Multi-mode prediction method and system for photovoltaic panel generation power, electronic equipment and medium

Abstract

The invention relates to the technical field of photovoltaic system prediction, in particular to a multi-mode prediction method and system for photovoltaic panel generation power, electronic equipment and a medium. According to the method, illumination data, atmospheric pollutant concentration data and meteorological parameters of the surface of the photovoltaic panel are obtained, and a real-time cleanliness index and an atmospheric transparency correction coefficient of the photovoltaic panel are calculated according to the illumination data, the atmospheric pollutant concentration data and the meteorological parameters; thirdly, inputting various data and parameters of the surface of the photovoltaic panel into the multi-modal prediction model, and outputting direct solar irradiance; and finally, calculating to obtain a power generation power prediction value based on the direct solar irradiance. By introducing a dual correction mechanism of a real-time cleanliness index and an atmospheric transparency correction coefficient, the multi-modal prediction model can dynamically adapt to environmental changes of different pollution degrees, and the problems of poor environmental adaptability and low prediction precision existing in a current photovoltaic power generation power prediction technology are solved. And the accuracy and environmental adaptability of photovoltaic power generation power prediction under the polluted weather condition are improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic system prediction technology, and in particular to a multimodal prediction method, system, electronic device and medium for photovoltaic panel power generation. Background Technology

[0002] With the continuous increase in the penetration rate of photovoltaic power generation in power distribution networks, the large-scale deployment of distributed photovoltaic power stations places higher demands on the accuracy of power generation forecasting. High-precision short-term forecasts not only help optimize grid dispatch and load balancing, but also provide data support for the operation and maintenance cleaning strategies of photovoltaic power stations. However, existing power forecasting methods mostly rely on traditional meteorological data and historical irradiance series, and make predictions through time series modeling or statistical learning. They generally fail to fully consider the combined attenuation effect of dust accumulation on photovoltaic panel surfaces and atmospheric pollutants (such as PM2.5 and PM10) on solar irradiance, resulting in significant prediction deviations in actual operation, especially under highly polluted weather conditions such as smog and sandstorms.

[0003] Dust accumulation directly reduces the light transmittance of photovoltaic panels. Atmospheric particulate matter not only weakens the effective irradiance reaching the ground through scattering and absorption, but also accelerates the accumulation of dirt on the panel surface, resulting in double losses. Long-term dust accumulation may also cause localized hot spots, affecting the lifespan of the modules. Some studies have attempted to introduce parameters such as aerosol optical depth (AOD) to correct for radiation attenuation along atmospheric paths. However, these methods mainly focus on atmospheric effects and do not simultaneously incorporate real-time monitoring of the panel surface contamination status, making it difficult to reflect the synergistic effect of dust and pollutants. Furthermore, traditional cleaning strategies rely on fixed cycles or manual experience, failing to dynamically respond to changes in contamination levels, easily leading to insufficient or excessive cleaning, resulting in power generation efficiency losses or resource waste.

[0004] In recent years, deep learning technology has made some progress in photovoltaic power prediction. For example, methods such as Long Short-Term Memory (LSTM) networks and Extreme Gradient Boosting (XGBoost) have been applied to temporal feature extraction and nonlinear relationship modeling. However, existing models still have limitations in data fusion, mostly modeling single factors and failing to fully integrate multi-source heterogeneous data such as atmospheric particulate matter concentration, panel dust level, meteorological parameters, and historical irradiance, lacking a collaborative representation of multimodal influences. Although some technologies have attempted to improve prediction capabilities, such as the photovoltaic power prediction method and system that integrates dust detection proposed in document 1 (CN116826703A), which calculates the transmittance through internal and external light sensors and introduces an atmospheric transparency model to assist in power prediction, its core still focuses on post-event power estimation and does not achieve real-time dynamic perception of pollution status and integrated monitoring of multiple parameters. Furthermore, it has shortcomings in adaptability, reliability, and low-power design in complex outdoor environments. Therefore, existing methods struggle to achieve long-term, stable, and coordinated sensing of dust and air pollutants in harsh environments, limiting further improvements in prediction accuracy and failing to provide a highly reliable data foundation for the intelligent operation and maintenance of photovoltaic power plants. Current photovoltaic power generation prediction technologies suffer from poor environmental adaptability and low prediction accuracy due to insufficient multimodal data fusion. Summary of the Invention

[0005] To address the aforementioned shortcomings or deficiencies, this invention provides a multimodal prediction method, system, electronic device, and medium for photovoltaic power generation, which can solve the problems of poor environmental adaptability and low prediction accuracy in current photovoltaic power generation prediction technologies.

[0006] This invention provides a multimodal prediction method for photovoltaic panel power generation, comprising: Acquire data on sunlight exposure, atmospheric pollutant concentrations, and meteorological parameters on the surface of photovoltaic panels.

[0007] The real-time cleanliness index of the photovoltaic panel is calculated based on the illumination data, and the atmospheric transparency correction coefficient is calculated based on the atmospheric pollutant concentration data and meteorological parameters.

[0008] Historical irradiance data, meteorological parameters, atmospheric pollutant concentration data, and real-time cleanliness indicators are input into a pre-defined multimodal prediction model. The model extracts the temporal characteristics of historical irradiance data and meteorological parameters, and models the nonlinear relationship between atmospheric pollutant concentration data and irradiance decay. Based on the extracted temporal characteristics and the modeled nonlinear relationship, the direct solar irradiance is output.

[0009] The direct solar irradiance is corrected based on the atmospheric transparency correction coefficient and the real-time cleanliness index to obtain the effective incident irradiance of the photovoltaic panel.

[0010] The predicted power generation value is calculated based on the effective incident irradiance of the photovoltaic panel and the electrical parameters of the photovoltaic panel.

[0011] According to a second aspect, the present invention provides a multi-mode prediction system for photovoltaic panel power generation, the system comprising: The multi-source heterogeneous data acquisition module is used to acquire solar irradiance data, atmospheric pollutant concentration data, and meteorological parameters from the surface of photovoltaic panels.

[0012] The power generation environment parameter calculation module is used to calculate the real-time cleanliness index of photovoltaic panels based on sunlight data, and to calculate the atmospheric transparency correction coefficient based on atmospheric pollutant concentration data and meteorological parameters.

[0013] The direct irradiance calculation module is used to input historical irradiance data, meteorological parameters, atmospheric pollutant concentration data, and real-time cleanliness indicators into a preset multimodal prediction model. Through the multimodal prediction model, the time-series characteristics of historical irradiance data and meteorological parameters are extracted, and the nonlinear relationship between atmospheric pollutant concentration data and irradiance decay is modeled. Based on the extracted time-series characteristics and the modeled nonlinear relationship, the direct solar irradiance is output.

[0014] The effective incident irradiance calculation module is used to correct the direct solar irradiance based on the atmospheric transparency correction coefficient and the real-time cleanliness index to obtain the effective incident irradiance of the photovoltaic panel.

[0015] The power generation prediction module is used to calculate the predicted power generation value based on the effective incident irradiance of the photovoltaic panel and the electrical parameters of the photovoltaic panel.

[0016] According to a third aspect, the present invention provides an electronic device comprising: At least one processor; and The memory that is communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform any of the multimodal prediction methods for photovoltaic power generation in the embodiments of the present invention.

[0017] According to another aspect of the present invention, a non-transient computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to execute a multimodal prediction method for photovoltaic power generation in any of the embodiments of the present invention.

[0018] The present invention provides a multimodal prediction method for photovoltaic (PV) panel power generation. The method first simultaneously acquires multi-dimensional information such as irradiance data, atmospheric pollutant concentration data, and meteorological parameters on the PV panel surface. Then, based on the irradiance data, a real-time cleanliness index of the PV panel, quantifying the impact of surface dust, is calculated. Based on the atmospheric pollutant concentration data and meteorological parameters, an atmospheric transparency correction coefficient, characterizing the attenuation effect of atmospheric particulate matter, is calculated using an aerosol optical thickness estimation model. Next, historical irradiance data, meteorological parameters, atmospheric pollutant concentration data, and the real-time cleanliness index are input into a preset multimodal prediction model. This model extracts the temporal dependence features of historical irradiance and meteorological conditions and models the complex nonlinear relationship between atmospheric pollutant concentration and irradiance attenuation. Based on the feature fusion results, a predicted value of direct solar irradiance is output. Then, the atmospheric transparency correction coefficient and the real-time cleanliness index are used to collaboratively correct atmospheric path attenuation and panel surface contamination attenuation, respectively, to obtain an accurate effective incident irradiance of the PV panel. Finally, combined with the PV panel's own electrical parameters, a high-precision short-term power generation prediction value is calculated.

[0019] In this technical solution, the present invention addresses the problem that existing prediction methods, as described in the background section, fail to comprehensively consider the coupling effect between atmospheric pollutants and aerosol particle dust. By introducing a dual correction mechanism of real-time cleanliness index and atmospheric transparency correction coefficient, and utilizing a multimodal prediction model to fuse temporal characteristics and nonlinear relationships, it achieves collaborative modeling and accurate compensation for the combined attenuation effect of dust deposition and aerosol particles. Therefore, the technical solution of this invention, through the introduction of a dual correction mechanism of real-time cleanliness index and atmospheric transparency correction coefficient, enables the model to dynamically adapt to environmental changes with different pollution levels. This solves the problems of poor environmental adaptability and low prediction accuracy in current photovoltaic power generation prediction technologies, improving the accuracy and environmental adaptability of photovoltaic power generation prediction under polluted weather conditions. Attached Figure Description

[0020] Figure 1 This is a structural block diagram of a photovoltaic panel surface dust contamination detection device according to an embodiment of the present invention; Figure 2 This is a flowchart of a multimodal prediction method for photovoltaic power generation according to an embodiment of the present invention; Figure 3 This is a flowchart of the adaptive weighting process of an LSTM-XGBoost hybrid model according to an embodiment of the present invention; Figure 4 This is a schematic diagram of an AOD estimation neural network model according to an embodiment of the present invention; Figure 5 This is a structural block diagram of a multimodal prediction system for photovoltaic panel power generation according to an embodiment of the present invention; Figure 6This is a block diagram of an electronic device used to implement embodiments of the present invention. Detailed Implementation

[0021] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0022] According to a first aspect, the present invention provides a multimodal prediction method for the power generation of a photovoltaic panel, which is executed by an intelligent monitoring device (hereinafter referred to as "the device") installed on the photovoltaic panel, which can be a power generation device in a distributed photovoltaic power generation system or a centralized photovoltaic power station.

[0023] The device should be able to run machine learning models and algorithms through local deployment or edge computing to perform real-time analysis and processing of multi-source heterogeneous data. It also integrates a multi-sensor fusion module to achieve collaborative monitoring and intelligent diagnosis of light intensity, atmospheric pollutant concentration, meteorological parameters, and component status. Specifically, the power equipment in the device includes, but is not limited to, self-powered photovoltaic units, a power management module, a data acquisition unit, and a communication transmission module.

[0024] In some embodiments, such as Figure 1 The diagram shows the structural block diagram of the photovoltaic panel surface dust pollution detection device of the present invention. As shown in the figure, the device collects illuminance data on the outer and inner sides of the photovoltaic glass panel through an external illuminance sensor and an internal illuminance sensor, respectively. By comparing and analyzing the difference in illuminance between the inner and outer sides, the surface cleanliness of the photovoltaic panel is calculated in real time, realizing a quantitative assessment of the degree of dust accumulation. At the same time, the air quality monitoring module collects multi-dimensional environmental parameters such as wind speed, temperature, weather conditions, latitude, longitude, altitude, and historical irradiance, and combines them with particulate matter concentration data, which are then input into the LSTM-XGBoost hybrid model and the AOD estimation model based on PM concentration, respectively. The LSTM-XGBoost model is used to extract the temporal characteristics of historical irradiance and meteorological parameters and model the nonlinear relationship between pollutants and irradiance attenuation to predict solar irradiance. The AOD estimation model calculates the atmospheric transparency correction coefficient through aerosol optical thickness. Finally, after correcting the predicted solar irradiance for atmospheric transparency and surface cleanliness, the effective direct irradiance actually received by the photovoltaic panel is obtained, and a high-precision short-term power generation prediction result is output by combining the electrical parameters of the photovoltaic panel. This process embodies the core technical path of multi-source data perception, model collaborative computation, and dual attenuation correction.

[0025] like Figure 2As shown, the method may include: Step S110: Obtain illumination data, atmospheric pollutant concentration data, and meteorological parameters on the surface of the photovoltaic panel.

[0026] Among them, the light intensity data can include the light intensity on the outside of the photovoltaic panel collected by an external light sensor and the light intensity on the inside of the photovoltaic panel collected by a built-in light sensor; the atmospheric pollutant concentration data mainly refers to the mass concentration of aerosol particulate matter such as PM2.5 and PM10; meteorological parameters include wind speed, wind direction, ambient temperature and relative humidity.

[0027] Specifically, the external light sensor uses a silicon photodiode (such as the BH1750 model) vertically mounted on the frame of the photovoltaic panel, while the internal sensor uses the same model and is mounted close to the back of the photovoltaic glass. Both are set to a sampling frequency of 1 Hz. Particulate matter concentration monitoring uses a laser scattering principle sensor (such as the SDS011 model), wind speed and direction are collected using an ultrasonic meteorological sensor (such as the WM030 model), and temperature and humidity are measured using a digital composite sensor (such as the SHT35 model).

[0028] For example, the device polls and collects data from each sensor via an RS485 bus, with a sampling interval of 5 minutes. Each acquisition lasts for 30 seconds, and the arithmetic mean is taken as the valid sample value. The collected raw data is first subjected to outlier filtering (using the Laida criterion to remove data that is more than 3 times the standard deviation), and missing values ​​are filled by linear interpolation of the data from the two preceding and following periods.

[0029] Step S120: Calculate the real-time cleanliness index of the photovoltaic panel based on the illumination data, and calculate the atmospheric transparency correction coefficient based on the atmospheric pollutant concentration data and meteorological parameters.

[0030] The real-time cleanliness index quantifies the light transmittance loss caused by dust accumulation on the photovoltaic panel surface, while the atmospheric transparency correction coefficient characterizes the attenuation effect of atmospheric aerosols on solar radiation. Specifically, the cleanliness index is calculated as the ratio of real-time transmittance to a baseline transmittance. The baseline transmittance is determined by averaging the ratio of internal and external irradiance over three consecutive sunny days with clean photovoltaic panels. The atmospheric transparency correction coefficient is calculated using an AOD estimation model, a three-layer feedforward neural network. Input features include PM2.5 concentration (unit: PM10 concentration (unit: ), relative humidity (unit: percentage) and temperature (unit: degrees Celsius).

[0031] For example, the device calculates the cleanliness index using the following formula (1): ; (1) in, The light intensity is 850 watts per square meter, collected by an external sensor. The light intensity is 769 watts per square meter, collected by the built-in sensor. The real-time cleanliness index was calculated based on a pre-calibrated baseline transmittance of 0.95. When calculating the atmospheric transparency correction factor, the measured PM2.5 concentration (35) will be used. PM10 concentration (68) The atmospheric transparency correction coefficient of 0.89 is obtained by inputting the relative humidity (45%) and temperature (28 degrees Celsius) into the trained neural network model.

[0032] Step S130: Input historical irradiance data, meteorological parameters, atmospheric pollutant concentration data and real-time cleanliness index into the preset multimodal prediction model. Through the multimodal prediction model, extract the time series characteristics of historical irradiance data and meteorological parameters, and model the nonlinear relationship between atmospheric pollutant concentration data and irradiance decay. Output the direct solar irradiance based on the extracted time series characteristics and the modeled nonlinear relationship.

[0033] The multimodal prediction model employs a hybrid LSTM-XGBoost architecture. The LSTM branch handles temporal features, while the XGBoost branch models nonlinear relationships. Weight adjustments are based on a segmented strategy for real-time PM2.5 concentration, with specific thresholds determined through optimization using historical data. This approach leverages the strengths of LSTM in capturing long-term dependencies in time series data, making it suitable for temporal features of irradiance and meteorological parameters. XGBoost, on the other hand, efficiently handles high-dimensional static features (such as pollutant concentrations) and prevents overfitting through regularization. Furthermore, the LSTM algorithm, with its gating mechanism, is well-suited for extracting temporal patterns from irradiance data, while the XGBoost algorithm effectively models the nonlinear mapping between pollutant concentration and irradiance decay through gradient boosting decision trees. The fusion of the two branches balances dynamic temporal and static nonlinear relationships, achieving complementary advantages and enhancing prediction robustness.

[0034] Specifically, the LSTM branch input consists of historical irradiance sequences and meteorological parameter sequences from the past 24 hours with a time resolution of 15 minutes. A single-layer LSTM network with 32 hidden units is used to extract temporal features. The XGBoost branch input consists of pollutant concentration data and cleanliness indicators at the current moment. The maximum tree depth is set to 6, and the learning rate is 0.1.

[0035] For example, the model training uses historical data from the past 6 months, inputting an irradiance sequence (24 hours long) at 72 consecutive time points into the LSTM branch, while simultaneously inputting the current PM2.5 concentration (28... Features such as cleanliness index (92%) are input into the XGBoost branch. The outputs of the two branches are fused through dynamic weights (currently set to LSTM weight 0.7 and XGBoost weight 0.3) to output the predicted value of direct solar irradiance for the next hour as 836 watts per square meter.

[0036] In some embodiments, the device can calculate the direct irradiance received by the photovoltaic panel using the following formula (2): (2) Formula (2) is a simplified direct irradiance calculation model used to quantify the attenuation of solar radiation by aerosol extinction effect. This indicates the direct irradiance received by the photovoltaic panel (unit: watts per square meter). The aerosol extinction transmittance is dimensionless. The relevant solar irradiance reaching the top of the atmosphere (unit: watts per square meter). This model focuses on the key attenuation path caused by aerosols, correcting the theoretical solar irradiance by aerosol extinction transmittance to obtain the actual direct irradiance reaching the photovoltaic panel surface.

[0037] Next, the device can calculate the transmittance of the aerosol extinction band using the following formula (1): (3) Formula (3) is an aerosol transmittance calculation model based on the Beer-Lambert law, used to quantify the attenuation of solar radiation by aerosols in a specific band. The transmissivity (dimensionless) represents the wavelength at which aerosol extinction occurs, and exp is a natural exponential function. denoted as α, where α is the optical quality of aerosol extinction (dimensionless), AOD is the optical thickness of aerosol (dimensionless), and τ(t) is the time-dependent band adjustment factor (dimensionless).

[0038] Furthermore, the device can calculate the optical quality of aerosol extinction using the empirical formula (4) shown below. : (4) The exponents and coefficients 0.16851, 0.18198, 95.318, and 1.9542 are empirical fitting parameters. By introducing a nonlinear correction term related to the zenith angle Z, this model can more accurately describe the radiative transfer path under actual atmospheric conditions compared to the traditional secant function sec(Z) calculation method. This provides key parameters for the subsequent precise application of aerosol optical thickness and atmospheric transparency correction.

[0039] Step S140: Correct the direct solar irradiance based on the atmospheric transparency correction coefficient and the real-time cleanliness index to obtain the effective incident irradiance of the photovoltaic panel.

[0040] The correction process includes two sequential steps: atmospheric attenuation correction and surface pollution attenuation correction.

[0041] Specifically, the device first uses an atmospheric transparency correction coefficient to correct for atmospheric path attenuation of the predicted irradiance, and then uses a cleanliness index to correct for surface light transmittance loss. The two corrections employ a multiplicative model to reflect the cascade relationship of the attenuation effect.

[0042] For example, the device first multiplies the predicted direct solar irradiance of 836 watts per square meter by an atmospheric transparency correction factor of 0.89 to obtain an atmospherically corrected irradiance of 745 watts per square meter, and then multiplies it by a cleanliness index of 95.2% (equivalent to 0.952) to finally obtain an effective incident irradiance of 709 watts per square meter for the photovoltaic panel.

[0043] Step S150: Calculate the predicted power generation value based on the effective incident irradiance of the photovoltaic panel and the electrical parameters of the photovoltaic panel.

[0044] Among them, the electrical parameters include inherent parameters such as the rated power of the photovoltaic panel, temperature coefficient, and nominal efficiency. The temperature coefficient can be determined according to the photovoltaic panel temperature characteristic test in the IEC 61215 standard.

[0045] In some embodiments, the device can calculate the predicted power generation value of the photovoltaic panel using the following formula (5): (5) Formula (5) is a basic photoelectric conversion model that comprehensively considers irradiance, temperature and electrical parameters, and is used to characterize the basic physical relationship of photovoltaic panels converting light energy into electrical energy. This represents the short-term power generation forecast of the photovoltaic panel (unit: watts). S represents the photoelectric conversion efficiency of the photovoltaic panel (dimensionless), and S represents the effective light-receiving area of ​​the photovoltaic panel (unit: square meters). Effective incident irradiance of photovoltaic panels (unit: watts per square meter). The value represents the operating temperature of the photovoltaic panel (in degrees Celsius), and the coefficient 0.005 is a typical empirical coefficient for the influence of temperature. This model incorporates the influence of module operating temperature... The relevant linear correction term quantifies the effect of temperature increase on the decrease in photoelectric conversion efficiency, providing a core calculation basis for accurate prediction of power generation.

[0046] Next, the device can calculate the operating temperature of the photovoltaic panel using the following formula (6): (6) Formula (6) is a photovoltaic panel operating temperature estimation model based on thermal balance, used to quantify the component temperature rise effect under the combined effect of ambient temperature and solar irradiance. The photovoltaic panel's operating temperature (in degrees Celsius) is represented by T, the ambient air temperature (in degrees Celsius), and k is the temperature coefficient characterizing the temperature rise caused by solar irradiation (in degrees Celsius). ), The direct irradiance received by the photovoltaic panel (unit: watts per square meter). The temperature coefficient k is an empirical parameter related to the module material and installation conditions, with a typical value range of 0.02 to 0.035. This model reflects the mechanism by which the ambient base temperature and the irradiation heating effect jointly determine the module's operating temperature through linear superposition, providing key input parameters for temperature correction in subsequent power generation prediction.

[0047] Alternatively, in other embodiments, the device can calculate the predicted output power of the photovoltaic panel using the following formula (7): (7) Formula (7) is a photoelectric conversion optimization model that integrates dynamic cleanliness factor and temperature compensation effect, used to accurately characterize the photovoltaic power generation under the combined effect of dust deposition and temperature change. This represents the short-term power output forecast of the photovoltaic panel (unit: watts). A dynamic cleanliness index that changes over time (dimensionless). S represents the nominal photoelectric conversion efficiency (dimensionless) of the photovoltaic panel, and S represents the effective light-receiving area of ​​the photovoltaic panel (unit: square meters). Let denoted as η(t), be the direct irradiance received by the photovoltaic panel (watts per square meter), T be the ambient air temperature (degrees Celsius), and k be the irradiation temperature coefficient (degrees Celsius·m² / watt). This model quantifies the optical loss caused by dust deposition by introducing a dynamic cleanliness factor η(t) and uses a linear temperature correction term. Reflects component operating temperature The impact on conversion efficiency enables accurate prediction of photovoltaic output power under complex environments.

[0048] Therefore, according to the above implementation method, the device first simultaneously acquires multi-dimensional information such as irradiance data, atmospheric pollutant concentration data, and meteorological parameters on the surface of the photovoltaic panel; then, it calculates a real-time cleanliness index of the photovoltaic panel based on the irradiance data to quantify the impact of surface dust, and calculates an atmospheric transparency correction coefficient characterizing the attenuation effect of atmospheric particulate matter based on the atmospheric pollutant concentration data and meteorological parameters through an aerosol optical thickness estimation model; further, it inputs historical irradiance data, meteorological parameters, atmospheric pollutant concentration data, and real-time cleanliness index into a preset multimodal prediction model, extracts the time-series dependence features of historical irradiance and meteorological conditions through the model, and models the complex nonlinear relationship between atmospheric pollutant concentration and irradiance attenuation, outputting a predicted value of direct solar irradiance based on the feature fusion result; next, it uses the atmospheric transparency correction coefficient and the real-time cleanliness index to perform synergistic correction of atmospheric path attenuation and panel surface contamination attenuation to obtain an accurate effective incident irradiance of the photovoltaic panel; finally, it calculates a high-precision short-term power generation prediction value by combining the electrical parameters of the photovoltaic panel itself.

[0049] Throughout the process, this embodiment addresses the problem described in the background art that existing prediction methods fail to comprehensively consider the coupling effect between atmospheric pollutants and aerosol particle dust. By introducing a dual correction mechanism of real-time cleanliness index and atmospheric transparency correction coefficient, and utilizing a multimodal prediction model to fuse temporal characteristics and nonlinear relationships, it achieves collaborative modeling and accurate compensation for the combined attenuation effect of dust deposition and aerosol particles. Therefore, the technical solution of this embodiment, by introducing a dual correction mechanism of real-time cleanliness index and atmospheric transparency correction coefficient, enables the model to dynamically adapt to environmental changes with different pollution levels. This solves the problems of poor environmental adaptability and low prediction accuracy in current photovoltaic power generation prediction technologies, improving the accuracy and environmental adaptability of photovoltaic power generation prediction under polluted weather conditions.

[0050] In some embodiments, a pollutant detection device is installed on the photovoltaic panel. The pollutant detection device is connected to the power output terminal of the photovoltaic panel. The pollutant detection device is equipped with an external light sensor, a built-in light sensor, a particulate matter concentration sensor, and a meteorological sensor; it acquires light data, air pollutant concentration data, and meteorological parameters from the surface of the photovoltaic panel, including: The surface light intensity of the photovoltaic panel is collected by an external light sensor, and the internal light intensity of the photovoltaic panel is collected by an internal light sensor. The surface light intensity and the internal light intensity are used as light data.

[0051] The external light sensor, a silicon photodiode (model BH1750), is mounted on the upper surface of the photovoltaic panel frame, 8 mm vertically from the glass panel, with a measurement range of 0-2000 watts per square meter and a sampling interval of 5 minutes. The built-in sensor of the same model is mounted on the back of the photovoltaic glass, and both transmit data synchronously via an RS485 bus.

[0052] Particulate matter concentration information in the atmospheric environment is collected by a particulate matter concentration sensor to obtain atmospheric pollutant concentration data.

[0053] The particulate matter concentration sensor (model SDS011) can be based on the laser scattering principle and installed on the windward support of the photovoltaic array to monitor PM2.5 (range 0-500). ) and PM10 (range 0~1000) The concentration is controlled by an insect-proof net and a miniature air pump to maintain a stable sampling airflow, with a sampling cycle of 15 minutes.

[0054] Atmospheric environmental parameters, including wind speed, wind direction, temperature, and humidity, are collected by meteorological sensors and used as meteorological parameters.

[0055] Among them, the meteorological sensor can be an integrated ultrasonic device (model WM030-ETH), installed at a height of 2 meters above the ground, with a wind speed measurement range of 0~60 meters per second (accuracy ±0.3 meters per second), a wind direction measurement range of 0~360 degrees (accuracy ±2 degrees), a temperature range of -40~85 degrees Celsius, and a relative humidity range of 0~100%RH (relative humidity, dimensionless).

[0056] Therefore, according to the above implementation method, the device can achieve synchronous acquisition of light, pollutant and meteorological data through multi-sensor collaborative acquisition.

[0057] In some embodiments, a real-time cleanliness index of the photovoltaic panel is calculated based on illumination data, and an atmospheric transparency correction coefficient is calculated based on atmospheric pollutant concentration data and meteorological parameters, including: The initial light intensity was obtained by the external light sensor and the internal light sensor when the photovoltaic panel was in a dust-free state.

[0058] The "dust-free state" refers to the condition where no visible contaminants are deposited on the surface of the photovoltaic panels after professional cleaning for three consecutive sunny days. When collecting initial light intensity data, the device must be used under conditions where the solar altitude angle is greater than 30 degrees and atmospheric transparency is good. Each data collection session lasts for 5 minutes, and the average value is taken.

[0059] The baseline transmittance of the photovoltaic panel is calculated based on the initial light intensity.

[0060] Specifically, the device can calculate the reference transmittance using the following formula (8): ; (8) in, The initial surface illumination intensity is collected by an external sensor. This represents the initial internal illumination intensity collected by the built-in sensor. The arithmetic mean of three consecutive valid data sets is used as the final baseline value during calculation.

[0061] The surface light intensity and internal light intensity are collected in real time by the external light sensor and the internal light sensor. The real-time transmittance is calculated based on the real-time collected surface light intensity and internal light intensity.

[0062] Specifically, the device can use the above formula (8) to calculate the real-time transmittance. At this time, For real-time surface illumination intensity, Real-time internal light intensity and real-time transmittance are available. The data collection frequency is set to once every 5 minutes, and three consecutive collections are taken, with the average value taken to eliminate the impact of instantaneous fluctuations.

[0063] The real-time cleanliness index is calculated based on the ratio of the real-time transmittance to the reference transmittance.

[0064] For example, when the real-time transmittance is measured Reference transmittance At that time, the cleanliness index calculated by the device is: This metric is transmitted to the data processor via the Modbus RTU protocol (a serial communication protocol used in industrial control) and stored in an SD card (a portable, pluggable flash memory storage device) for subsequent analysis.

[0065] Atmospheric pollutant concentration data and meteorological parameters are used as input parameters and input into a preset aerosol optical thickness estimation model. The aerosol optical thickness estimation model processes the input parameters and outputs the current atmospheric transparency correction coefficient.

[0066] For example, the device will collect PM2.5 concentration (38) in real time. The relative humidity (45%) and temperature (25 degrees Celsius) are input into the trained BP neural network model, and the model outputs an AOD value of 0.21. Then, the atmospheric transparency correction coefficient of 0.89 is calculated by the following formula (9): (9) The typical value for airmass is 1.5.

[0067] Therefore, according to the above embodiments, the device can accurately quantify the degree of pollution on the surface of the photovoltaic panel through light intensity difference analysis, and accurately assess the irradiance attenuation caused by air pollution based on a multi-parameter fusion machine learning model.

[0068] In some embodiments, the multimodal prediction model is configured with a first prediction branch and a second prediction branch; the multimodal prediction model extracts the temporal features of historical irradiance data and meteorological parameters, models the nonlinear relationship between atmospheric pollutant concentration data and irradiance decay, and outputs direct solar irradiance based on the extracted temporal features and the modeled nonlinear relationship, including: The first prediction branch extracts the time-series features of historical irradiance data and meteorological parameters to obtain the first prediction result.

[0069] Specifically, the first prediction branch can adopt an LSTM structure, with inputs including historical irradiance sequences (unit: watts per square meter) over the past 24 hours at a time resolution of 15 minutes and meteorological parameter sequences (including temperature, humidity, and wind speed). The LSTM network is configured with 32 hidden units, which learn temporal dependencies through its gating mechanism, and finally outputs the first prediction result for the next hour through a fully connected layer.

[0070] The second prediction branch is used to model the nonlinear relationship between atmospheric pollutant concentration data and irradiance decay to obtain the second prediction result.

[0071] Specifically, the second prediction branch can employ the XGBoost algorithm, with input features including the current PM2.5 concentration (unit: The model uses instantaneous values ​​of PM10 concentration, real-time cleanliness index (unit: percentage), and meteorological parameters. The maximum tree depth is set to 6, and the learning rate is 0.1. Gradient boosting of the decision tree fits the complex nonlinear mapping between pollutant concentration and irradiance decay.

[0072] In fact, the second prediction branch of the multimodal prediction model uses the XGBoost algorithm, based on its unique technical advantages and the matching consideration of the branch's functional positioning. For example... Figure 3In the illustrated LSTM-XGBoost hybrid model architecture, "branch" specifically refers to the functional module division within the model: the first branch (LSTM) is responsible for processing time-series data, while the second branch is specifically used to model the complex nonlinear relationship between static features such as pollutant concentration and cleanliness indicators and irradiance decay. The XGBoost algorithm, as the specific machine learning tool for implementing this branch's function, efficiently handles high-dimensional discrete features through its gradient boosting decision tree structure and suppresses overfitting through regularization, perfectly meeting the second branch's requirement for accurate modeling of nonlinear relationships. This "branch-algorithm" correspondence reflects the synergy between structural layering (branch defining functional scope) and technology selection (algorithm implementing specific functions) in model design. It is not a conflict of terminology, but rather a targeted design for optimizing prediction accuracy.

[0073] The first and second prediction results are dynamically weighted and fused to obtain the fused prediction value.

[0074] For example, the dynamic weighting strategy is adjusted based on the current atmospheric pollutant concentration level: when the PM2.5 concentration is below 35... When the PM2.5 concentration is above 75, the weight of the first prediction branch is set to 0.8, and the weight of the second prediction branch is 0.2; At that time, the weight of the first branch was adjusted to 0.4, and the weight of the second branch was increased to 0.6 to enhance the representation weight of the pollutant impact. The weighting formula is: the fused prediction value is... ,in, , As weight, , These are the prediction results for the two branches, respectively.

[0075] The fused predictions are used as the output of the predicted direct solar irradiance.

[0076] For example, at a certain prediction time, the first branch outputs a predicted irradiance of 850 watts per square meter, and the second branch outputs 820 watts per square meter. The device adjusts the irradiance based on the current PM2.5 concentration (50 watts per square meter). The dynamic weights are calculated as follows: , The final fusion prediction value is: (Watts per square meter). The result is encapsulated in JSON (JavaScript Object Notation) format and written to a real-time database.

[0077] Therefore, according to the above embodiments, the device can simultaneously capture the temporal variation of irradiance and the nonlinear decay effect of pollutants through a dual-branch collaborative prediction mechanism.

[0078] In some embodiments, the direct solar irradiance is corrected based on an atmospheric transparency correction factor and a real-time cleanliness index to obtain the effective incident irradiance of the photovoltaic panel, including: The predicted direct solar irradiance is corrected for atmospheric attenuation using an atmospheric transparency correction coefficient, resulting in the irradiance after atmospheric attenuation correction.

[0079] Specifically, the atmospheric attenuation correction uses a multiplicative model to reflect the scattering and absorption of solar radiation by aerosol particles in the atmospheric path. Multiplying the predicted direct solar irradiance from the multimodal prediction model by the atmospheric transparency correction factor yields the theoretical irradiance before reaching the photovoltaic panel surface.

[0080] The effective incident irradiance of the photovoltaic panel is obtained by multiplying the irradiance corrected for atmospheric attenuation with the real-time cleanliness index.

[0081] Specifically, this step is used to correct for optical losses caused by dust accumulation on the photovoltaic panel surface. The cleanliness index is used in the calculation as a percentage and needs to be converted to a decimal coefficient (i.e., divided by 100) before multiplication to finally obtain the effective irradiance that can be used for photoelectric conversion.

[0082] For example, suppose that at a certain moment the model predicts a direct solar irradiance of 900 watts per square meter, the current atmospheric transparency correction factor is 0.85, and the real-time cleanliness index is 88%. Then the calculation process of the device is as follows: 1. The irradiance after atmospheric attenuation correction is: 900 (watts per square meter) × 0.85 = 765 (watts per square meter); 2. The effective incident irradiance of the photovoltaic panel is: 765 (watts per square meter) × (88 / 100) = 765 watts per square meter × 0.88 = 673.2 (watts per square meter).

[0083] Therefore, according to the above embodiments, the device can accurately quantify the combined impact of atmospheric path attenuation and surface contamination attenuation on solar irradiance through two-stage series correction.

[0084] In some embodiments, the electrical parameters of the photovoltaic panel include photoelectric conversion efficiency, light-receiving area, and temperature influence coefficient; the predicted power generation value is calculated based on the effective incident irradiance of the photovoltaic panel and the electrical parameters of the photovoltaic panel, including: The initial power generation is calculated based on the effective incident irradiance, photoelectric conversion efficiency, and light-receiving area of ​​the photovoltaic panel.

[0085] Specifically, the initial power generation is calculated using the following physical model (expressed as Equation (10)): (10) in, A represents the nominal photoelectric conversion efficiency of the photovoltaic panel (unit: percentage), and A represents the effective light-receiving area of ​​the photovoltaic panel (unit: square meters). This represents the effective incident irradiance of the photovoltaic panel (unit: watts per square meter). The percentage value of the photoelectric conversion efficiency must be converted to decimal form for calculation.

[0086] The current ambient temperature data is obtained, and the initial power generation is corrected by temperature compensation based on the temperature influence coefficient to obtain the compensated power generation.

[0087] Specifically, the temperature compensation correction is calculated using the following linear model (expressed as Equation (11)): (11) Where γ is the temperature effect coefficient (unit: per degree Celsius). The current ambient temperature (unit: degrees Celsius). The standard test condition temperature is 25 degrees Celsius. This correction reflects the characteristic that the power generation efficiency of photovoltaic panels decreases as temperature increases.

[0088] The compensated power generation is used as the predicted power generation value.

[0089] For example, for a monocrystalline silicon photovoltaic panel with a rated power of 320 watts ( (A=1.6 square meters), when the effective incident irradiance is measured It is 613.36 watts per square meter, with an ambient temperature of At a temperature of 35 degrees Celsius, with a temperature influence coefficient γ = -0.004 / degree Celsius, the initial power generation is: The temperature compensation correction amount is: The power generation after compensation is: .

[0090] Therefore, according to the above embodiments, the device can achieve accurate prediction of power generation by combining physical models with temperature compensation.

[0091] In some embodiments, before inputting historical irradiance data, meteorological parameters, air pollutant concentration data, and real-time cleanliness indicators into a preset multimodal prediction model, the method further includes: The training dataset is constructed by acquiring sample irradiance data, sample meteorological parameters, sample atmospheric pollutant concentration data, and sample cleanliness indicators within a preset historical time period.

[0092] Generally, the device can select a continuous time period that includes different seasons and weather conditions as the historical time period, such as complete data for the most recent 12 months. The temporal resolution of the sample data is set to 15 minutes, and each sample point includes the irradiance, temperature, humidity, wind speed, PM2.5 concentration, PM10 concentration, and cleanliness index at that moment.

[0093] The training dataset is divided into a training set, a validation set, and a test set.

[0094] Specifically, the device can be divided into three sets: a training set for learning model parameters, a validation set for hyperparameter tuning and training process monitoring, and a test set for final evaluation of the model's generalization ability. The data must be divided in chronological order to prevent future information leakage.

[0095] Construct an initial multimodal prediction model that includes a temporal feature extraction module and a nonlinear relationship modeling module.

[0096] For example, the time-series feature extraction module uses a single-layer LSTM network with 32 hidden units to process the irradiance and meteorological parameter sequences for a historical 24-hour period (96 time points); the nonlinear relationship modeling module uses the XGBoost algorithm with a maximum tree depth of 6 and a learning rate of 0.1 to process static features such as pollutant concentration and cleanliness indicators. In the hybrid model architecture, the "nonlinear relationship modeling module" refers to a functional unit specifically handling the complex mapping between pollutant concentration and irradiance decay, and the "XGBoost algorithm" is the core technical tool for implementing this module. Figure 3 As shown, the nonlinear relationship modeling module, through the gradient boosting decision tree structure of XGBoost, can effectively learn the nonlinear relationship between static features such as PM2.5 concentration and cleanliness index and irradiance attenuation. Its regularization characteristics can prevent overfitting. This "module-algorithm" correspondence reflects the hierarchical design of functional division and implementation methods, and the two logically form a complete technical closed loop.

[0097] The initial multimodal prediction model is iteratively trained using the training set, and the model performance is monitored during the training process using the validation set.

[0098] For example, the training batch size is set to 32, the optimizer uses Adam (adaptive moment estimation), the initial learning rate is set to 0.001, and the training epochs are 100. After each training epoch, the mean square error (MSE) is calculated on the validation set as a monitoring metric. The device can calculate the MES using the following formula (12): (12) in, For the true value, is the predicted value, and n is the number of samples.

[0099] When the model evaluation metrics on the validation set reach the preset convergence condition, the training process of the initial multimodal prediction model is terminated, and the trained multimodal prediction model is obtained.

[0100] For example, the preset convergence condition can be set so that the mean squared error of the validation set no longer decreases after 10 consecutive training epochs (early stopping). When the mean squared error of the validation set is 45.2 watts per square meter after the 85th training epoch, and fluctuates within the range of 45.2 ± 0.5 for the subsequent 10 epochs, training is terminated and the current model parameters are saved. The final mean squared error evaluated on the test set is 48.7 watts per square meter, and the mean absolute percentage error is 5.3%.

[0101] In some embodiments, Figure 3 The diagram illustrates an embodiment of the LSTM-XGBoost hybrid model's adaptive weighting process. The model training begins with loading the original dataset, which contains multi-dimensional time-series data from the past six months, with a time resolution of 15 minutes, including historical irradiance, meteorological parameters, pollutant concentrations, and cleanliness indices. Data preprocessing is then performed, including outlier removal (using the 3σ criterion), linear imputation of missing values, and max-min normalization, forming standardized training samples. The preprocessed data is then input in parallel to the XGBoost and LSTM model construction branches: the XGBoost branch takes current-time features (such as PM2.5 concentration and cleanliness indices) as input, sets the maximum tree depth to 6, and the learning rate to 0.1, outputting prediction result 1 after training; the LSTM branch takes the past 24-hour historical sequence (96 time points) as input, configures 32 hidden units, and outputs prediction result 2. The prediction results from both branches enter the adaptive weighting stage. The weights are dynamically adjusted based on the mean squared error of the validation set over the first five rounds (e.g., XGBoost weights α iterate from 0.3 to 0.7). The optimal weight combination is determined by checking if it is optimal (convergence criterion: a decrease in validation set error of less than 1% over 10 consecutive rounds). If optimal, the process returns to the residual weighting improvement stage, recalculating the weight allocation based on the prediction residuals. If optimal, the weights are fixed (e.g., final α = 0.6, LSTM weight β = 0.4), and the process enters the LSTM-XGBoost hybrid prediction stage. Finally, the prediction results are evaluated using the root mean square error (RMSE) and mean absolute percentage error (MAPE). For example, on the test set, the accuracy metrics are RMSE ≤ 50 W / m² and MAPE ≤ 5.5%, completing the entire training and validation process of the hybrid model.

[0102] In some embodiments, Figure 4 The figure shows a schematic diagram of an AOD estimation neural network model according to an embodiment of the present invention. As shown, the model is a typical three-layer feedforward neural network structure, with the input layer containing five neurons, each corresponding to one of the five key temporal input features: This represents the PM2.5 mass concentration at the current time t (unit: ), The PM10 mass concentration at time t (unit: ), H(t) represents relative humidity (in percentage), and H(t) represents the aerosol standard scale height (in meters). Represents ambient air temperature (unit: degrees Celsius). The hidden layer consists of N neuron nodes ( to The hidden layer is composed of neurons connected by weighted connections. The model is fully connected to all input layer nodes and uses the sigmoid function as the activation function for nonlinear transformation to extract features. The output layer is a single neuron node AOD, which aggregates the outputs of all hidden layer nodes through a linear function to finally generate the dimensionless predicted value of AOD. This model takes the aforementioned multi-source monitoring data as input and learns the complex nonlinear mapping relationship between it and AOD through training. For example, when a set of measured data is input... At that time, the model can output an estimated AOD value of 0.38, providing core parameters for subsequent accurate calculation of the atmospheric transparency correction coefficient.

[0103] Therefore, according to the above embodiments, the device can obtain a high-performance prediction model through a systematic model training process.

[0104] The sensor selection, data preprocessing method, and neural network structure described in this embodiment are merely illustrative and are not intended to limit the scope of protection of this invention. Those skilled in the art can adjust the parameters and algorithm implementation methods according to actual application scenarios.

[0105] Figure 5 This is a structural block diagram of a multimodal prediction system for photovoltaic panel power generation according to an embodiment of the present invention.

[0106] like Figure 5 As shown, the multimodal prediction system for photovoltaic panel power generation includes: The multi-source heterogeneous data acquisition module 210 is used to acquire irradiance data, atmospheric pollutant concentration data and meteorological parameters on the surface of the photovoltaic panel.

[0107] The power generation environment parameter calculation module 220 is used to calculate the real-time cleanliness index of the photovoltaic panel based on the illumination data, and to calculate the atmospheric transparency correction coefficient based on the atmospheric pollutant concentration data and meteorological parameters.

[0108] The direct irradiance calculation module 230 is used to input historical irradiance data, meteorological parameters, atmospheric pollutant concentration data and real-time cleanliness index into a preset multimodal prediction model. Through the multimodal prediction model, the time series characteristics of historical irradiance data and meteorological parameters are extracted, and the nonlinear relationship between atmospheric pollutant concentration data and irradiance decay is modeled. Based on the extracted time series characteristics and the modeled nonlinear relationship, the direct solar irradiance is output.

[0109] The effective incident irradiance calculation module 240 is used to correct the direct solar irradiance based on the atmospheric transparency correction coefficient and the real-time cleanliness index to obtain the effective incident irradiance of the photovoltaic panel.

[0110] The power generation prediction module 250 is used to calculate the power generation prediction value based on the effective incident irradiance of the photovoltaic panel and the electrical parameters of the photovoltaic panel.

[0111] The specific functions and examples of each module and submodule of the device in this embodiment of the invention can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.

[0112] According to embodiments of the present invention, the above-described method of the present invention can be applied to an electronic device and a readable storage medium.

[0113] Figure 6 A schematic block diagram of an example electronic device 600 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0114] like Figure 6 As shown, the electronic device 600 includes a computing unit 601, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 602 or a computer program loaded from a storage unit 608 into a random access memory (RAM) 603. The RAM 603 may also store various programs and data required for the operation of the electronic device 600. The computing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0115] Multiple components in electronic device 600 are connected to I / O interface 605, including: input unit 606, such as keyboard, mouse, etc.; output unit 607, such as various types of displays, speakers, etc.; storage unit 608, such as disk, optical disk, etc.; and communication unit 609, such as network card, modem, wireless transceiver, etc. Communication unit 609 allows electronic device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0116] The computing unit 601 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above, such as a multimodal prediction method for photovoltaic panel power generation. For example, in some embodiments, a multimodal prediction method for photovoltaic panel power generation can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 600 via ROM 602 and / or communication unit 609. When the computer program is loaded into RAM 603 and executed by the computing unit 601, one or more steps of the multimodal prediction method for photovoltaic panel power generation described above can be performed. Alternatively, in other embodiments, the computing unit 601 may be configured, by any other suitable means (e.g., by means of firmware), to perform a multimodal prediction method for photovoltaic panel power generation.

[0117] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0118] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0119] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0120] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT or LCD monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual, auditory, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0121] The systems and technologies described herein can be implemented in computing systems that include back-end components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include front-end components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0122] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0123] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0124] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this invention should be included within the scope of protection of this invention.

Claims

1. A multi-modal prediction method of photovoltaic panel power generation, characterized in that, The method comprises: acquiring light data, atmospheric pollutant concentration data and meteorological parameters on the surface of a photovoltaic panel; calculating a real-time cleanliness index of the photovoltaic panel according to the light data, and calculating an atmospheric transparency correction coefficient according to the atmospheric pollutant concentration data and the meteorological parameters; inputting historical irradiance data, the meteorological parameters, the atmospheric pollutant concentration data and the real-time cleanliness index into a preset multi-modal prediction model, extracting time sequence features of the historical irradiance data and the meteorological parameters by the multi-modal prediction model, modeling a nonlinear relationship between the atmospheric pollutant concentration data and irradiance attenuation, and outputting solar direct irradiance based on the extracted time sequence features and the modeled nonlinear relationship; correcting the solar direct irradiance according to the atmospheric transparency correction coefficient and the real-time cleanliness index to obtain photovoltaic panel effective incident irradiance; calculating a power generation prediction value according to the photovoltaic panel effective incident irradiance and electrical parameters of the photovoltaic panel.

2. The method of claim 1, wherein, The photovoltaic panel is provided with a pollutant detection device connected to a power supply output end of the photovoltaic panel, and the pollutant detection device is configured with an external light sensor, an internal light sensor, a particulate matter concentration sensor and a meteorological sensor. The light data, atmospheric pollutant concentration data and meteorological parameters on the surface of the photovoltaic panel are acquired by: collecting surface light intensity on the outside of the photovoltaic panel by the external light sensor, and collecting internal light intensity on the inside of the photovoltaic panel by the internal light sensor, and taking the surface light intensity and the internal light intensity as the light data; collecting particulate matter concentration information in the atmospheric environment by the particulate matter concentration sensor to obtain the atmospheric pollutant concentration data; collecting atmospheric environmental parameters including wind speed, wind direction, temperature and humidity by the meteorological sensor as the meteorological parameters.

3. The method of claim 2, wherein, The real-time cleanliness index of the photovoltaic panel is calculated according to the light data, and the atmospheric transparency correction coefficient is calculated according to the atmospheric pollutant concentration data and the meteorological parameters, which comprises: acquiring initial light intensity collected by the external light sensor and the internal light sensor when the photovoltaic panel is in a dust-free state; calculating a reference light transmission ratio of the photovoltaic panel according to the initial light intensity; acquiring surface light intensity and internal light intensity collected by the external light sensor and the internal light sensor in real time, and calculating a real-time light transmission ratio according to the surface light intensity and the internal light intensity collected in real time; calculating the real-time cleanliness index according to the ratio of the real-time light transmission ratio to the reference light transmission ratio; inputting the atmospheric pollutant concentration data and the meteorological parameters as input parameters into a preset aerosol optical depth estimation model, processing the input parameters by the aerosol optical depth estimation model, and outputting the current atmospheric transparency correction coefficient.

4. The method of claim 3, wherein, The multi-modal prediction model is configured with a first prediction branch and a second prediction branch; the multi-modal prediction model is used to extract time sequence features of the historical irradiance data and the meteorological parameters, model a nonlinear relationship between the atmospheric pollutant concentration data and irradiance attenuation, and output solar direct irradiance based on the extracted time sequence features and the modeled nonlinear relationship, including: The first prediction branch is used to extract time sequence features of the historical irradiance data and the meteorological parameters, and obtain a first prediction result; The second prediction branch is used to model a nonlinear relationship between the atmospheric pollutant concentration data and irradiance attenuation, and obtain a second prediction result; The first prediction result and the second prediction result are dynamically weighted and fused to obtain a fused prediction value; The fused prediction value is output as a predicted solar direct irradiance.

5. The method of claim 4, wherein, The atmospheric transparency correction coefficient and the real-time cleanliness index are used to correct the solar direct irradiance to obtain a photovoltaic panel effective incident irradiance, including: The atmospheric transparency correction coefficient is used to correct the predicted solar direct irradiance to obtain an irradiance corrected by atmospheric attenuation; The irradiance corrected by atmospheric attenuation is multiplied by the real-time cleanliness index to obtain the photovoltaic panel effective incident irradiance.

6. The method of claim 5, wherein, The electrical parameters of the photovoltaic panel include photoelectric conversion efficiency, light-receiving area, and temperature influence coefficient; the photovoltaic panel effective incident irradiance and the electrical parameters of the photovoltaic panel are used to calculate a power generation power prediction value, including: The photovoltaic panel effective incident irradiance, the photoelectric conversion efficiency, and the light-receiving area are used to calculate an initial power generation power; Current environmental temperature data are obtained, and the initial power generation power is temperature-compensated and corrected in combination with the temperature influence coefficient to obtain a compensated power generation power; The compensated power generation power is output as the power generation power prediction value.

7. The method of claim 1, wherein, Before the historical irradiance data, the meteorological parameters, the atmospheric pollutant concentration data, and the real-time cleanliness index are input into a preset multi-modal prediction model, the method further includes: Sample irradiance data, sample meteorological parameters, sample atmospheric pollutant concentration data, and sample cleanliness indexes in a preset historical time period are obtained to form a training data set; The training data set is divided into a training set, a validation set, and a test set; An initial multi-modal prediction model including a time sequence feature extraction module and a nonlinear relationship modeling module is constructed; The training set is used to iteratively train the initial multi-modal prediction model, and the model performance in the training process is monitored through the validation set; In response to the model evaluation index on the validation set reaching a preset convergence condition, the training process of the initial multi-modal prediction model is terminated, and a trained multi-modal prediction model is obtained.

8. A multi-modal prediction system for photovoltaic panel power generation, characterized in that, The system includes: A multi-source heterogeneous data acquisition module is configured to acquire light irradiation data, atmospheric pollutant concentration data, and meteorological parameters on a photovoltaic panel surface; The power generation environment parameter calculation module is configured to calculate a real-time cleanliness index of the photovoltaic panel according to the illumination data, and calculate an atmospheric transparency correction coefficient according to the atmospheric pollutant concentration data and the meteorological parameters; The direct irradiance calculation module is configured to input historical irradiance data, the meteorological parameters, the atmospheric pollutant concentration data and the real-time cleanliness index into a preset multi-modal prediction model, extract time sequence features of the historical irradiance data and the meteorological parameters through the multi-modal prediction model, model a nonlinear relationship between the atmospheric pollutant concentration data and irradiance attenuation, and output solar direct irradiance based on the extracted time sequence features and the modeled nonlinear relationship; The effective incident irradiance calculation module is configured to correct the solar direct irradiance according to the atmospheric transparency correction coefficient and the real-time cleanliness index, to obtain photovoltaic panel effective incident irradiance. The power generation power prediction value generation module is configured to calculate a power generation power prediction value according to the photovoltaic panel effective incident irradiance and electrical parameters of the photovoltaic panel.

9. An electronic device, comprising: Comprise: At least one processor; And The memory is in communication connection with the at least one processor; Wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method of any one of claims 1-7.

10. A non-transitory computer-readable storage medium having stored thereon computer instructions, wherein, Computer instructions for causing a computer to execute the method according to any one of claims 1-7.

Images