Method and system for predicting power of string type photovoltaic inverter

Abstract

The invention discloses a string type photovoltaic inverter power prediction method and system, and relates to the technical field of photovoltaic inverters. According to the method, through dynamic abnormal data processing, multi-source time sequence alignment and multi-dimensional feature construction, the problems of large dimensional difference and time sequence asynchronization of original data are solved, and the generated fusion feature vector provides high-quality input for modeling and lays a prediction precision foundation; a dynamic response model is constructed in combination with a physical principle and data driving, targeted adjustment is achieved through a multi-working-condition feature library, and transient changes and non-ideal factor influences are accurately captured; the efficiency mapping model dynamic calibration mechanism further improves the power evaluation accuracy, solves the problem of poor working condition adaptability of a traditional model, constructs ultra-short-term, short-term and medium-term prediction sub-models, dynamically fuses the sub-models, combines error analysis and uncertainty quantification, outputs a multi-confidence prediction result, and improves the accuracy of power evaluation. Different time scale requirements of power grid dispatching, plan making and the like are met, and photovoltaic consumption and power grid stability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic inverters, and in particular to a power prediction method and system for string photovoltaic inverters. Background Art

[0002] String photovoltaic inverters are widely used due to their high flexibility. However, their output power fluctuates significantly due to environmental factors such as irradiance and temperature, as well as non-ideal factors such as component aging and shading, posing challenges to grid scheduling and power consumption. Chinese patent application publication number CN106156466A discloses a method for plotting a predicted output power curve for a photovoltaic inverter, comprising the following steps: S1. Selecting an initial reference curve; S2. Correcting the reference curve based on the measured actual output power of the photovoltaic inverter on that day and the actual output power on the day corresponding to the initial reference curve; S3. Obtaining the current time, the actual output power of the photovoltaic inverter, and the power of the reference curve, inputting prediction coefficients, and calculating the output power at the next time; S4. Concatenating the calculated predicted output powers into a predicted output power curve; S5. Determining whether the current time is between a first preset time and a second preset time; if so, executing step S3 after the first preset time period. This method corrects the reference curve and incorporates prediction coefficients to calculate the output power at the next time, ensuring that the photovoltaic inverter output power curve more closely matches the actual output power curve.

[0003] Although the above patent improves the fit of the photovoltaic inverter output power curve, the following problems still exist: 1. Correction is based solely on the overall inverter output power, ignoring the differences in power characteristics of each PV string due to geographical location, component aging, and shading conditions. This makes it difficult to accurately reflect the distributed characteristics of string inverters. 2. The benchmark curve correction relies on historical data and cannot effectively capture transient power fluctuations caused by non-ideal factors such as sudden changes in irradiance and high-temperature aging. The lack of a dynamic adjustment mechanism leads to significant errors under extreme working conditions. 3. The time scale is single, and it only provides short-term power forecasts at adjacent moments. It cannot meet the needs of diverse scenarios such as real-time grid dispatching and daily planning. It also does not involve the quantification of uncertainty in the prediction results, making it difficult to support decision reliability assessment. Summary of the Invention

[0004] The purpose of the present invention is to provide a power prediction method and system for string photovoltaic inverters, combining physical principles and data-driven construction of a dynamic response model, achieving targeted adjustment through a multi-operating condition feature library, accurately capturing transient changes and the influence of non-ideal factors, constructing ultra-short-term, short-term, and medium-term prediction sub-models and dynamically integrating them, combining error analysis and uncertainty quantification, and outputting multi-confidence prediction results to meet the different time scale requirements of grid scheduling, planning, etc., improve photovoltaic absorption and grid stability, and solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions: A method for predicting the power of a string photovoltaic inverter includes: Photovoltaic inverter parameter collection: Based on a distributed sensor network, the operating parameters of the photovoltaic inverter are collected in real time, and external environmental data is obtained at the same time, and the obtained operating parameters and external environmental data are pre-processed; The photovoltaic inverter dynamic response model is constructed based on the pre-processed operating parameters and external environmental data to extract the transient and steady-state characteristics of the photovoltaic inverter output power as key factors change. Photovoltaic inverter power evaluation: Establishing a mapping relationship between the input power and output power of the photovoltaic inverter, and calculating the actual output power of the photovoltaic inverter based on the conversion efficiency and internal losses of the photovoltaic inverter at different working nodes; The prediction results are integrated to obtain the historical operating parameters of the string photovoltaic inverter. The real-time operating parameters and external environmental data are combined to build a power prediction model. The ultra-short-term prediction, short-term prediction and medium-term prediction results are integrated to generate multi-time-scale collaborative prediction results through a weighted average strategy.

[0006] Furthermore, the photovoltaic inverter parameters are collected and pre-processed, including: Dynamic abnormal data processing: Build an adaptive threshold model based on the distribution characteristics of historical data. Use this model to monitor the deviation between the collected operating parameters and external environmental data in real time. Identify abnormal values ​​based on the data deviation, perform smoothing, and perform polynomial fitting interpolation or labeling on missing data. Multi-source data timing alignment: Establish a reference time axis through a distributed clock synchronization protocol to perform time alignment and synchronization of electrical parameters and external environment data; Multi-dimensional feature construction: Based on the pre-processed electrical parameters and external environment data, dynamic trend features, cross-domain correlation features, and status assessment features are generated, and quantified to generate a fusion feature vector of electrical parameters and external environment data.

[0007] Furthermore, the dynamic response model of the photovoltaic inverter is constructed, including: Key variable screening and correlation analysis: Based on the fused feature vector, the correlation between each feature and the actual output power of the PV inverter is calculated. The core input variables are screened based on the preset correlation threshold, and the correlation strength between the core input variables is determined. The correlation strength is compared with the preset redundant correlation threshold to eliminate redundant features. Basic response model construction: Based on the circuit principle of photovoltaic modules, a basic response model between the output power of the photovoltaic inverter and the selected core input variables is established to obtain the basic response law of the photovoltaic inverter output power as the core input variables change; Dynamic response model optimization: Based on dynamic trend characteristics and cross-domain correlation characteristics, the deviations generated by the basic response model when predicting the output power of the photovoltaic inverter are identified and quantified. The deviations are quantified and used as correction items to correct the basic response model. The output power value of the dynamic response model of the photovoltaic inverter after correction is used as the final output.

[0008] Furthermore, the construction of the photovoltaic inverter dynamic response model also includes: Collect PV inverter operating data in different seasons, weather types, and time periods, and build a multi-operating condition feature library by combining corresponding core input variables, dynamic trend features, and cross-domain correlation features. Obtain historical operating data of string photovoltaic inverters, extract operating condition features from the historical operating data, and perform similarity matching with a multi-operating condition feature library to determine different operating conditions in the historical operating data; Based on the actual output power of the photovoltaic inverter under different working conditions in the historical operating data, the adaptability of the modified dynamic response model is verified, and the deviation rate between the output power of the dynamic response model and the actual output power is calculated; When the deviation rate exceeds the preset deviation threshold, the core input variables, dynamic trend characteristics and cross-domain correlation characteristics of the corresponding working conditions are obtained, the mutual correlation strength between the core input variables is recalculated, and the parameters of the basic response model and the weights of the correction items are adjusted.

[0009] Further, the photovoltaic inverter power evaluation includes: Basic data acquisition: Based on the output power value output by the dynamic response model and combined with the real-time collected operating parameters, the input power of the photovoltaic inverter is determined. At the same time, the output voltage, current and operating temperature data of the photovoltaic inverter AC side are collected; Constructing an efficiency mapping model: Based on the input power, AC side output voltage, current, and operating temperature data of the photovoltaic inverter, construct an efficiency mapping model for the photovoltaic inverter; Loss quantification and actual output power calculation: Decompose the internal losses of the photovoltaic inverter and combine the conversion efficiency output by the efficiency mapping model with the quantified loss value to calculate the actual output power of the photovoltaic inverter.

[0010] Furthermore, the photovoltaic inverter power evaluation also includes dynamic calibration of the mapping relationship: Based on the multi-operating condition feature library, establishing an efficiency mapping model parameter subset for different operating condition categories; Collect PV inverter operating parameters in real time and match them to corresponding working conditions, extract corresponding efficiency mapping model parameter subsets to calibrate the efficiency mapping model; When the deviation between the actual output power calculated value and the actual measured value within the sampling period exceeds a preset deviation threshold, and the sampling period exceeds a preset sampling period range, the parameter recalibration mechanism is triggered.

[0011] Furthermore, the prediction result integration also includes: Based on historical and real-time operating parameters, ultra-short-term, short-term, and medium-term prediction sub-models are constructed, and the power prediction results of each prediction sub-model are quantitatively evaluated. Analyze the probability distribution characteristics and main occurrence periods of prediction errors, locate the high-incidence error conditions by combining the multi-condition feature library, and identify the key factors that cause the errors; According to the prediction error analysis results, each prediction sub-model is optimized to generate prediction intervals under different confidence levels. Combined with the probability distribution characteristics of the prediction error, a comprehensive prediction result including the optimal prediction value, uncertainty interval and confidence level is output.

[0012] Furthermore, after generating the multi-time-scale collaborative prediction results by the weighted average strategy, the method further includes: Continuously obtaining in real time the multimodal application of the multi-time-scale collaborative prediction results; Based on the multimodal application, the application value score of the multi-time scale collaborative prediction result is calculated by the following formula: in, Score the application value of the multi-time-scale collaborative prediction results, For the multimodal application case described in The ratings under the application value rating indicators are: For the The preset scoring weights corresponding to the application value scoring indicators are: For the The preset error coefficient corresponding to each application value scoring indicator, is the total number of scoring indicators, Score the application value of the previous calculation, The maximum value of the historical application value scores calculated within the recent preset time. and is the preset base weight, is a preset constant; When the application value score is lower than a preset score threshold, extracting an optimization basis feature set from the multimodal application situation based on a preset feature extraction template; Based on the optimization basis feature set, matching the corresponding model optimization strategy from a preset model optimization strategy library; Optimizing the power prediction model based on the model optimization strategy; Based on the optimized power prediction model, the ultra-short-term prediction, short-term prediction and medium-term prediction results are re-fused on a time scale, and a new multi-time-scale collaborative prediction result is re-generated through a weighted average strategy.

[0013] Furthermore, after generating the multi-time-scale collaborative prediction results by the weighted average strategy, the method further includes: The edge computing nodes deployed at the photovoltaic station side continuously receive the actual output power measurement values ​​of the photovoltaic inverters from the sensor network; The continuously received actual output power measurement values ​​of the photovoltaic inverter are compared point by point with the multi-time-scale collaborative prediction results; Based on the results of point-by-point comparison, a time series prediction error matrix is ​​constructed; Based on the constructed time series prediction error matrix, a multi-dimensional error tracing analysis is performed; Based on the results of the multi-dimensional error tracing analysis, predictive compensation is performed on the power prediction model; Based on the power prediction model after prediction compensation, the ultra-short-term prediction, short-term prediction and medium-term prediction results are re-time-scale integrated, and a new multi-time-scale collaborative prediction result is re-generated through a weighted average strategy; The multi-dimensional error tracing analysis based on the constructed time series prediction error matrix includes: Identify the correlation between external environmental data and forecast deviations; Extract the historical attenuation curve of inverter conversion efficiency and quantify the impact coefficient of device aging on power mapping relationship.

[0014] The present invention provides another technical solution, a string photovoltaic inverter power prediction system, comprising: The distributed parameter acquisition module is configured to collect the operating parameters and environmental data of the photovoltaic inverters in real time based on the sensors deployed on each photovoltaic inverter, and perform preprocessing to generate a fused feature vector containing dynamic trend features, cross-domain correlation features, and status assessment features; a dynamic modeling module configured to screen core input variables and eliminate redundant features based on the fused feature vector, establish a basic response model between the photovoltaic inverter output power and the core input variables, and modify the basic response model based on dynamic trend features and cross-domain correlation features to generate a dynamic response model of the photovoltaic inverter; The operating condition feature library module is configured to build a multi-operating condition feature library based on historical operating data, and store multi-operating condition feature data, including core input variables, dynamic trend features, cross-domain correlation features and corresponding model parameter subsets under different operating conditions; a power evaluation module configured to determine the input power of the photovoltaic inverter based on the output power value output by the dynamic response model, construct an efficiency mapping model of the photovoltaic inverter, and dynamically calibrate the efficiency mapping model based on a multi-operating condition feature library; The multi-time-scale prediction integration module is configured to construct ultra-short-term, short-term, and medium-term prediction sub-models, perform quantitative evaluation of the power prediction results of each prediction sub-model, optimize each prediction sub-model based on the prediction error analysis results, and fuse the multi-time-scale prediction results to generate a comprehensive prediction result that includes the optimal prediction value, uncertainty interval, and confidence level.

[0015] Compared with the prior art, the present invention has the following beneficial effects: Through dynamic abnormal data processing, multi-source time series alignment and multi-dimensional feature construction, the problems of large dimensional differences and time series asynchrony in the original data are solved. The generated fusion feature vector provides high-quality input for modeling, laying the foundation for prediction accuracy; combining physical principles and data-driven to build a dynamic response model, and achieving targeted adjustment through a multi-operating condition feature library to accurately capture transient changes and the influence of non-ideal factors; the dynamic calibration mechanism of the efficiency mapping model further improves the accuracy of power assessment and solves the problem of poor adaptability of traditional models to operating conditions. Ultra-short-term, short-term and medium-term prediction sub-models are constructed and dynamically integrated, combined with error analysis and uncertainty quantification, to output multi-confidence prediction results to meet the different time scale requirements of grid dispatching, planning, etc., and improve photovoltaic absorption and grid stability.

[0016] The system continuously optimizes itself to cope with changes in different environments and load conditions, improving the accuracy of power prediction and system stability.

[0017] Through edge computing nodes, high-precision and real-time photovoltaic inverter power prediction and compensation can be achieved, which can accurately identify and correct prediction errors, improve prediction accuracy, quickly respond to environmental changes and equipment status, enhance system adaptability, extend equipment life, and optimize grid scheduling through multi-time scale prediction fusion, providing strong technical support. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1This is a flow chart of the power prediction method for a string photovoltaic inverter of the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0020] To address the technical issues that existing technologies do not consider the differences in group-level characteristics, are difficult to reflect distributed characteristics, are not adaptable to complex working conditions such as sudden changes in irradiance, lack dynamic adjustment of prediction coefficients, and have a single time scale that cannot meet the needs of multiple scenarios and cannot support diversified decision-making, please refer to Figure 1 , this embodiment provides the following technical solutions: A method for predicting the power of a string photovoltaic inverter includes: Photovoltaic inverter parameter acquisition: Based on a distributed sensor network, the system collects the operating parameters of the photovoltaic inverters in real time, including the open-circuit voltage, short-circuit current, operating temperature, and DC bus voltage of each photovoltaic inverter. It also obtains external environmental data, including irradiance, temperature, humidity, and numerical weather forecast data. The acquired operating parameters and external environmental data are preprocessed to eliminate the influence of different dimensions and facilitate subsequent model processing. The dynamic response model of the photovoltaic inverter is constructed based on the pre-processed operating parameters and external environmental data. The transient and steady-state characteristics of the photovoltaic inverter output power are extracted as the key factors such as irradiance and temperature change. Photovoltaic inverter power evaluation: Establishing a mapping relationship between the input power and output power of the photovoltaic inverter, and calculating the actual output power of the photovoltaic inverter based on the conversion efficiency and internal losses of the photovoltaic inverter at different working nodes; The prediction results are integrated to obtain the historical operating parameters of the string photovoltaic inverter. The real-time operating parameters and external environmental data are combined to build a power prediction model that adapts to different prediction time scales. The ultra-short-term prediction, short-term prediction and medium-term prediction results are time-scale fused, and a weighted average strategy is used to generate multi-time-scale collaborative prediction results.

[0021] In this embodiment, multi-source data preprocessing is used to solve the problems of dimensional differences and time series asynchrony, and the generated fusion feature vector provides high-quality input for modeling. Its dynamic feature construction accurately adapts to the string characteristics, achieving more refined data processing than the existing technology; on this basis, the physical principles and data-driven construction model are integrated to accurately capture transient and steady-state characteristics, and solve the problem of poor adaptability of traditional models to working conditions through multi-working condition adjustment, significantly improving the prediction accuracy in complex scenarios; further combining loss quantification and dynamic calibration mechanism, breaking through the roughness of traditional efficiency evaluation, relying on the adaptation of multi-working condition parameter subsets to achieve accurate output of actual output power; finally, multi-time scale fusion is used to meet diverse needs, uncertainty quantification improves decision reliability, and weighted fusion strategy significantly enhances adaptability to power grid scheduling, forming a complete technical closed loop from data processing to decision support.

[0022] In this embodiment, the photovoltaic inverter parameter collection and preprocessing include: Dynamic abnormal data processing: Build an adaptive threshold model based on the distribution characteristics of historical data. Use this model to monitor the deviation between the collected operating parameters and external environmental data in real time. Identify abnormal values ​​based on the data deviation, perform smoothing, and perform polynomial fitting interpolation or labeling on missing data. Multi-source data timing alignment: A distributed clock synchronization protocol is used to establish a reference time axis, aligning and synchronizing electrical parameters with external environmental data. Using the DC-side voltage jump signal of the photovoltaic inverter as the time anchor, data with different sampling frequencies are mapped to a unified time scale. Synchronization deviations are corrected through quadratic curve fitting to ensure that the time axis error does not exceed one-tenth of the sampling period. Multi-dimensional feature construction: Based on the pre-processed electrical parameters and external environmental data, dynamic trend features, cross-domain correlation features, and status assessment features are generated. For example, the average irradiance, irradiance change rate, and harmonic content of voltage and current within a specific time period are calculated and quantified to generate a fusion feature vector of electrical parameters and external environmental data, which is directly used to construct the input layer of the dynamic response model of the photovoltaic inverter.

[0023] In this embodiment, the dynamic trend feature is as follows: the irradiance change slope, voltage and current fluctuation amplitude, and temperature cumulative change within three consecutive sampling windows are calculated, and the weight of recent data is strengthened through a sliding weighting algorithm; In this embodiment, cross-domain correlation features: construct a coupling coefficient matrix between environmental parameters and electrical parameters, calculate interactive features such as irradiance-open circuit voltage correlation and temperature-short circuit current sensitivity, and quantify the impact of environmental factors on the electrical characteristics of photovoltaic inverters; In this embodiment, the status assessment feature is as follows: the component performance attenuation coefficient is calculated based on the open-circuit voltage temperature coefficient and the operating temperature deviation value, and the health status feature vector is generated in combination with the short-circuit current change rate to characterize the long-term operating status of the photovoltaic inverter.

[0024] In this embodiment, the eigenvalues ​​are mapped to the interval [0,1], and dynamic weight factors are assigned to different types of features. The weight values ​​are dynamically adjusted according to their mutual information values ​​with historical power data. The processed data are directly used to construct the input layer of the dynamic response model of the photovoltaic inverter.

[0025] In this embodiment, the photovoltaic inverter dynamic response model is constructed, including: Key variable screening and correlation analysis: Based on the fused feature vector, the correlation between each feature and the actual output power of the photovoltaic inverter is calculated. Based on the preset correlation threshold, irradiance, operating temperature, open-circuit voltage, fill factor, and component performance attenuation coefficient are selected as core input variables. The correlation strength between the core input variables is determined, and the correlation strength is compared with the preset redundant correlation threshold to eliminate redundant features. Basic response model construction: Based on the principles of photovoltaic module circuits, a basic response model is established between the output power of the photovoltaic inverter and the selected core input variables. The basic response law of the photovoltaic inverter output power as the core input variables change is obtained, such as the power-irradiance curve and the power-temperature curve. Dynamic response model optimization: Based on dynamic trend characteristics and cross-domain correlation characteristics, the deviations caused by ignoring rapid dynamic changes and interactions between photovoltaic inverters / between photovoltaic inverters and the environment when predicting the output power of the basic response model are identified and quantified. The deviations are quantified and used as correction terms to correct the basic response model to improve the model's response accuracy to actual operating conditions. The output power value of the corrected dynamic response model of the photovoltaic inverter is used as the final output, which reflects the comprehensive response of the photovoltaic inverter output power to the core input variables and their dynamic changes at the current and adjacent moments.

[0026] In this embodiment, the photovoltaic inverter dynamic response model is constructed, further comprising: Collect PV inverter operating data from different seasons, weather types (e.g., sunny, cloudy, rainy, snowy), and time periods (e.g., morning, noon, evening), and build a multi-operating condition feature library by combining corresponding core input variables, dynamic trend features, and cross-domain correlation features. Obtain historical operating data of string photovoltaic inverters, extract operating condition features from the historical operating data, and perform similarity matching with a multi-operating condition feature library to determine different operating conditions in the historical operating data; Based on the actual output power of the photovoltaic inverter under different working conditions in the historical operating data, the adaptability of the modified dynamic response model is verified, and the deviation rate between the output power of the dynamic response model and the actual output power is calculated; When the deviation rate exceeds the preset deviation threshold, the core input variables, dynamic trend characteristics and cross-domain correlation characteristics of the corresponding working conditions are obtained, the mutual correlation strength between the core input variables is recalculated, and the parameters of the basic response model and the weights of the correction items are adjusted.

[0027] In this embodiment, the operating condition identification results are used to classify historical operating data or model training data into corresponding operating condition categories. For each identified operating condition category, a data subset within that category is used to re-estimate or optimize the parameters (such as the temperature coefficient, irradiance response coefficient, etc.) and dynamic correction terms (such as correction factors based on dynamic trends and cross-domain correlations) in the basic response model (e.g., a power calculation model based on circuit principles). For example, for sunny noon operating conditions, the model's temperature coefficient may need to be adjusted to more accurately reflect performance degradation under high temperatures; for cloudy weather, the weight of the dynamic correction term for rapidly changing irradiance may need to be increased. Through this targeted adjustment based on the operating conditions, the dynamic response model can dynamically adjust its internal parameters and correction strategies based on the currently identified operating conditions, thereby maintaining high prediction accuracy under all types of operating conditions.

[0028] In this embodiment, the photovoltaic inverter power evaluation includes: Basic data acquisition: Based on the output power value output by the dynamic response model and combined with the real-time collected operating parameters, the input power of the photovoltaic inverter is determined. At the same time, the output voltage, current and operating temperature data of the photovoltaic inverter AC side are collected; Constructing an efficiency mapping model: Based on the input power, AC side output voltage, current, and operating temperature data of the photovoltaic inverter, an efficiency mapping model is constructed to describe the conversion efficiency curve of the photovoltaic inverter under different load rates and temperature conditions; Loss quantification and actual output power calculation: The internal losses of the PV inverter are decomposed into fixed losses and variable losses. Fixed losses include core losses and control circuit losses. Loss coefficients are fitted based on the PV inverter nameplate parameters and operating temperature. Variable losses include switching losses and conduction losses. Loss values ​​are calculated using the RMS input current and switching frequency. Combined with the conversion efficiency output by the efficiency mapping model and the quantified loss values, the actual output power of the PV inverter is calculated: actual output power = PV inverter input power × conversion efficiency - total losses. Dynamic calibration of the mapping relationship: Based on the multi-operating condition feature library, efficiency mapping model parameter subsets are established for different operating condition categories (such as high load rate-high temperature, low load rate-normal temperature, etc.); photovoltaic inverter operating parameters are collected in real time and matched with corresponding operating conditions, and the corresponding efficiency mapping model parameter subsets are extracted to dynamically calibrate the efficiency mapping model; and when the deviation between the actual output power calculated value and the actual measured value within the sampling period exceeds a preset deviation threshold, and the sampling period exceeds a preset sampling period range, the parameter recalibration mechanism is triggered, the operating parameters of the photovoltaic inverter are re-acquired, and the polynomial fitting coefficients are updated to ensure the timeliness of the mapping relationship.

[0029] In this embodiment, the efficiency mapping model accurately depicts the conversion efficiency laws under different load rates and temperatures, and the loss quantification mechanism refines the calculation of fixed and variable losses, thereby improving the reliability of the actual output power results. The dynamic calibration of the mapping relationship is combined with a subset of multi-operating condition parameters to achieve model adaptation under different operating conditions. The recalibration mechanism triggered by excessive deviations ensures long-term timeliness, solving problems such as traditional evaluation ignoring operating condition differences, rough loss quantification, and time-varying model failure. It provides high-precision actual output power data for power prediction and ensures the accuracy of multi-time scale predictions.

[0030] In this embodiment, the prediction result integration also includes: Based on historical and real-time operating parameters, ultra-short-term (0-1 hour), short-term (1-24 hours), and medium-term (24-72 hours) prediction sub-models are constructed, and the power prediction results of each prediction sub-model are quantitatively evaluated. Analyze the probability distribution characteristics and main occurrence periods of prediction errors. Combined with the multi-condition feature library, locate the conditions where errors are most likely to occur. Identify the key factors that cause the errors, which may be limitations of the model itself, data quality issues, or sudden changes in the external environment. Based on the results of the prediction error analysis, each prediction sub-model is optimized, such as adjusting model parameters, adding or modifying features, and improving the model structure. The optimization process can combine online learning and offline retraining strategies, generate prediction intervals under different confidence levels through Monte Carlo simulation, and combine the probability distribution characteristics of the prediction error to output a comprehensive prediction result containing the optimal prediction value, uncertainty interval and confidence level, so that users can make decisions based on the reliability of the prediction.

[0031] In this embodiment, the ultra-short-term prediction sub-model inputs dynamic trend features, focusing on capturing the transient characteristics of rapid changes in irradiance; the short-term prediction sub-model combines cross-domain correlation features with numerical weather forecast data to explore the power change patterns within the daily cycle; the medium-term prediction sub-model introduces long-term trend parameters in the state assessment features to match the historical operating conditions during the same period.

[0032] In this embodiment, after generating the multi-time-scale collaborative prediction result by the weighted average strategy, the method further includes: Continuously obtaining in real time the multimodal application of the multi-time-scale collaborative prediction results; Based on the multimodal application, the application value score of the multi-time scale collaborative prediction result is calculated by the following formula: in, Score the application value of the multi-time-scale collaborative prediction results, For the multimodal application case described in The ratings under the application value rating indicators are: For the The preset scoring weights corresponding to the application value scoring indicators are: For the The preset error coefficient corresponding to each application value scoring indicator, is the total number of scoring indicators, Score the application value of the previous calculation, The maximum value of the historical application value scores calculated within the recent preset time. and is the preset base weight, is a preset constant; When the application value score is lower than a preset score threshold, extracting an optimization basis feature set from the multimodal application situation based on a preset feature extraction template; Based on the optimization basis feature set, matching the corresponding model optimization strategy from a preset model optimization strategy library; Optimizing the power prediction model based on the model optimization strategy; Based on the optimized power prediction model, the ultra-short-term prediction, short-term prediction and medium-term prediction results are re-fused on a time scale, and a new multi-time-scale collaborative prediction result is re-generated through a weighted average strategy.

[0033] In this embodiment, multimodal application refers to the performance of the power prediction model of the photovoltaic inverter in multiple application scenarios under different operating conditions. These application scenarios involve different geographical locations, climatic conditions, load conditions, etc., or prediction results on different time scales, such as ultra-short-term, short-term, and medium-term predictions. In order to more accurately predict power output, it is necessary to track and analyze the application conditions under these different conditions in real time. By collecting the operating data and external environmental data of the photovoltaic inverter in real time, a dynamic monitoring system can be constructed that can not only capture the transient response of the photovoltaic inverter, but also obtain its stability performance in various environments.

[0034] Once the multimodal application data is captured, the system calculates an application value score for the multi-timescale collaborative forecast results based on this data. The score for the application value score metric represents the performance score for a specific application scenario. The preset error coefficient for the application value score metric describes the error range or sensitivity of the metric. The weight of the application value score metric represents the degree of influence of the metric on the overall score. This weight can be adjusted based on actual needs. The previously calculated application value score represents the application value achieved during the previous time period. The maximum application value score calculated historically within the most recent preset period is used to normalize the previously calculated application value score for comparison. Two basic weight parameters are used to balance the current forecast results with historical forecast results. These weights control the weight ratio between new and historical data. A preset constant is used for final normalization to ensure the score value is within a reasonable range. Based on these calculated application value scores, the system continuously adjusts the power forecast model to achieve optimal forecast results.

[0035] When the calculated application value score falls below the preset threshold, it indicates that the existing prediction model may be biased or deficient. In this case, the system extracts optimization-based feature sets from real-time multimodal application data based on specific feature extraction templates. These feature sets contain potential model issues, such as model failure under certain environmental conditions or unstable performance within a specific time period.

[0036] The information extracted from the optimization feature set is used to match the corresponding optimization strategy from the model optimization strategy library. These optimization strategies can include adjusting algorithm parameters, updating data processing methods, and changing the weights of the weighted average strategy. Through these optimization measures, the system can improve the accuracy and reliability of the prediction model.

[0037] In the optimized prediction model, the ultra-short-term prediction, short-term prediction and medium-term prediction results will be re-fused on a time scale, and combined with a new weighted average strategy to generate new multi-time scale collaborative prediction results.

[0038] Through this process, the system continuously optimizes itself to cope with changes in different environments and load conditions, improving the accuracy of power prediction and the stability of the system.

[0039] In this embodiment, after generating the multi-time-scale collaborative prediction result by the weighted average strategy, the method further includes: The edge computing nodes deployed at the photovoltaic station side continuously receive the actual output power measurement values ​​of the photovoltaic inverters from the sensor network; The continuously received actual output power measurement values ​​of the photovoltaic inverter are compared point by point with the multi-time-scale collaborative prediction results; Based on the results of point-by-point comparison, a time series prediction error matrix is ​​constructed; Based on the constructed time series prediction error matrix, a multi-dimensional error tracing analysis is performed; Based on the results of the multi-dimensional error tracing analysis, predictive compensation is performed on the power prediction model; Based on the power prediction model after prediction compensation, the ultra-short-term prediction, short-term prediction and medium-term prediction results are re-time-scale integrated, and a new multi-time-scale collaborative prediction result is re-generated through a weighted average strategy; The multi-dimensional error tracing analysis based on the constructed time series prediction error matrix includes: Identify the correlation between external environmental data and forecast deviations; Extract the historical attenuation curve of inverter conversion efficiency and quantify the impact coefficient of device aging on power mapping relationship.

[0040] In this embodiment, edge computing nodes are deployed at the photovoltaic station to receive real-time data from a distributed sensor network. The output power of the photovoltaic inverter is a key parameter of the photovoltaic power generation system and is affected by various factors, such as environmental conditions, the operating status of the inverter itself, and equipment aging. The sensor network is responsible for collecting various operating parameters of the photovoltaic inverter in real time, such as voltage, current, frequency, temperature, and power, and transmitting this data to the edge computing nodes. The edge computing nodes not only collect data but also perform local processing and analysis.

[0041] Edge computing nodes typically possess strong processing capabilities, enabling local real-time data processing, reducing reliance on remote servers and enabling low-latency data feedback. Actual output power measurements collected through the sensor network are fundamental data for measuring the performance of photovoltaic inverters. Edge computing nodes continuously receive this real-time data, performing preprocessing and simple analysis to facilitate subsequent comparison and compensation calculations.

[0042] Once the edge computing node collects the actual output power measurements of the PV inverter, it is next necessary to compare and analyze them with the existing power prediction model. This power prediction model uses historical operating parameters, external environmental data, and the inverter's dynamic response model to predict the PV inverter's power output. However, due to various factors (such as equipment aging, environmental changes, and sudden weather changes), there may be discrepancies between the actual output power and the predicted value. Therefore, the real-time actual output power needs to be compared point by point with the model's predictions. This comparison process generates an error matrix, which contains the difference between the predicted and actual power at each point in time. The error matrix not only helps to understand the accuracy of the current model but also reveals error trends at specific time points. For example, during certain specific periods, factors such as weather changes and changes in inverter load can cause significant deviations between the predicted and actual power. These discrepancies are quantified in the error matrix, providing a data foundation for subsequent error source analysis.

[0043] After the error matrix is ​​constructed, the next key step is to conduct a multi-dimensional error source analysis. This process deeply analyzes various possible influencing factors to identify the root cause of power forecast errors. The output power of PV inverters is affected by external environmental factors such as temperature, humidity, radiation intensity, and wind speed. Therefore, it is first necessary to identify the relationship between external environmental data (such as weather changes, light intensity, and temperature fluctuations) and forecast errors. For example, weather forecast systems may contain errors, causing actual weather conditions to differ from predicted conditions, thereby affecting the inverter's power output. By analyzing the correlation between this external data and power forecast errors, the cause of the forecast inaccuracy can be identified and guidance for subsequent compensation adjustments. Inverter conversion efficiency gradually decreases over time and with equipment aging. In particular, after long-term operation, inverter performance may decline to a certain extent. This factor has a significant impact on the power mapping relationship. Therefore, by extracting the historical decay curve of the inverter conversion efficiency and quantifying its impact coefficient on power output, a more accurate basis for error analysis can be provided. Equipment aging often leads to reduced conversion efficiency, resulting in actual output power being lower than predicted. This analysis can promptly identify signs of equipment performance degradation and provide decision support for equipment maintenance or replacement.

[0044] Based on the results of multi-dimensional error source analysis, the power prediction model is compensated. The goal of compensation is to adjust the original prediction model based on the actual output power measurements and the analyzed error sources. This compensation typically involves correcting known errors and adjusting the model's input parameters or algorithms to enable the model to more accurately predict future power output.

[0045] Once the model has been compensated, the results of the ultra-short-term, short-term, and medium-term forecasts can be combined using a weighted averaging strategy. This strategy dynamically adjusts the weights of the forecasts at each time scale based on the forecast errors at each time scale, thereby generating a new multi-timescale collaborative forecast. This process not only improves the accuracy of the forecast results but also reduces the uncertainty caused by timescale variations.

[0046] After performing prediction compensation, a new power prediction model is reconstructed, and a weighted average strategy is used to generate new multi-timescale collaborative prediction results. Ultimately, these new prediction results can be provided to the photovoltaic site management system to help optimize inverter operation scheduling and fault warning.

[0047] Through edge computing nodes, high-precision and real-time photovoltaic inverter power prediction and compensation can be achieved, which can accurately identify and correct prediction errors, improve prediction accuracy, quickly respond to environmental changes and equipment status, enhance system adaptability, extend equipment life, and optimize grid scheduling through multi-time scale prediction fusion, providing strong technical support.

[0048] In order to better demonstrate the implementation of the string photovoltaic inverter power prediction method, the present invention provides a string photovoltaic inverter power prediction system, including: The distributed parameter acquisition module is configured to collect the operating parameters and environmental data of the photovoltaic inverters in real time based on the sensors deployed on each photovoltaic inverter, and perform preprocessing to generate a fused feature vector containing dynamic trend features, cross-domain correlation features, and status assessment features; a dynamic modeling module configured to screen core input variables and eliminate redundant features based on the fused feature vector, establish a basic response model between the photovoltaic inverter output power and the core input variables, and modify the basic response model based on dynamic trend features and cross-domain correlation features to generate a dynamic response model of the photovoltaic inverter; The operating condition feature library module is configured to build a multi-operating condition feature library based on historical operating data, and store multi-operating condition feature data, including core input variables, dynamic trend features, cross-domain correlation features and corresponding model parameter subsets under different operating conditions; a power evaluation module configured to determine the input power of the photovoltaic inverter based on the output power value output by the dynamic response model, construct an efficiency mapping model of the photovoltaic inverter, and dynamically calibrate the efficiency mapping model based on a multi-operating condition feature library; The multi-time-scale prediction integration module is configured to construct ultra-short-term, short-term, and medium-term prediction sub-models, perform quantitative evaluation of the power prediction results of each prediction sub-model, optimize each prediction sub-model based on the prediction error analysis results, and fuse the multi-time-scale prediction results to generate a comprehensive prediction result that includes the optimal prediction value, uncertainty interval, and confidence level.

[0049] In this embodiment, the distributed parameter acquisition module ensures data quality and feature validity, laying the foundation for modeling; the dynamic modeling module combines physical principles with data-driven correction to accurately reflect the power response under different working conditions; the working condition feature library module provides a unified benchmark to improve the system response speed and stability; the power evaluation module ensures accurate calculation of the actual output power through efficiency mapping and loss quantification; the multi-time scale prediction integration module realizes multi-dimensional precise prediction, solves problems such as data quality, model adaptability, and working condition differences, and comprehensively improves prediction accuracy and practicality.

[0050] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for predicting the power of a string photovoltaic inverter, characterized in that: include: Photovoltaic inverter parameter collection: Based on a distributed sensor network, the operating parameters of the photovoltaic inverter are collected in real time, and external environmental data is obtained at the same time; And pre-process the acquired operating parameters and external environment data; The photovoltaic inverter dynamic response model is constructed based on the pre-processed operating parameters and external environmental data to extract the transient and steady-state characteristics of the photovoltaic inverter output power as key factors change. Photovoltaic inverter power evaluation: Establishing a mapping relationship between the input power and output power of the photovoltaic inverter, and calculating the actual output power of the photovoltaic inverter based on the conversion efficiency and internal losses of the photovoltaic inverter at different working nodes; The prediction results are integrated to obtain the historical operating parameters of the string photovoltaic inverter. The real-time operating parameters and external environmental data are combined to build a power prediction model. The ultra-short-term prediction, short-term prediction and medium-term prediction results are integrated to generate multi-time-scale collaborative prediction results through a weighted average strategy.

2. The method for predicting power of a string photovoltaic inverter according to claim 1, wherein: Preprocessing of PV inverter parameter collection, including: Dynamic abnormal data processing: Build an adaptive threshold model based on the distribution characteristics of historical data. Use this model to monitor the deviation between the collected operating parameters and external environmental data in real time. Identify abnormal values ​​based on the data deviation, perform smoothing, and perform polynomial fitting interpolation or labeling on missing data. Multi-source data timing alignment: Establish a reference time axis through a distributed clock synchronization protocol to perform time alignment and synchronization of electrical parameters and external environment data; Multi-dimensional feature construction: Based on the pre-processed electrical parameters and external environment data, dynamic trend features, cross-domain correlation features, and status assessment features are generated, and quantified to generate a fusion feature vector of electrical parameters and external environment data.

3. The method for predicting the power of a string photovoltaic inverter according to claim 2, wherein: The construction of the dynamic response model of the photovoltaic inverter includes: Key variable screening and correlation analysis: Based on the fused feature vector, the correlation between each feature and the actual output power of the PV inverter is calculated. The core input variables are screened based on the preset correlation threshold, and the correlation strength between the core input variables is determined. The correlation strength is compared with the preset redundant correlation threshold to eliminate redundant features. Basic response model construction: Based on the circuit principle of photovoltaic modules, a basic response model between the output power of the photovoltaic inverter and the selected core input variables is established to obtain the basic response law of the photovoltaic inverter output power as the core input variables change; Dynamic response model optimization: Based on dynamic trend characteristics and cross-domain correlation characteristics, the deviations generated by the basic response model when predicting the output power of the photovoltaic inverter are identified and quantified. The deviations are quantified and used as correction items to correct the basic response model. The output power value of the dynamic response model of the photovoltaic inverter after correction is used as the final output.

4. The method for predicting the power of a string photovoltaic inverter according to claim 3, wherein: The construction of the photovoltaic inverter dynamic response model also includes: Collect PV inverter operating data in different seasons, weather types, and time periods, and build a multi-operating condition feature library by combining corresponding core input variables, dynamic trend features, and cross-domain correlation features. Obtain historical operating data of string photovoltaic inverters, extract operating condition features from the historical operating data, and perform similarity matching with a multi-operating condition feature library to determine different operating conditions in the historical operating data; Based on the actual output power of the photovoltaic inverter under different working conditions in the historical operating data, the adaptability of the modified dynamic response model is verified, and the deviation rate between the output power of the dynamic response model and the actual output power is calculated; When the deviation rate exceeds the preset deviation threshold, the core input variables, dynamic trend characteristics and cross-domain correlation characteristics of the corresponding working conditions are obtained, the mutual correlation strength between the core input variables is recalculated, and the parameters of the basic response model and the weights of the correction items are adjusted.

5. The method for predicting power of a string photovoltaic inverter according to claim 4, wherein: Photovoltaic inverter power evaluation, including: Basic data acquisition: Based on the output power value output by the dynamic response model and combined with the real-time collected operating parameters, the input power of the photovoltaic inverter is determined. At the same time, the output voltage, current and operating temperature data of the photovoltaic inverter AC side are collected; Constructing an efficiency mapping model: Based on the input power, AC side output voltage, current, and operating temperature data of the photovoltaic inverter, construct an efficiency mapping model for the photovoltaic inverter; Loss quantification and actual output power calculation: Decompose the internal losses of the photovoltaic inverter and combine the conversion efficiency output by the efficiency mapping model with the quantified loss value to calculate the actual output power of the photovoltaic inverter.

6. The method for predicting the power of a string photovoltaic inverter according to claim 5, wherein: PV inverter power evaluation, including dynamic calibration of mapping relationships: Based on the multi-operating condition feature library, establishing an efficiency mapping model parameter subset for different operating condition categories; Collect PV inverter operating parameters in real time and match them to corresponding working conditions, extract corresponding efficiency mapping model parameter subsets to calibrate the efficiency mapping model; When the deviation between the actual output power calculated value and the actual measured value within the sampling period exceeds a preset deviation threshold, and the sampling period exceeds a preset sampling period range, the parameter recalibration mechanism is triggered.

7. The method for predicting the power of a string photovoltaic inverter according to claim 6, wherein: The prediction results integration also includes: Based on historical and real-time operating parameters, ultra-short-term, short-term, and medium-term prediction sub-models are constructed, and the power prediction results of each prediction sub-model are quantitatively evaluated. Analyze the probability distribution characteristics and main occurrence periods of prediction errors, locate the high-incidence error conditions by combining the multi-condition feature library, and identify the key factors that cause the errors; According to the prediction error analysis results, each prediction sub-model is optimized to generate prediction intervals under different confidence levels. Combined with the probability distribution characteristics of the prediction error, a comprehensive prediction result including the optimal prediction value, uncertainty interval and confidence level is output.

8. The method for predicting power of a string photovoltaic inverter according to claim 1, wherein: After generating the multi-time-scale collaborative prediction results by the weighted average strategy, the method further includes: Continuously obtaining in real time the multimodal application of the multi-time-scale collaborative prediction results; Based on the multimodal application, the application value score of the multi-time scale collaborative prediction result is calculated by the following formula: in, Score the application value of the multi-time-scale collaborative prediction results, For the multimodal application case described in The ratings under the application value rating indicators are: For the The preset scoring weights corresponding to the application value scoring indicators are: For the The preset error coefficient corresponding to each application value scoring indicator, is the total number of scoring indicators, Score the application value of the previous calculation, The maximum value of the historical application value scores calculated within the recent preset time. and is the preset base weight, is a preset constant; When the application value score is lower than a preset score threshold, extracting an optimization basis feature set from the multimodal application situation based on a preset feature extraction template; Based on the optimization basis feature set, matching a corresponding model optimization strategy from a preset model optimization strategy library; Optimizing the power prediction model based on the model optimization strategy; Based on the optimized power prediction model, the ultra-short-term prediction, short-term prediction and medium-term prediction results are re-fused on a time scale, and a new multi-time-scale collaborative prediction result is re-generated through a weighted average strategy.

9. The method for predicting power of a string photovoltaic inverter according to claim 1, wherein: After generating the multi-time-scale collaborative prediction results by the weighted average strategy, the method further includes: The edge computing nodes deployed at the photovoltaic station side continuously receive the actual output power measurement values ​​of the photovoltaic inverters from the sensor network; The continuously received actual output power measurement values ​​of the photovoltaic inverter are compared point by point with the multi-time-scale collaborative prediction results; Based on the results of point-by-point comparison, a time series prediction error matrix is ​​constructed; Based on the constructed time series prediction error matrix, a multi-dimensional error tracing analysis is performed; Based on the results of the multi-dimensional error tracing analysis, predictive compensation is performed on the power prediction model; Based on the power prediction model after prediction compensation, the ultra-short-term prediction, short-term prediction and medium-term prediction results are re-time-scale integrated, and a new multi-time-scale collaborative prediction result is re-generated through a weighted average strategy; The multi-dimensional error tracing analysis based on the constructed time series prediction error matrix includes: Identify the correlation between external environmental data and forecast deviations; Extract the historical attenuation curve of inverter conversion efficiency and quantify the impact coefficient of device aging on power mapping relationship.

10. A string photovoltaic inverter power prediction system, applied to the string photovoltaic inverter power prediction method according to claim 1, characterized in that: include: The distributed parameter acquisition module is configured to collect the operating parameters and environmental data of the photovoltaic inverters in real time based on the sensors deployed on each photovoltaic inverter, and perform preprocessing to generate a fused feature vector containing dynamic trend features, cross-domain correlation features, and status assessment features; a dynamic modeling module configured to screen core input variables and eliminate redundant features based on the fused feature vector, establish a basic response model between the photovoltaic inverter output power and the core input variables, and modify the basic response model based on dynamic trend features and cross-domain correlation features to generate a dynamic response model of the photovoltaic inverter; The operating condition feature library module is configured to build a multi-operating condition feature library based on historical operating data, and store multi-operating condition feature data, including core input variables, dynamic trend features, cross-domain correlation features and corresponding model parameter subsets under different operating conditions; a power evaluation module configured to determine the input power of the photovoltaic inverter based on the output power value output by the dynamic response model, construct an efficiency mapping model of the photovoltaic inverter, and dynamically calibrate the efficiency mapping model based on a multi-operating condition feature library; The multi-time-scale prediction integration module is configured to construct ultra-short-term, short-term, and medium-term prediction sub-models, perform quantitative evaluation of the power prediction results of each prediction sub-model, optimize each prediction sub-model based on the prediction error analysis results, and fuse the multi-time-scale prediction results to generate a comprehensive prediction result that includes the optimal prediction value, uncertainty interval, and confidence level.

Images