of the nth photovoltaic module within a preset future time period. Finally, it lets... If n = n + 1, return to the step of obtaining the multi-source time-series data of the nth photovoltaic module in the photovoltaic system, until n is greater than the total number of photovoltaic modules in the photovoltaic system. Based on the predicted power time-series data of all photovoltaic modules within a preset future time, determine the predicted total power time-series data of the photovoltaic system within a preset future time. That is, by constructing coupled feature vector time-series data and power prediction for each photovoltaic module, considering multi-factor coupling and individual module differences, it effectively solves the problems of low prediction accuracy, significant calculation errors and power estimation under complex operating conditions caused by insufficient consideration of multi-factor coupling and neglect of individual module differences in the existing technology. It improves the power prediction accuracy, reduces the calculation error and power estimation deviation under complex operating conditions, and ultimately achieves the comprehensive technical effect of improving the power generation efficiency of the photovoltaic system, enhancing grid connection stability and maximizing economic benefits. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] in: Figure 1 This is a schematic diagram of a photovoltaic system power prediction method according to an embodiment of this application; Figure 2 This is a schematic diagram of a photovoltaic system power prediction device in an embodiment of this application; Figure 3 This is a diagram showing the internal structure of a computer device in some embodiments. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Against the backdrop of accelerated global energy transition, photovoltaic (PV) power, as a core pillar of clean energy, has experienced explosive growth in installed capacity. However, the output power of PV systems is dynamically affected by various factors, exhibiting strong volatility and high uncertainty. This poses severe challenges to precise power scheduling, grid connection stability control, and maximizing power generation revenue. The core computational needs in the PV field focus on two main areas: dynamic power prediction and maximum power point (MPP) tracking. The accuracy and real-time performance of these calculations directly determine the system's power generation efficiency, grid connection security, and economic returns, making them a core focus of current PV technology research and development.

[0023] Existing photovoltaic (PV) system power prediction schemes suffer from two fundamental flaws: First, schemes based on statistical models (such as ARIMA and SVM) or basic machine learning models (such as LSTM) primarily rely on historical power generation data and average sunshine and temperature data provided by weather stations. Prediction is achieved by fitting data patterns (e.g., collecting average sunshine and temperature data from the past year and historical power generation data to train an SVM model for 24-hour power prediction the following day). Such schemes have low prediction accuracy and significant calculation errors under complex operating conditions. Second, schemes using simplified series-parallel models assume that all components within the same string channel have identical characteristics. These schemes exhibit significant power estimation biases. These flaws directly limit the achievement of goals such as improving PV system power generation efficiency, ensuring grid connection stability, and maximizing economic benefits.

[0024] To address the aforementioned issues, this application proposes a photovoltaic system power prediction method that considers multi-factor coupling and individual component differences. This method effectively solves the problems of low prediction accuracy, significant calculation errors and power estimation under complex operating conditions caused by insufficient consideration of multi-factor coupling and neglect of individual component differences in the prior art. It improves power prediction accuracy and reduces calculation errors and power estimation deviations under complex operating conditions. The specific implementation principle will be described in detail in the following embodiments.

[0025] This application provides a method for predicting the power of a photovoltaic system in its first aspect.

[0026] Please see Figure 1 This is a schematic diagram of a photovoltaic system power prediction method according to an embodiment of this application. The method includes: Step 110: Obtain the multi-source timing data of the nth photovoltaic module in the photovoltaic system, where the initial value of n is 1.

[0027] Regarding the selection of multi-source time-series data types, in some embodiments, multi-source time-series data includes environmental dynamic time-series data, component characteristic time-series data, and system runtime time-series data. Among them, environmental dynamic time-series data includes irradiance time-series data, ambient temperature time-series data, wind speed time-series data, cloud cover ratio time-series data, and shadow movement speed time-series data; component characteristic time-series data includes component short-circuit current time-series data, component open-circuit voltage time-series data, component fill factor time-series data, and component attenuation coefficient time-series data; and system runtime time-series data includes photovoltaic inverter output voltage time-series data and photovoltaic inverter output current time-series data.

[0028] Regarding the acquisition method of multi-source time series data, in some embodiments, the multi-source raw time series data of the nth photovoltaic module can be acquired, and then each dimension of the raw time series data in the multi-source raw time series data of the nth photovoltaic module is subjected to data cleaning and standardization processing to obtain the multi-source time series data of the nth photovoltaic module. The data cleaning includes outlier removal and missing value interpolation.

[0029] For outlier removal, missing value interpolation, and standardization, in some embodiments, a 3-standard-deviation removal principle can be used to remove outliers in each dimension of the multi-source original time series data, and a linear interpolation method can be used to complete the missing data in each dimension of the multi-source original time series data, and a normalization method can be used to standardize the data in each dimension of the multi-source original time series data to the [0, 1] interval.

[0030] Step 120: Determine the first coupling coefficient time series data based on the irradiance time series data, ambient temperature time series data, and module attenuation coefficient time series data in the multi-source time series data of the nth photovoltaic module.

[0031] It should be noted that, considering that both strong light and high temperature will accelerate the degradation of photovoltaic modules, this application quantifies the coupling effect between these three factors based on time-series data of light intensity, time-series data of ambient temperature, and time-series data of module degradation coefficient, namely, the determined first coupling coefficient time-series data.

[0032] Step 130: Determine the second coupling coefficient time series data based on the module short-circuit current time series data and cloud shading ratio time series data in the multi-source time series data of the nth photovoltaic module.

[0033] It should be noted that, considering the differences in production errors and aging rates among various photovoltaic modules, i.e., the influence of module shape on short-circuit current, this application quantifies the coupling effect between these two factors based on the time series data of module short-circuit current and the time series data of cloud shading ratio, i.e., the determined second coupling coefficient time series data.

[0034] Step 140: Construct coupling feature vector time series data based on the multi-source time series data of the nth photovoltaic module, the first coupling coefficient time series data, and the second coupling coefficient time series data.

[0035] In some embodiments, the time series data of coupled feature vectors can be constructed by sequentially concatenating the following data from the multi-source time series data of the nth photovoltaic module: irradiance time series data, ambient temperature time series data, wind speed time series data, cloud cover ratio time series data, shadow movement speed time series data, module short-circuit current time series data, module open-circuit voltage time series data, module fill factor time series data, photovoltaic inverter output voltage time series data, photovoltaic inverter output current time series data, as well as the first coupling coefficient time series data and the second coupling coefficient time series data, to obtain the coupled feature vector time series data.

[0036] Step 150: Input the coupled feature vector time series data into the preset power prediction model to obtain the module prediction power time series data of the nth photovoltaic module within the preset future time.

[0037] The preset future duration can be obtained and set in advance by the operator based on a large amount of experience, experiments or statistics. Of course, it can also be set in advance by the operator according to actual needs. The preset power prediction model here can be a model obtained by the operator in advance, which is used to predict the predicted power time series data of the nth photovoltaic module within the preset future duration based on the input coupling feature vector time series data.

[0038] Regarding the value of the preset future duration, in some embodiments, this application preferably sets the preset future duration to a short period of 15 minutes to 2 hours and / or an ultra-short period of 1 minute to 15 minutes.

[0039] In some embodiments, the training method for the preset power prediction model can be achieved by acquiring the historical coupling feature vector time-series data of each photovoltaic module and the historical module power time-series data within a preset time period. Then, the historical coupling feature vector time-series data of each photovoltaic module and the historical module power time-series data within the preset time period are sequentially input into the initial power prediction model for training. Once the training reaches a certain level, the preset power prediction model can be obtained. The method for determining when the model can be applied after training can be determined by setting a loss function. The preset time period is equal to the preset future time period.

[0040] Furthermore, in addition to training only one preset power prediction model for all photovoltaic modules, in some other embodiments, a preset power prediction model can be trained for each photovoltaic module to achieve more accurate power prediction.

[0041] Regarding the model type of the initial power prediction model, in some embodiments, the operator may choose an existing machine learning model or a deep learning network model according to actual needs, and no limitation is made here.

[0042] Step 160: Let n = n + 1, return to the step of obtaining the multi-source time series data of the nth photovoltaic module in the photovoltaic system, until n is greater than the total number of photovoltaic modules in the photovoltaic system. Based on the component predicted power time series data of all photovoltaic modules in the preset future time, determine the predicted total power time series data of the photovoltaic system in the preset future time.

[0043] In some embodiments, the predicted total power time series data can be determined by obtaining the series-parallel topology information of the photovoltaic system's components, and then determining the predicted total power time series data of the photovoltaic system within the preset future time period based on the series-parallel topology information and the predicted power time series data of all photovoltaic components within the preset future time period.

[0044] Furthermore, in some embodiments, based on the component series-parallel topology information, the predicted power time series data of all photovoltaic modules within a preset future time period can be aggregated at the same time step according to the series-parallel relationship to obtain the predicted total power time series data of the photovoltaic system within a preset future time period.

[0045] In this embodiment, by constructing coupled feature vector time-series data and power prediction for each photovoltaic module, multi-factor coupling and individual module differences are considered. This effectively solves the problems of low prediction accuracy, significant calculation errors and power estimation under complex operating conditions caused by insufficient consideration of multi-factor coupling and neglect of individual module differences in the prior art. It improves power prediction accuracy, reduces calculation errors and power estimation deviations under complex operating conditions, and ultimately achieves the comprehensive technical effect of improving photovoltaic system power generation efficiency, enhancing grid connection stability and maximizing economic benefits.

[0046] In addition to the aforementioned benefits such as improving power prediction accuracy, reducing calculation errors and power estimation deviations, increasing photovoltaic system power generation efficiency, enhancing grid connection stability, and maximizing economic benefits, this photovoltaic system power prediction method also has the following advantages: Improved data quality: When acquiring multi-source time-series data, the original multi-source time-series data is first cleaned and standardized. Outlier removal and missing value interpolation ensure the integrity and accuracy of the data, avoiding interference from outlier and missing data in subsequent analysis. This provides a reliable foundation for constructing accurate coupled feature vector time-series data. Furthermore, the normalization method standardizes the data to […]. Within the interval [0, 1], the influence of dimensions between different data dimensions is eliminated, enabling different types of data to be compared and analyzed on the same scale, further improving data quality and usability; Comprehensive data utilization: The selected multi-source time series data covers information from multiple aspects such as environmental dynamics, component characteristics, and system operation, comprehensively considering various factors affecting photovoltaic module power. This comprehensive data utilization method can more accurately reflect the actual operating status of the photovoltaic system and provide rich information support for power prediction; Enhanced model adaptability: By determining the time series data of the first coupling coefficient and the second coupling coefficient, the different factors are quantified. This approach addresses the coupling effects between photovoltaic (PV) modules and incorporates these coupling coefficients into the time-series data of the coupling feature vector. This allows the pre-set power prediction model to better adapt to complex and changing operating conditions because the model considers not only the influence of individual factors but also the interactions between them, thereby improving the model's predictive ability under different environmental conditions. Furthermore, a pre-set power prediction model can be trained for each PV module, fully considering individual module differences. Different PV modules may have different performance due to factors such as production errors and aging. Training a model separately for each PV module can more accurately capture its characteristics, further improving the accuracy of power prediction. The model training process is optimized: when training the pre-set power prediction model, a loss function is set to judge the training level, allowing for timely understanding of the model's training effect. The training strategy can be adjusted based on changes in the loss function to ensure the model is trained to its optimal state, improving model performance and generalization ability. Finally, this method facilitates system maintenance and management: it can obtain detailed information for each PV module, including multi-source time-series data and predicted power time-series data. Analysis of this data can promptly identify abnormalities in PV modules, such as performance degradation and faults, facilitating targeted maintenance and management by operation and maintenance personnel, extending the service life of PV modules, and reducing operation and maintenance costs.Supporting Intelligent Decision-Making: Accurate time-series data of total power prediction for photovoltaic (PV) systems can provide crucial information for grid dispatching and energy management. Grid dispatching departments can use this data to rationally schedule power generation, optimize grid operation, and improve energy efficiency. Energy management departments can use this data for energy storage and allocation, achieving rational energy distribution and utilization, and supporting the development of smart energy systems. Promoting the Development of the PV Industry: This method improves the accuracy and reliability of PV system power prediction, helping to reduce the uncertainty of PV power generation, enhancing investor confidence in PV projects, and promoting investment and development in the PV industry. Simultaneously, accurate power prediction also helps improve the competitiveness of PV systems in the electricity market and promotes the integration of the PV industry with other energy industries.

[0047] In one feasible implementation, the timing data of the first coupling coefficient is determined using the following formula: ; in, The first coupling coefficient is the first coupling coefficient at time step t in the time series data. This refers to the component attenuation coefficient at time step t in the component attenuation coefficient time series data. Let represent the light intensity at time step t in the time series data of light intensity. Let t be the ambient temperature at time step t in the ambient temperature time series data.

[0048] In this embodiment of the application, a precise calculation method is provided for quantifying the coupling effect of illumination, temperature and component attenuation through mathematical formulas, thereby improving the accuracy of the timing data of the first coupling coefficient and assisting in power prediction.

[0049] Understandably, this formula precisely quantifies the coupling relationship: it explicitly provides the calculation method for the first coupling coefficient at each time step in the time-series data, incorporating the three key factors of module degradation coefficient, light intensity, and ambient temperature into a unified calculation framework. Through this precise mathematical expression, it can accurately quantify the complex coupling effect of strong light and high temperature accelerating photovoltaic module degradation, avoiding errors caused by relying solely on experience or simple estimations, and providing more reliable basic data for subsequent power prediction. It also adapts to complex operating conditions: in actual photovoltaic system operation, light intensity and ambient temperature constantly change with time and weather conditions, and the degree of impact on module degradation varies at different times. This formula considers the specific data at each time step, enabling… Real-time reflection of these dynamic changes makes the calculated first coupling coefficient time-series data more closely reflect actual conditions. Whether it is a sunny and hot daytime or a cloudy and cold period, the first coupling coefficient can be accurately calculated, thereby improving the adaptability and accuracy of the power prediction model under various complex operating conditions. Enhanced model interpretability: The clear formula makes the calculation process of the first coupling coefficient transparent, which facilitates the understanding and analysis of the contribution of different factors to the coupling effect. Researchers and maintenance personnel can intuitively understand the interaction between light, temperature and component degradation based on the parameters in the formula, providing a theoretical basis for further optimizing photovoltaic system design and improving operation and maintenance strategies, and helping to improve the performance and reliability of the entire photovoltaic system.

[0050] In one feasible implementation, step 130 in the above embodiment, determining the second coupling coefficient time series data based on the module short-circuit current time series data and the cloud shading ratio time series data in the multi-source time series data of the nth photovoltaic module, includes: determining the maximum module short-circuit current, minimum module short-circuit current, and average module short-circuit current in the module short-circuit current time series data; determining the module mismatch degree based on the maximum module short-circuit current, minimum module short-circuit current, and average module short-circuit current; and determining the second coupling coefficient time series data based on the module mismatch degree and the cloud shading ratio time series data.

[0051] In this embodiment, the maximum, minimum and average short-circuit currents of the components are extracted from the short-circuit current time-series data to construct the component mismatch degree, accurately quantify the coupling effect between component mismatch and cloud shading, and improve the accuracy of power prediction under complex operating conditions.

[0052] Understandably, precise quantification of module mismatch—by extracting the maximum, minimum, and average short-circuit current from the module short-circuit current time-series data, a module mismatch degree is constructed. This objectively reflects the performance inconsistencies of photovoltaic modules caused by production errors, aging differences, etc., avoiding the bias caused by the simplified model assuming that all modules have completely identical characteristics; Coupling the impact of cloud shading—by combining the module mismatch degree with the cloud shading ratio time-series data to calculate the second coupling coefficient time-series data, both the inherent differences of the modules and the dynamic capture of the impact of instantaneous changes in illumination caused by cloud movement on the short-circuit current, making the second coupling coefficient... The data can reflect the combined effects of component mismatch and cloud shading in real time; improve adaptability to complex operating conditions: this calculation method is particularly suitable for complex scenarios such as cloudy weather and local shading. By quantifying the interaction between component mismatch and cloud shading, it effectively solves the problem of significant prediction errors in traditional methods under varying operating conditions, providing more discriminative feature inputs for preset power prediction models; enhance model interpretability: the explicit calculation process of component mismatch makes the physical meaning of the second coupling coefficient clearer. Operation and maintenance personnel can locate abnormal components by analyzing the trend of mismatch changes, providing a basis for decision-making for photovoltaic array optimization configuration and targeted maintenance.

[0053] In one feasible implementation, the timing data of the component mismatch and the second coupling coefficient are determined using a formula: ; in, For component mismatch, The maximum component short-circuit current, For the minimum component short-circuit current, This represents the average component short-circuit current. The second coupling coefficient is the second coupling coefficient at time step t in the time series data. This represents the cloud cover percentage at time step t in the time series data of cloud cover percentage.

[0054] In this embodiment, the coupling relationship between component mismatch and cloud occlusion is accurately quantified by mathematical formulas, thereby improving the accuracy of power prediction under complex operating conditions.

[0055] Understandably, precise quantification of module mismatch—by explicitly calculating the maximum, minimum, and average module short-circuit currents—objectively reflects the performance inconsistencies of photovoltaic modules caused by production errors and aging differences. This avoids the biases arising from the simplified model's assumption that all modules have completely identical characteristics, providing a more accurate basis for the second coupling coefficient. Dynamic coupling with cloud shading—combining module mismatch with time-series data on cloud shading percentages—allows for real-time calculation of the second coupling coefficient using a formula. This considers both inherent module differences and dynamically captures the impact of instantaneous changes in illumination caused by cloud movement on the short-circuit current, enabling the second coupling coefficient to accurately reflect both module mismatch and cloud shading. Composite effects; Enhanced adaptability to complex operating conditions: This formula is particularly suitable for complex scenarios such as cloudy weather and localized shading. By quantifying the interaction between component mismatch and cloud cover, it effectively solves the problem of significant prediction errors in traditional methods under varying operating conditions. It provides more discriminative feature inputs for the preset power prediction model, improving the model's prediction capability under complex operating conditions; Improved model interpretability: The explicit calculation process of component mismatch and the second coupling coefficient makes the physical meaning of the preset power prediction model clearer. Maintenance personnel can analyze the trend of mismatch changes to locate components with abnormal performance, providing a basis for decision-making for photovoltaic array optimization and targeted maintenance, thus enhancing the model's interpretability and practicality.

[0056] In one feasible implementation, the method in the above embodiments further includes: acquiring in real time the real-time output power of all photovoltaic modules in the k-th string channel of the m-th photovoltaic inverter in the photovoltaic system, and acquiring the module series-parallel topology information of the photovoltaic system, wherein the initial values ​​of m and k are both 1; determining the real-time total output power of the k-th string channel of the m-th photovoltaic inverter based on the real-time output power of all photovoltaic modules in the k-th string channel of the m-th photovoltaic inverter and the module series-parallel topology information; and based on the module predicted power time-series data of all photovoltaic modules within a preset future time period, using... Based on the component series-parallel topology information, the predicted total output power of the k-th string channel of the m-th photovoltaic inverter is determined, where the predicted total output power of the k-th string channel of the m-th photovoltaic inverter and the real-time total output power are the power at the same time step; the real-time relative error between the real-time total output power and the predicted total output power of the k-th string channel of the m-th photovoltaic inverter is determined; if the average value of the real-time relative error over a consecutive preset number of preset sampling periods is greater than a preset percentage, the difference between the real-time total output power and the predicted total output power of the k-th string channel of the m-th photovoltaic inverter is determined. The system calculates the real-time absolute error; obtains the current maximum power point step size coefficient of the photovoltaic system; corrects the current maximum power point step size coefficient based on the real-time absolute error; determines the conductance increment based on the corrected current maximum power point step size coefficient, the real-time total output power and real-time total output voltage of the kth string channel of the m-th photovoltaic inverter; and adjusts the real-time total output voltage of the kth string channel of the m-th photovoltaic inverter based on the conductance increment and the corrected current maximum power point step size coefficient when the conductance increment is not equal to 0. Let k = k + 1, return to the step of obtaining the real-time output power of all photovoltaic modules in the k-th string channel of the m-th photovoltaic inverter in the photovoltaic system, until k is greater than the total number of string channels of the m-th photovoltaic inverter, and adjust the real-time total output voltage of all string channels of the m-th photovoltaic inverter; let m = m + 1, return to the step of obtaining the real-time output power of all photovoltaic modules in the k-th string channel of the m-th photovoltaic inverter in the photovoltaic system, until m is greater than the total number of photovoltaic inverters in the photovoltaic system, and adjust the real-time total output voltage of all string channels of all photovoltaic inverters.

[0057] The preset values, preset sampling periods, and preset percentages can all be preset by the operator based on extensive experience, experiments, or statistics. Alternatively, they can be preset by the operator according to actual needs. Regarding the values ​​of the preset value, preset sampling period, and preset percentage, in some embodiments, this application preferably sets the preset value to 5, the preset sampling period to 1 second, and the preset percentage to 5%.

[0058] Regarding the determination of real-time total output power and predicted total output power, in some embodiments, based on the component series-parallel topology information, according to the series-parallel relationship, the real-time output power of all photovoltaic modules in the kth string channel of the m-th photovoltaic inverter is summarized at the same time step to obtain the real-time total output power of the kth string channel of the m-th photovoltaic inverter. Additionally, the predicted power time-series data of all photovoltaic modules located in the kth string channel within a preset future time period are summarized at the same time step to obtain the predicted total output power of the kth string channel of the m-th photovoltaic inverter.

[0059] In some embodiments, the method for determining the corrected current maximum power point step size factor can be achieved using a formula. Determine the corrected step size coefficient for the current maximum power point; where, This is the corrected step size factor for the current maximum power point. This is the step size factor for the current maximum power point. To preset the learning rate, For real-time absolute error, It is the partial derivative of the real-time absolute error with respect to the step size coefficient of the current maximum power point.

[0060] Regarding the value of the preset learning rate, in some embodiments, this application preferably sets the preset learning rate to 0.001.

[0061] In some embodiments, the determination of the conductance increment can be achieved using a formula. Determine the conductance increment; where, For the increase in conductivity, This is the corrected step size factor for the current maximum power point. Let be the real-time total output power of the k-th string channel of the m-th photovoltaic inverter. Let be the real-time total output voltage of the k-th string channel of the m-th photovoltaic inverter. Let be the derivative of the real-time total output power of the k-th string channel of the m-th photovoltaic inverter with respect to the real-time total output voltage.

[0062] It should be noted that adjusting the real-time total output voltage of all string channels of the m-th photovoltaic inverter will cause the real-time total output power of all string channels of the m-th photovoltaic inverter to converge to the current maximum power point. In turn, adjusting the real-time total output voltage of all string channels of all photovoltaic inverters will cause the real-time total output power of all string channels of the m-th photovoltaic inverter to converge to the current maximum power point.

[0063] In other embodiments, when the conductance increment is equal to 0, it indicates that the real-time total output power of the k-th string channel of the m-th photovoltaic inverter has converged to the current maximum power point, and therefore, no further steps need to be performed.

[0064] It is worth noting that the mainstream traditional maximum power point (MPP) tracking calculation schemes are the perturbation-observation method and the incremental conductance method. The perturbation-observation method determines the adjustment direction by periodically fine-tuning the output voltage or current of the photovoltaic module and comparing the power change trend. The incremental conductance method is based on the principle that the conductance increment at the maximum power point is equal to the negative conductance, and determines the adjustment strategy by calculating the conductance change rate. Both of these traditional MPP tracking calculation schemes use a fixed maximum power point step size for adjustment, which results in power oscillations in steady state, leading to tracking lag under dynamic conditions. This application optimizes the maximum power point step size coefficient through dynamic response, shortening the calculation response time when there are sudden changes in operating conditions such as illumination and shading. This solves the problem of power oscillations in steady state and tracking lag under dynamic conditions caused by traditional MPP tracking calculation schemes, ensuring real-time performance to meet the requirements of grid-connected scheduling and efficiency optimization.

[0065] In the embodiments of this application, precise and efficient maximum power point tracking is achieved by dynamically correcting the MPP step size coefficient, which significantly improves the system's power generation efficiency and stability.

[0066] Understandably, dynamic response optimization involves calculating the relative error between the real-time total output power and the predicted total output power. When the error exceeds 5% for five consecutive sampling periods (1-second intervals), a maximum power point step size correction is triggered. This allows the photovoltaic system to quickly respond to sudden changes in illumination (such as cloud movement) or shading changes, minimizing the tracking delay of the traditional fixed step size method and significantly improving power capture efficiency under dynamic conditions. Steady-state oscillation suppression utilizes a gradient descent method to dynamically adjust the maximum power point step size. Adjustment automatically terminates when the conductance increment is zero, effectively eliminating the traditional... The MPP tracking calculation scheme reduces power oscillations in steady state, improving power generation stability; it is adaptable to all operating conditions: by combining component-level power prediction time-series data with component series-parallel topology information, it can accurately locate the string channels that need adjustment, and maintain high MPP tracking accuracy even in locally shaded scenarios; it has a closed-loop self-optimization mechanism: by correcting the maximum power point step size coefficient in reverse through real-time absolute error, it forms a closed-loop control of prediction, tracking, and correction, which keeps the power prediction error growth within a very small range during long-term system operation, significantly outperforming the error drift of traditional MPP tracking calculation schemes.

[0067] In one feasible implementation, obtaining the current maximum power point step size coefficient of the photovoltaic system in the above embodiments includes: obtaining the real-time illuminance change rate and real-time shadow movement speed of the photovoltaic system in real time; and determining the current maximum power point step size coefficient based on the real-time illuminance change rate and real-time shadow movement speed.

[0068] Regarding the determination of the current maximum power point step size coefficient, in some embodiments, when the real-time illumination intensity change rate is greater than the upper limit of the preset illumination change rate range and the real-time shadow movement speed is greater than the upper limit of the preset shadow movement speed range, the first preset maximum power point step size coefficient is used as the current maximum power point step size coefficient; when the real-time illumination intensity change rate is within the preset illumination change rate range and the real-time shadow movement speed is within the preset shadow movement speed range, the second preset maximum power point step size coefficient is used as the current maximum power point step size coefficient; when the real-time illumination intensity change rate is less than the lower limit of the preset illumination change rate range and the real-time shadow movement speed is less than the lower limit of the preset shadow movement speed range, the third preset maximum power point step size coefficient is used as the current maximum power point step size coefficient. The preset illumination change rate range, preset shadow movement speed range, first preset maximum power point step size coefficient, second preset maximum power point step size coefficient, and third preset maximum power point step size coefficient can all be obtained and preset by the operator based on extensive experience, experiments, or statistics. Alternatively, they can be preset by the operator according to actual needs.

[0069] Regarding the values ​​of the preset illumination change rate range, preset shadow movement speed range, first preset maximum power point step size coefficient, second preset maximum power point step size coefficient, and third preset maximum power point step size coefficient, in some embodiments, this application preferably sets the preset illumination change rate range to [10, 50], the preset shadow movement speed range to [0.1, 0.5], the first preset maximum power point step size coefficient to 0.01, the second preset maximum power point step size coefficient to 0.005, and the third preset maximum power point step size coefficient to 0.002.

[0070] In this embodiment, the maximum power point step size coefficient is adaptively adjusted by dynamically sensing changes in illumination and shadow, which significantly improves the power tracking accuracy and system stability under complex operating conditions.

[0071] Understandably, the system features dynamic environment adaptation: by monitoring the rate of change of light intensity and the speed of shadow movement in real time, it can accurately identify typical operating conditions such as sunny days, cloudy days, and partial occlusion. This allows the system to quickly switch to a larger maximum power point step size coefficient for rapid response when light intensity changes abruptly (such as rapid cloud movement), and automatically switch to a smaller maximum power point step size coefficient to eliminate oscillations under steady-state conditions, significantly improving power capture efficiency compared to the traditional fixed step size method. Multi-scenario optimized control: A three-dimensional mapping relationship is established between the rate of change of light intensity, the speed of shadow movement, and the maximum power point step size coefficient. Three maximum power point step size coefficients are set for different scenarios, maintaining high MPP tracking accuracy even under severe convective weather. Fast convergence mechanism: When real-time light intensity is detected... When the rate of change of illumination intensity exceeds the upper limit of the preset illumination rate of change range, and the real-time shadow movement speed exceeds the upper limit of the preset shadow movement speed range, the maximum maximum power point step size coefficient is immediately activated, enabling the system to complete the power response in a very short time. This improves the response speed compared to traditional maximum power point tracking calculations and effectively suppresses power drops under dynamic operating conditions. Steady-state accuracy is guaranteed: In steady-state scenarios where the real-time illumination rate of change is less than the lower limit of the preset illumination rate of change range, and the real-time shadow movement speed is less than the lower limit of the preset shadow movement speed range, the minimum maximum power point step size coefficient is automatically switched to, keeping the steady-state power oscillation amplitude within a very small range. This significantly improves the stability of power generation compared to traditional fixed maximum power point step size coefficients.

[0072] In one feasible implementation, adjusting the real-time total output voltage of the k-th string channel of the m-th photovoltaic inverter based on the conductance increment and the corrected current maximum power point step size coefficient in the above embodiments includes: obtaining reference values ​​of the open-circuit voltage of all photovoltaic modules in the k-th string channel of the m-th photovoltaic inverter; determining the reference value of the total open-circuit voltage of the k-th string channel of the m-th photovoltaic inverter based on the reference values ​​of the open-circuit voltage of all photovoltaic modules in the k-th string channel of the m-th photovoltaic inverter and the module series-parallel topology information; determining the voltage change based on the reference value of the total open-circuit voltage of the k-th string channel of the m-th photovoltaic inverter and the corrected current maximum power point step size coefficient; and adjusting the real-time total output voltage of the k-th string channel of the m-th photovoltaic inverter based on the voltage change and the conductance increment.

[0073] Regarding the method for determining the total open-circuit voltage reference value of the components, in some embodiments, the open-circuit voltage reference values ​​of all photovoltaic modules in the kth string channel of the mth photovoltaic inverter can be summarized at the same time step according to the series-parallel topology information of the components and the series-parallel relationship, so as to obtain the total open-circuit voltage reference value of the components in the kth string channel of the mth photovoltaic inverter.

[0074] Regarding the method for determining the voltage change, in some embodiments, the product of the reference value of the total open-circuit voltage of the kth string channel of the mth photovoltaic inverter and the corrected current maximum power point step size coefficient can be used as the voltage change.

[0075] In the embodiments of this application, the coordinated control of the component's total open-circuit voltage reference value and the correction step size coefficient enables precise dynamic adjustment of the output voltage, significantly improving power tracking accuracy and system stability.

[0076] Understandably, the system offers several advantages: Precise voltage regulation: By aggregating the open-circuit voltage reference values ​​of all photovoltaic modules within the string channel, a precise total open-circuit voltage reference value is obtained. This, combined with a corrected maximum power point (MPP) step size factor, calculates voltage changes, ensuring a high degree of match between voltage adjustment amplitude and system state. This significantly reduces voltage adjustment errors compared to traditional methods with fixed MPP step size factors. Dynamic response optimization: When the conductance increment is non-zero, dynamic adjustments are made based on real-time calculated voltage changes and conductance increments. This allows the system to quickly respond to sudden changes in illumination (such as cloud movement) or shading changes, minimizing the tracking delay of traditional methods and greatly improving power capture efficiency under dynamic conditions. Steady-state accuracy assurance: By leveraging the benchmark effect of the total open-circuit voltage reference value, the voltage adjustment amplitude is controlled within a small range under steady-state conditions, effectively eliminating power oscillations near the maximum power point as in traditional methods, significantly improving power generation stability. Full-condition adaptability: By combining module-level parameters and module string-parallel topology information, the string channel requiring adjustment can be precisely located, maintaining high MPP tracking accuracy even in partially shaded scenarios. In one feasible implementation, adjusting the real-time total output voltage of the k-th string channel of the m-th photovoltaic inverter according to the voltage change and conductance increment in the above embodiments includes: increasing the real-time total output voltage of the k-th string channel of the m-th photovoltaic inverter according to the voltage change when the conductance increment is greater than 0; and decreasing the real-time total output voltage of the k-th string channel of the m-th photovoltaic inverter according to the voltage change when the conductance increment is less than 0.

[0077] It should be noted that, in this application, when the conductance increment is greater than 0, the real-time total output voltage of the kth string channel of the m-th photovoltaic inverter is increased by the voltage change amount, while when the conductance increment is less than 0, the real-time total output voltage of the kth string channel of the m-th photovoltaic inverter is decreased by the voltage change amount.

[0078] In this embodiment, the voltage adjustment direction is dynamically determined by the conductance increment, thereby achieving precise closed-loop control of the output voltage and significantly improving power tracking accuracy and system stability.

[0079] Understandably, the system offers several advantages: Dynamic Response Optimization: By judging the real-time conductance increment, the output voltage is automatically increased when the conductance increment is greater than 0 and automatically decreased when it is less than 0, forming a complete closed-loop control mechanism. This enables the system to respond quickly to sudden changes in illumination or shadow changes, significantly reducing tracking delay compared to traditional maximum power point tracking (MPPT) calculation schemes. Steady-State Accuracy Guarantee: Based on the zero-point judgment of the conductance increment, the voltage adjustment is automatically terminated, effectively eliminating power oscillations near the maximum power point in traditional MPT calculation schemes. This keeps the steady-state power fluctuation range within a very small range, significantly improving power generation stability. Full-Condition Adaptability: This strategy is applicable to both steady-state and dynamic conditions. It maintains high-precision tracking when illumination intensity changes slowly and maintains high MPP tracking accuracy even in sudden scenarios such as rapidly moving clouds, significantly improving adaptability compared to traditional single control modes. Closed-Loop Self-Optimization Mechanism: Through the coordinated regulation of conductance increment and voltage change, a closed-loop control chain of detection, judgment, and adjustment is formed, effectively suppressing the growth of power prediction error over long-term system operation. This is significantly better than the error drift characteristics of traditional MPT calculation schemes.

[0080] In one feasible implementation, the method in the above embodiments further includes: correcting the attention layer parameters of the multi-head attention layer in the preset power prediction model based on the real-time absolute error to obtain a corrected preset power prediction model; and using the corrected preset power prediction model as the preset power prediction model.

[0081] For the correction process, in some embodiments, a formula can be used. Correct the attention layer parameters of the multi-head attention layer; among which, and The corrected attention layer parameters for the h-th attention head in a multi-head attention layer. and The attention layer parameters of the h-th attention head in the multi-head attention layer are... To preset the learning rate, For real-time absolute error, and It is the partial derivative of the real-time absolute error with respect to the attention layer parameters of the h-th attention head in the multi-head attention layer.

[0082] In the embodiments of this application, the attention parameters of the attention layer are dynamically corrected by real-time absolute error, which significantly improves the adaptability and long-term stability of the power prediction model.

[0083] Understandably, the dynamic model optimization achieves dynamic fine-tuning of model parameters by calculating the gradient relationship between the real-time absolute error and the attention parameters, enabling the preset power prediction model to continuously adapt to changes in the operating state of the photovoltaic system, significantly improving prediction accuracy compared to traditional static power prediction models. The error suppression mechanism establishes a closed-loop error feedback control system, automatically triggering parameter correction when a prediction deviation is detected, effectively suppressing the cumulative effect of errors and keeping the growth of prediction errors under long-term system operation within a minimal range. Enhanced adaptability is achieved through dynamic adjustment of the attention parameters, allowing the model to automatically learn the feature weight allocation patterns under different operating conditions, maintaining high-precision prediction capabilities even in complex scenarios such as cloudy skies and partial shadows. The closed-loop self-optimization characteristic forms a complete closed loop of prediction, error detection, parameter correction, and re-prediction, enabling the model to continuously learn and exhibiting stronger environmental adaptability and robustness compared to traditionally trained power prediction models.

[0084] In one feasible implementation, the preset power prediction model in the above embodiment includes an input layer, a hidden layer, a multi-head attention layer, and an output layer connected in sequence. The hidden layer includes two stacked LSTM layers, and the output layer is a fully connected layer. The input layer is used to receive coupled feature vector time-series data. The hidden layer is used to perform time-series capture processing on the coupled feature vector time-series data to obtain first feature vector time-series data and second feature vector time-series data, and then concatenates the first feature vector time-series data and the second feature vector time-series data to obtain deep feature vector time-series data. The multi-head attention layer is used to determine the feature weight time-series data of each attention head based on the deep feature vector time-series data, and then perform weight allocation processing on the deep feature vector time-series data based on the feature weight time-series data of each attention head to obtain the target feature vector time-series data of each attention head, and then perform feature fusion processing on the target feature vector time-series data of each attention head to obtain fused feature vector time-series data. The output layer is used to perform linear transformation processing on the fused feature vector time-series data to obtain the module prediction power time-series data of the nth photovoltaic module within a preset future time period.

[0085] It should be noted that the first feature vector time series data is the time series data output by the first LSTM layer in the hidden layer, while the second feature vector time series data is the time series data output by the first LSTM layer in the hidden layer.

[0086] Regarding the number of attention heads in a multi-head attention layer, in some embodiments, this application preferably sets the number of attention heads in the multi-head attention layer to 4.

[0087] In some embodiments, the determination of the time-series data of the feature weights for each attention head can be achieved using formulas. Determine the temporal data of the feature weights for each attention head; Let h be the feature weights of the h-th attention head at the t-th time step in the time series data. For natural numbers, and Let h be the attention layer parameters of the h-th attention head. Let be the depth feature vector at time step t in the time series data. This represents the total number of time steps in the deep feature vector time series data.

[0088] In this embodiment of the application, by constructing a deep learning architecture that combines stacked dual-layer LSTM with multi-head attention, the accuracy and environmental adaptability of photovoltaic power prediction under complex time-series patterns are significantly improved.

[0089] Understandably, the architecture offers several advantages: Deep time-series feature mining: The dual-layer LSTM structure, through stacked time-series modeling, automatically extracts long-term and short-term dependencies from multi-source time-series data. Compared to single-layer LSTM models, it possesses stronger time-series pattern capture capabilities, making it particularly suitable for complex scenarios where photovoltaic power is dynamically affected by multiple factors such as sunlight and temperature. Multi-dimensional feature fusion: The multi-head attention mechanism, through parallel computation of multiple attention heads, can simultaneously focus on feature patterns at different time scales (such as daily periodic fluctuations and instantaneous mutations). Dynamic allocation of feature weights enables adaptive fusion of multi-dimensional information, effectively addressing the problem of insufficient feature utilization in traditional power prediction models. Enhanced adaptability to complex operating conditions: This architecture combines the time-series memory capability of LSTM with the feature-focusing advantages of the attention mechanism, maintaining high-precision predictions even under complex conditions such as cloudy skies and partial shading. It exhibits stronger environmental adaptability than traditional statistical and basic machine learning power prediction models. Enhanced model interpretability: Through visual analysis of attention weights, the contribution of features at different time steps to the prediction results can be intuitively displayed, providing interpretable decision-making basis for photovoltaic system operation and maintenance optimization, while also aiding in model debugging and performance optimization.

[0090] In a second aspect, this application provides a photovoltaic system power prediction device.

[0091] Please see Figure 2 This is a schematic diagram of a photovoltaic system power prediction device according to an embodiment of this application. The device 210 includes: The acquisition module 211 is used to acquire the multi-source time-series data of the nth photovoltaic module in the photovoltaic system, where the initial value of n is 1; The first determining module 212 is used to determine the first coupling coefficient time series data based on the irradiance time series data, ambient temperature time series data and module attenuation coefficient time series data in the multi-source time series data of the nth photovoltaic module; The second determining module 213 is used to determine the second coupling coefficient time series data based on the component short-circuit current time series data and the cloud shading ratio time series data in the multi-source time series data of the nth photovoltaic module. Module 214 is used to construct coupling feature vector time series data based on the multi-source time series data of the nth photovoltaic module, the time series data of the first coupling coefficient, and the time series data of the second coupling coefficient. The model prediction module 215 is used to input the coupled feature vector time series data into the preset power prediction model to obtain the module prediction power time series data of the nth photovoltaic module within a preset future time period; The iterative aggregation module 216 is used to set n=n+1 and return to the step of obtaining the multi-source time series data of the nth photovoltaic module in the photovoltaic system until n is greater than the total number of photovoltaic modules in the photovoltaic system. Based on the component predicted power time series data of all photovoltaic modules in the preset future time, the predicted total power time series data of the photovoltaic system in the preset future time is determined.

[0092] In this embodiment of the application, the relevant contents of the above-mentioned acquisition module 211, first determination module 212, second determination module 213, construction module 214, model prediction module 215 and iterative summarization module 216 can be found in the following references. Figure 1 The contents of the illustrated embodiments will not be repeated here.

[0093] It should be noted that the device 210 of this application also includes other modules. It can be understood that the method of this application and the device 210 have a one-to-one correspondence. Therefore, the other modules of the device 210 of this application are the contents corresponding to the method of this application in the above embodiments.

[0094] In this embodiment, by constructing coupled feature vector time-series data and power prediction for each photovoltaic module, multi-factor coupling and individual module differences are considered. This effectively solves the problems of low prediction accuracy, significant calculation errors and power estimation under complex operating conditions caused by insufficient consideration of multi-factor coupling and neglect of individual module differences in the prior art. It improves power prediction accuracy, reduces calculation errors and power estimation deviations under complex operating conditions, and ultimately achieves the comprehensive technical effect of improving photovoltaic system power generation efficiency, enhancing grid connection stability and maximizing economic benefits.

[0095] In addition to the aforementioned benefits such as improving power prediction accuracy, reducing calculation errors and power estimation deviations, increasing photovoltaic system power generation efficiency, enhancing grid connection stability, and maximizing economic benefits, this photovoltaic system power prediction device also has the following advantages: Improved data quality: When acquiring multi-source time-series data, the original multi-source time-series data is first cleaned and standardized. Through outlier removal and missing value interpolation, the integrity and accuracy of the data are ensured, avoiding interference from outlier and missing data in subsequent analysis. This provides a reliable foundation for constructing accurate coupled feature vector time-series data. Furthermore, the data is standardized to […]. Within the interval [0, 1], the influence of dimensions between different data dimensions is eliminated, enabling different types of data to be compared and analyzed on the same scale, further improving data quality and usability; Comprehensive data utilization: The selected multi-source time series data covers information from multiple aspects such as environmental dynamics, component characteristics, and system operation, comprehensively considering various factors affecting photovoltaic module power. This comprehensive data utilization method can more accurately reflect the actual operating status of the photovoltaic system and provide rich information support for power prediction; Enhanced model adaptability: By determining the time series data of the first coupling coefficient and the second coupling coefficient, the different factors are quantified. This approach, which incorporates coupling coefficients into the coupling feature vector time series data, allows the preset power prediction model to better adapt to complex and changing operating conditions. The model considers not only the influence of individual factors but also the interactions between them, thus improving its predictive ability under different environmental conditions. Furthermore, a preset power prediction model can be trained for each photovoltaic module, fully considering individual module differences. Different photovoltaic modules may exhibit different performance due to production errors, aging, and other factors. Training a separate model for each module allows for more accurate capture of its characteristics, further improving the accuracy of power prediction. The model training process is optimized by setting a loss function to assess the training level during model training. This allows for timely understanding of the model's training effect and adjustment of the training strategy based on changes in the loss function, ensuring the model reaches its optimal state and improving its performance and generalization ability. Finally, the device facilitates system maintenance and management by acquiring detailed information for each photovoltaic module, including multi-source time series data and predicted power time series data. Analysis of this data allows for timely detection of abnormalities in photovoltaic modules, such as performance degradation and malfunctions, facilitating targeted maintenance and management by operations and maintenance personnel, extending the lifespan of photovoltaic modules, and reducing maintenance costs.Supporting Intelligent Decision-Making: Accurate time-series data of total power prediction for photovoltaic (PV) systems can provide crucial information for grid dispatching and energy management. Grid dispatching departments can use this data to rationally schedule power generation, optimize grid operation, and improve energy efficiency. Energy management departments can use this data for energy storage and allocation, achieving rational energy distribution and utilization, and supporting the development of smart energy systems. Promoting the Development of the PV Industry: This device improves the accuracy and reliability of PV system power prediction, helping to reduce the uncertainty of PV power generation, enhancing investor confidence in PV projects, and promoting investment and development in the PV industry. Simultaneously, accurate power prediction also helps improve the competitiveness of PV systems in the electricity market and promotes the integration of the PV industry with other energy industries.

[0096] In a third aspect, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform a photovoltaic system power prediction method as described in any of the first aspects.

[0097] This application provides a computer device in a fourth aspect, including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform a photovoltaic system power prediction method as described in any of the first aspects.

[0098] Figure 3 The diagram illustrates the internal structure of a computer device in some embodiments. This computer device may specifically be a terminal, a server, or a gateway. Figure 3 As shown, the computer device includes a processor, memory, and network interface connected via a system bus.

[0099] The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium of the computer device stores an operating system and may also store a computer program. When executed by a processor, this computer program causes the processor to perform the steps in the above method embodiments. The internal memory may also store a computer program, which, when executed by a processor, causes the processor to perform the steps in the above method embodiments. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0100] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods.

[0101] Any references to memory, storage, database, or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0102] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0103] The embodiments described above are merely examples of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application.

Claims

1. A method for predicting the power of a photovoltaic system, characterized in that, The method includes: Obtain the multi-source time-series data of the nth photovoltaic module in the photovoltaic system, where the initial value of n is 1; Based on the irradiance time-series data, ambient temperature time-series data, and module attenuation coefficient time-series data in the multi-source time-series data of the nth photovoltaic module, determine the first coupling coefficient time-series data; The second coupling coefficient time series data is determined based on the component short-circuit current time series data and cloud shading ratio time series data in the multi-source time series data of the nth photovoltaic module. Based on the multi-source time-series data of the nth photovoltaic module, the time-series data of the first coupling coefficient, and the time-series data of the second coupling coefficient, a coupling feature vector time-series data is constructed. The time series data of the coupled feature vector is input into the preset power prediction model to obtain the time series data of the predicted power of the nth photovoltaic module within a preset future time period; Let n = n + 1, return to the step of obtaining the multi-source time series data of the nth photovoltaic module in the photovoltaic system, until n is greater than the total number of photovoltaic modules in the photovoltaic system. Based on the component predicted power time series data of all photovoltaic modules in the preset future time, determine the predicted total power time series data of the photovoltaic system in the preset future time.

2. The photovoltaic system power prediction method according to claim 1, characterized in that, The timing data of the first coupling coefficient is determined using the following formula: ； in, The first coupling coefficient is the first coupling coefficient at time step t in the time series data of the first coupling coefficient. The component attenuation coefficient is the component attenuation coefficient at time step t in the component attenuation coefficient time series data. The light intensity at time step t in the light intensity time series data is... The ambient temperature at time step t in the ambient temperature time series data is given.

3. The photovoltaic system power prediction method according to claim 1, characterized in that, The step of determining the second coupling coefficient time series data based on the module short-circuit current time series data and cloud shading ratio time series data from the multi-source time series data of the nth photovoltaic module includes: Determine the maximum, minimum, and average component short-circuit current in the component short-circuit current timing data; The component mismatch is determined based on the maximum component short-circuit current, the minimum component short-circuit current, and the average component short-circuit current. The second coupling coefficient time series data is determined based on the component mismatch degree and the cloud occlusion ratio time series data.

4. The photovoltaic system power prediction method according to claim 3, characterized in that, The timing data for determining the component mismatch and the second coupling coefficient are obtained using the following formula: ； in, The component mismatch degree, The maximum component short-circuit current, The short-circuit current of the minimum component. The average component short-circuit current. The second coupling coefficient is the second coupling coefficient at time step t in the time series data. The cloud occlusion percentage is the cloud occlusion percentage at time step t in the time series data of cloud occlusion percentage.

5. The photovoltaic system power prediction method according to claim 1, characterized in that, The method further includes: The system acquires the real-time output power of all photovoltaic modules in the k-th string channel of the m-th photovoltaic inverter in the photovoltaic system, and acquires the module series-parallel topology information of the photovoltaic system, where the initial values ​​of m and k are both 1. Based on the real-time output power of all photovoltaic modules in the kth string channel of the mth photovoltaic inverter and the serial-parallel topology information of the modules, determine the real-time total output power of the kth string channel of the mth photovoltaic inverter. Based on the time-series data of the predicted power of all photovoltaic modules within a preset future time period, and the information on the series and parallel topology of the modules, the predicted total output power of the kth string channel of the mth photovoltaic inverter is determined, wherein the predicted total output power of the kth string channel of the mth photovoltaic inverter and the real-time total output power are the power at the same time step. Determine the real-time relative error between the real-time total output power and the predicted total output power of the k-th string channel of the m-th photovoltaic inverter; If the average value of the real-time relative error over a number of consecutive preset sampling periods is greater than a preset percentage, determine the real-time absolute error between the real-time total output power of the k-th string channel of the m-th photovoltaic inverter and the predicted total output power. Obtain the current maximum power point step size coefficient of the photovoltaic system; Based on the real-time absolute error, the current maximum power point step size coefficient of the photovoltaic system is corrected to obtain the corrected current maximum power point step size coefficient. The conductivity increment is determined based on the corrected current maximum power point step size coefficient, and the real-time total output power and real-time total output voltage of the kth string channel of the mth photovoltaic inverter. When the conductance increment is not equal to 0, the real-time total output voltage of the kth string channel of the mth photovoltaic inverter is adjusted according to the conductance increment and the corrected current maximum power point step size coefficient. Let k = k + 1, return to the step of obtaining the real-time output power of all photovoltaic modules in the kth string channel of the mth photovoltaic inverter in the photovoltaic system, until k is greater than the total number of string channels of the mth photovoltaic inverter, and adjust the real-time total output voltage of all string channels of the mth photovoltaic inverter. Let m = m + 1, then return to the step of obtaining the real-time output power of all photovoltaic modules in the k-th string channel of the m-th photovoltaic inverter in the photovoltaic system, until m is greater than the total number of photovoltaic inverters in the photovoltaic system, and adjust the real-time total output voltage of all string channels of all photovoltaic inverters.

6. The photovoltaic system power prediction method according to claim 5, characterized in that, The step size factor for obtaining the current maximum power point of the photovoltaic system includes: Real-time acquisition of the real-time rate of change of light intensity and the real-time shadow movement speed of the photovoltaic system; The step size coefficient of the current maximum power point is determined based on the real-time light intensity change rate and the real-time shadow movement speed.

7. The photovoltaic system power prediction method according to claim 5, characterized in that, The step of adjusting the real-time total output voltage of the k-th string channel of the m-th photovoltaic inverter based on the conductance increment and the corrected current maximum power point step size coefficient includes: Obtain the reference values ​​of the open-circuit voltage of all photovoltaic modules in the kth string channel of the mth photovoltaic inverter; Based on the reference values ​​of the open-circuit voltage of all photovoltaic modules in the kth string channel of the mth photovoltaic inverter, and the information on the series-parallel topology of the modules, the reference value of the total open-circuit voltage of the modules in the kth string channel of the mth photovoltaic inverter is determined. The voltage change is determined based on the reference value of the total open-circuit voltage of the kth string channel of the mth photovoltaic inverter and the corrected current maximum power point step size coefficient. The real-time total output voltage of the k-th string channel of the m-th photovoltaic inverter is adjusted based on the voltage change and the conductance increment.

8. The photovoltaic system power prediction method according to claim 7, characterized in that, The step of adjusting the real-time total output voltage of the k-th string channel of the m-th photovoltaic inverter based on the voltage change and the conductance increment includes: When the conductivity increment is greater than 0, the real-time total output voltage of the kth string channel of the mth photovoltaic inverter is increased according to the voltage change. When the conductivity increment is less than 0, the real-time total output voltage of the kth string channel of the mth photovoltaic inverter is reduced according to the voltage change.

9. The photovoltaic system power prediction method according to claim 5, characterized in that, The method further includes: Based on the real-time absolute error, the attention layer parameters of the multi-head attention layer in the preset power prediction model are corrected to obtain the corrected preset power prediction model. The modified preset power prediction model is used as the preset power prediction model.

10. The photovoltaic system power prediction method according to claim 1, characterized in that, The preset power prediction model includes an input layer, a hidden layer, a multi-head attention layer and an output layer connected in sequence. The hidden layer includes two stacked LSTM layers and the output layer is a fully connected layer. The input layer is used to receive the temporal data of the coupled feature vector; The hidden layer is used to perform time-series capture processing on the coupled feature vector time-series data to obtain first feature vector time-series data and second feature vector time-series data, and then concatenates the first feature vector time-series data and the second feature vector time-series data to obtain deep feature vector time-series data; The multi-head attention layer is used to determine the feature weight time-series data of each attention head based on the deep feature vector time-series data, and to perform weight allocation processing on the deep feature vector time-series data based on the feature weight time-series data of each attention head to obtain the target feature vector time-series data of each attention head, and to perform feature fusion processing on the target feature vector time-series data of each attention head to obtain fused feature vector time-series data. The output layer is used to perform linear transformation on the time series data of the fused feature vector to obtain the time series data of the predicted power of the nth photovoltaic module within the preset future time period.

Images