Photovoltaic power prediction method and system based on clear sky constraint and cloud state perception

Abstract

The application discloses a kind of photovoltaic power prediction method and system based on clear sky constraint and cloud state perception, belong to new energy power generation prediction technical field.The method obtains photovoltaic power station historical data, meteorological data and sky image;Based on clear sky irradiation model, calculate clear sky theoretical power;According to actual output power and clear sky theoretical power, construct normalized attenuation state;Extract sky image cloud disturbance feature and clear sky index CSI;According to clear sky theoretical power and solar incidence condition, carry out physical effective domain discrimination, determine effective and invalid sample;Based on effective sample, train multi-time step prediction model, input cloud disturbance feature, CSI, meteorological prediction feature and historical normalized attenuation state, obtain future multi-time step normalized attenuation state prediction sequence;Combined with future time clear sky theoretical power, reconstruct and obtain photovoltaic power prediction result.The application converts prediction to normalized attenuation state space, improves physical consistency and the perception ability to short time scale cloud change.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation prediction technology, and in particular, a photovoltaic power prediction method and system based on clear sky constraints and cloud state perception. Background Technology

[0002] Existing photovoltaic (PV) power prediction methods typically include physical modeling, statistical modeling, and machine learning methods. Physical modeling methods are based on calculations using solar radiation, module temperature, and PV module characteristics, offering strong interpretability but heavily relying on the accuracy of meteorological parameters. Statistical and machine learning models can learn nonlinear mapping relationships from historical data, but in the absence of physical constraints, these models need to simultaneously learn the impact of multiple factors, such as solar cycle variations, meteorological changes, and the PV system's operating status, on output power from historical samples.

[0003] When photovoltaic (PV) power is directly used as the prediction target, the model needs to simultaneously learn seasonal variations, solar geometric variations, power plant installed capacity constraints, and meteorological attenuation effects. Since PV power is simultaneously affected by deterministic solar motion and random meteorological disturbances, directly establishing a power prediction model would cause the model to simultaneously bear the learning tasks of theoretical irradiance changes and weather attenuation changes, making it difficult for the model to effectively distinguish between predictable variation factors and random disturbance factors. Existing technologies also utilize clear-sky irradiance, clear-sky index, or cloud cover information to assist in PV power prediction; however, these methods typically use this information as input features of the prediction model, or apply simple boundary constraints after the prediction results are generated. They do not fully utilize clear-sky theoretical power to normalize the prediction target, nor do they fully utilize real-time or near-real-time cloud state change information to enhance the model's ability to represent short-term cloud disturbances.

[0004] Especially under conditions of rapid cloud movement, abrupt changes in cloud cover, and localized cloud obstruction, relying solely on historical power data and conventional numerical weather prediction data is insufficient. Furthermore, conventional numerical weather prediction data has limited spatial and temporal resolution, restricting its ability to describe short-term temporal changes such as rapid cloud movement, cloud obstruction, and abrupt changes in cloud cover. Sky images can directly characterize the spatial distribution of clouds above photovoltaic power plants, and continuous sky images can also provide information such as cloud movement direction, speed, and changes in clouds near the sun. Therefore, how to combine the dynamic cloud information represented by sky images with meteorological forecast information and clear-sky physical information to improve the prediction model's ability to perceive short-term cloud disturbances is also a problem that needs to be solved in photovoltaic power prediction. Summary of the Invention

[0005] This invention addresses the problems of existing photovoltaic power prediction methods that directly establish power input-output mappings, causing the model to simultaneously undertake the learning tasks of theoretical irradiance changes and random weather disturbances, as well as the difficulty of conventional prediction models in reflecting rapid changes in cloud conditions in a timely manner. It proposes a photovoltaic power prediction method and system based on clear sky constraints and cloud condition perception.

[0006] To achieve the above technical objectives, this invention discloses a photovoltaic power prediction method based on clear-sky constraints and cloud state perception, comprising the following steps: S1: Acquire historical actual output power, historical meteorological data, predicted meteorological data, time information, site parameters, and sky image data of photovoltaic power plants; perform time synchronization, anomaly processing, and missing value processing on data with time attributes to construct input features. S2, calculate the solar geometric characteristics based on the site parameters and corresponding time information, and calculate the clear sky horizontal irradiance at the corresponding time based on the clear sky irradiance model. Combine the installed capacity and component parameters in the site parameters to calculate the clear sky theoretical power, which serves as the physical benchmark for prediction. S3. Based on the ratio between the historical actual output power and the theoretical power of clear sky at the corresponding time, a normalized attenuation state is constructed; at the same time, features are extracted from continuous historical sky images to obtain cloud disturbance features, and the clear sky index (CSI) is obtained based on the ratio between the actual irradiance and the horizontal irradiance of clear sky at the corresponding time. S4. Based on the theoretical power of clear sky and the cosine value of solar zenith angle, establish the physical effective domain discrimination rule for the normalized decay state. Determine the samples that simultaneously satisfy the theoretical power of clear sky being greater than or equal to a preset power threshold and the cosine value of solar zenith angle being greater than or equal to a preset solar incidence threshold as valid samples, and use the normalized decay state corresponding to the valid samples as the valid training label for the subsequent multi-time step prediction model; exclude the samples that do not meet the conditions as invalid samples. S5 trains a multi-time-step prediction model based on effective training labels. It takes cloud disturbance features, clear sky index (CSI) and its historical variation features, predicted meteorological features, solar geometric features, clear sky theoretical power and historical normalized decay state as inputs, and obtains the normalized decay state prediction sequence for multiple future time steps through the trained multi-time-step prediction model. S6. Based on the normalized decay state prediction sequence for multiple future time steps and the corresponding clear-sky theoretical power at future moments, the photovoltaic power prediction results for multiple future time steps are reconstructed.

[0007] Furthermore, the site parameters in S1 include at least geographical location parameters, installed capacity, and photovoltaic module parameters; the solar geometric features in S2 include one or more of the following: solar altitude angle, solar zenith angle, solar azimuth angle, and solar zenith angle cosine value.

[0008] Furthermore, the clear-sky irradiance model in S2 adopts the Ineichen clear-sky model, which estimates the global horizontal solar irradiance under clear-sky conditions based on the sun's position, geographical location, and atmospheric condition parameters at the target time; the formula for calculating the theoretical power of clear-sky conditions is: , in, The rated installed capacity of the power station for Horizontal irradiance under clear skies at all times The reference irradiance under the standard test conditions (STC) for photovoltaic modules is typically taken as... , for Temperature correction factor at any time.

[0009] Furthermore, in S3, the normalized decay state is constructed by the ratio of the historical actual output power to the theoretical power under clear skies at the corresponding moment, and the calculation formula is expressed as: , For normalized decaying state variables, This represents the actual photovoltaic output power. Theoretical power for clear skies, To prevent extremely small positive numbers with a denominator of zero.

[0010] Furthermore, in S4, the physical effective domain discrimination rule is expressed as follows: , in, for The result of the physical effective domain determination at time t. This indicates that the corresponding time point is a valid sample. This indicates that the corresponding time point is an invalid sample; for Theoretical power of clear sky at any given moment. To preset the theoretical power threshold for clear skies, for The zenith angle of the sun at that moment. To preset the solar incidence threshold; for For samples that do not meet the requirements, their corresponding normalized decay states are not used as valid labels for model training, and such invalid samples are excluded from model training.

[0011] Furthermore, in S3, cloud disturbance features include image spatial features characterizing cloud coverage and spatial distribution, as well as cloud motion features obtained based on optical flow information between consecutive sky images; the Clear Sky Index (CSI) calculation formula is: , in, for The clear sky index at any given time. for The actual total irradiance of the horizontal surface at any given time. for The horizontal irradiance of a clear sky at any given moment.

[0012] Furthermore, in S5, the normalized decay state, cloud disturbance characteristics, clear sky index (CSI) and its historical variation characteristics at the current and historical times, as well as the predicted meteorological characteristics, solar geometric characteristics and clear sky theoretical power corresponding to each future target time are constructed as model inputs, and the normalized decay state of the next N consecutive time steps is constructed as model outputs, establishing a mapping relationship between the input features and the normalized decay state of multiple future time steps.

[0013] Furthermore, in S6, the predicted photovoltaic power value for the k-th time step is: , in, For the future Predicted photovoltaic power at each time step This represents the normalized decay state predicted at the corresponding time step. This corresponds to the theoretical power of clear skies at a future time.

[0014] A system used in a photovoltaic power prediction method based on clear-sky constraints and cloud state perception includes: The data processing module acquires and processes historical actual output power, historical meteorological data, forecast meteorological data, time information, site parameters, and sky image data of photovoltaic power plants, providing a data foundation for the system. The clear-sky theoretical power calculation module obtains solar geometric characteristics, clear-sky horizontal irradiance, and clear-sky theoretical power based on station parameters and time information, serving as the physical benchmark for system prediction; The normalized decay state construction module constructs normalized decay states based on historical actual output power and corresponding clear sky theoretical power, and determines valid and invalid samples according to the physical effective domain discrimination rules. The normalized decay states corresponding to the valid samples are used as valid training labels, and invalid samples are excluded from model training. The cloud state feature construction module obtains cloud disturbance features based on continuous sky image data, and obtains the Clear Sky Index (CSI) and its historical variation features based on the actual total horizontal irradiance and clear sky horizontal irradiance. The multi-time-step prediction module trains a multi-time-step prediction model based on effective training labels, and obtains a normalized decay state prediction sequence for multiple future time steps through the trained multi-time-step prediction model based on cloud state characteristics, predicted meteorological characteristics, solar geometric characteristics, clear sky index (CSI), clear sky theoretical power, and historical normalized decay state. The power reconfiguration module obtains photovoltaic power prediction results for multiple future time steps based on the normalized decay state prediction sequence and the theoretical power under clear skies at the corresponding future time.

[0015] Furthermore, the multi-time-step prediction module employs a gradient boosting tree model, a random forest model, or a neural network model to establish a nonlinear mapping relationship between input features and the normalized decay states of multiple future time steps. Beneficial Effects. This invention uses the theoretical power under clear skies as the physical benchmark for the theoretical power generation capacity of photovoltaic power plants, and constructs a normalized decay state based on the actual photovoltaic power and the theoretical power under clear skies, transforming photovoltaic power prediction from an absolute power space to a physically normalized state space. This invention uses the theoretical power under clear skies to account for theoretical power changes caused by deterministic factors such as solar position and seasonal variations, enabling the multi-time-step prediction model to primarily learn the normalized decay state changes caused by meteorological, cloud, and system state changes, thus improving the physical interpretability of the prediction process. This invention introduces an effective domain discrimination based on the theoretical power under clear skies and solar incidence conditions during the normalized decay state training label construction stage, reducing normalized state anomalies caused by excessively low theoretical power under low light conditions. This invention integrates cloud disturbance features extracted from sky images, the Clear Skies Index (CSI) and its historical variation features, predicted meteorological features, and solar geometric features, enabling the prediction model to simultaneously utilize future meteorological change information and current cloud spatial and motion state information, enhancing its ability to represent short-term cloud changes. This invention employs a multi-time-step prediction structure to obtain the normalized decay state for multiple future time steps, and combines this with the theoretical power under clear skies at the corresponding future times for power reconstruction, thereby obtaining future multi-time-step photovoltaic power prediction results with a clear physical benchmark. The results can more timely reflect short-timescale weather disturbances such as rapid cloud movement and changes in cloud cover. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the photovoltaic power prediction method based on clear sky constraints and cloud state perception according to the present invention.

[0017] Figure 2 This is a schematic diagram of the photovoltaic power multi-step prediction system in an embodiment of the present invention.

[0018] Figure 3 This is a schematic diagram of normalized attenuation state prediction and power reconstruction in an embodiment of the present invention.

[0019] Figure 4This is a schematic diagram of the fusion input of cloud disturbance features and Clear Sky Index (CSI) of sky image in an embodiment of the present invention. Detailed Implementation

[0020] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0021] Example 1: Multi-step photovoltaic power prediction method: like Figure 1 As shown, this embodiment provides a photovoltaic power prediction method based on clear sky constraints and cloud state perception, including steps S1 to S6. For ease of explanation, this embodiment uses a 15-minute resolution, 24-hour prediction of a photovoltaic power station as an example, but the invention is not limited thereto.

[0022] S1. Data Acquisition and Preprocessing: This process acquires historical actual output power, historical meteorological data, predicted meteorological data, time information, site parameters, and sky imagery data from photovoltaic (PV) power plants. Site parameters include geographical location, altitude, installed capacity, and PV module parameters. Geographical location includes longitude and latitude; PV module parameters include module model, rated power, and temperature coefficient. Historical meteorological data includes historical observation information such as actual total horizontal irradiance and ambient temperature; predicted meteorological data includes weather forecast information such as irradiance and ambient temperature for future target time periods. Sky imagery data consists of continuous image sequences captured by all-sky imagers or similar equipment. The historical actual output power, historical meteorological data, and sky imagery data, which have time attributes, are synchronized according to a unified time base, and outlier identification, missing value handling, unit standardization, and time sorting are performed. For example, data is resampled at 15-minute intervals, and missing values ​​are filled using linear interpolation or the average of preceding and following times. Time synchronization ensures that power, meteorological, and imagery data are aligned to the same timestamp, and training samples are constructed based on the predicted time resolution.

[0023] S2. Solar geometry and theoretical power calculation under clear skies: Calculate the solar altitude angle, solar zenith angle, solar azimuth angle, and cosine of the solar zenith angle based on the geographical location of the photovoltaic power station and the target time. Solar zenith angle With solar altitude angle Both conditions are met: , The cosine of the solar zenith angle, as an important physical characteristic characterizing the geometric conditions of solar incidence, is expressed as: , Based on the power station's location, date, and time, the theoretical solar irradiance under cloudless conditions was calculated using a clear-sky irradiance model. The total horizontal irradiance under clear-sky conditions consists of direct irradiance, diffuse irradiance, and surface reflected irradiance, namely: , in, The total irradiance on a clear sky horizontal plane. The horizontal direct irradiance under clear sky conditions. Sky diffuse irradiance, This refers to surface reflected irradiance. Horizontal direct irradiance is determined by the normal direct irradiance under clear sky conditions. Determined together with the solar zenith angle: , When the sun is below the horizon, the clear-sky irradiance is set to zero to avoid non-physical irradiance at night. After obtaining the total clear-sky horizontal irradiance, it is further converted into the corresponding clear-sky theoretical power based on the photovoltaic power station's installed capacity and module rated parameters. The rated power is then calculated under standard module test conditions. and standard test irradiance Based on this, the theoretical power of clear sky can be constructed: , in, The rated installed capacity of the power station The total irradiance on a clear sky horizontal plane. This is a relative correction factor that takes into account factors such as component temperature, component performance, and system losses. When detailed component temperature and system parameters are unavailable, it can be set as follows: =1, and this result is used as the theoretical power upper bound or physical prior value based on solar irradiance conditions. The clear-sky irradiance model can use Ineichen, Haurwitz, or other existing models; then, combined with installed capacity and temperature correction, the clear-sky theoretical power is calculated. The clear-sky theoretical power is used to represent the theoretically achievable power level of a photovoltaic power station under corresponding solar position and clear-sky conditions, and serves as the physical benchmark for subsequent normalized degradation state construction and predicted power reconstruction. Based on the aforementioned clear-sky horizontal irradiance, photovoltaic power station installed capacity, and temperature correction coefficient, the clear-sky theoretical power at the corresponding target time is calculated, expressed as: , in, Horizontal irradiance under clear skies, For the installed capacity of photovoltaic power plants, This is the temperature correction factor.

[0024] S3. Construction of Normalized Decay State and Cloud Perturbation Features: A normalized decay state is constructed based on the relationship between the historical actual output power of a photovoltaic power plant and the theoretical power under clear skies at the corresponding time. This state is used to characterize the deviation of the actual power generation capacity from the theoretically achievable capacity under clear skies. The normalized decay state is expressed as: , in, The normalized decay state at time t, This represents the actual photovoltaic output power. This corresponds to the theoretical power under clear skies at that moment. A positive number is set to prevent the denominator from being zero. To further characterize the impact of cloud cover and motion changes on photovoltaic output over short timescales, feature extraction is performed on continuous sky image data to obtain cloud disturbance features. The cloud disturbance features include image spatial features characterizing the degree of cloud cover and spatial distribution, as well as cloud motion features characterizing the direction, speed, and motion state of clouds in the solar neighborhood.

[0025] In one implementation, spatial features are extracted from continuous sky images, and cloud motion features are obtained based on optical flow information between adjacent sky images. Furthermore, the position of the sun in the sky images is combined to determine a solar neighborhood region, and motion features related to the clouds' proximity to or distance from the sun are extracted from this solar neighborhood region. Additionally, the actual total horizontal irradiance from historical meteorological observation data is used as a basis for further analysis. Horizontal irradiance under clear sky at the corresponding time obtained from S2 Constructing the Clear Sky Index (CSI): , The Clear Sky Index (CSI) is used to characterize the degree of attenuation of actual solar radiation relative to theoretical solar radiation under clear sky conditions, and the current and historical CSI and their variation characteristics are used as input information for subsequent normalized attenuation state prediction.

[0026] S4. Physical effective domain determination and tag processing: Because the theoretical power under clear-sky conditions is relatively small at sunrise, sunset, and low solar altitude angles, directly calculating the ratio between actual photovoltaic power and theoretical power under clear-sky conditions can easily lead to an amplified normalized degradation state. Therefore, effective domain discrimination is performed on the normalized degradation state during the model training phase. A physical effective domain discrimination function is established: , in, for The result of the physical effective domain determination at time t. This indicates that the corresponding time point is a valid sample. This indicates that the corresponding time point is an invalid sample; for Theoretical power of clear sky at any given moment. To preset the theoretical power threshold for clear skies, for The zenith angle of the sun at that moment. This is a preset solar incidence threshold. For For samples, their corresponding normalized decay states are not used as valid labels for model training.

[0027] Will The normalized decay states corresponding to the effective samples are used as effective training labels for subsequent multi-timestep prediction models. For samples that do not meet the criteria, their corresponding normalized decay states are not used as valid labels for model training, and such invalid samples are excluded from model training. For example, Set at 5% of the rated installed capacity. Set to 0.1. This step is a physical constraint in the training sample and prediction label construction stage, used to reduce the normalization state anomaly caused by the small theoretical power of clear sky under low light conditions, rather than a simple post-processing pruning of the prediction results.

[0028] S5. Multi-timestep normalized decay state prediction: The effective training labels obtained based on S4 (i.e. A multi-time-step prediction model is trained using the normalized decay states corresponding to the samples. Model inputs include: the normalized decay states at the current and historical times, cloud disturbance characteristics, Clear Sky Index (CSI) and its historical variation characteristics; predicted meteorological characteristics, solar geometric characteristics, and theoretical clear sky power for each future target time. The model output is a predicted sequence of normalized decay states for N consecutive future time steps. In this embodiment, the prediction time resolution is set to 15 minutes, the prediction range is set to the next 24 hours, and N=96. The multi-time-step prediction model uses a gradient boosting tree model, a random forest model, or a neural network model to establish a nonlinear mapping relationship between the input features and the normalized decay states for multiple future time steps.

[0029] During model training, a training set is constructed using historical valid samples. The model parameters are optimized by minimizing the loss function between the predicted sequence and the actual normalized decay state sequence. The loss function can be the mean squared error. After training, in the prediction phase, current and historical state information and sky image information are acquired. Cloud disturbance features are extracted, and current and historical CSI are calculated. Simultaneously, meteorological prediction data, solar geometric features, and theoretical power of clear skies for each future target time are obtained. These are input into the trained multi-time-step prediction model to obtain the normalized decay state prediction sequence for the next 96 time steps. , in, This represents the normalized decay state prediction value at the k-th time step in the future.

[0030] The multi-time-step prediction model can adopt gradient boosting tree model, random forest model, neural network model or other models that can achieve multi-time-step prediction. This invention does not limit the specific model form.

[0031] S6, Photovoltaic power reconfiguration: Based on the normalized decay state prediction values ​​obtained from S5 for multiple future time steps and the theoretical power under clear skies at the corresponding future time moments obtained from S2, the photovoltaic power prediction results for multiple future time steps are reconstructed.

[0032] For the k-th time step in the future, the predicted value of photovoltaic power is expressed as: , in, For the future Predicted photovoltaic power at each time step This represents the normalized decay state predicted at the corresponding time step. This corresponds to the theoretical power of clear skies at a future time.

[0033] By using the above methods, deterministic changes such as solar position and seasonal variations are mainly represented by the theoretical power under clear skies, while changes in power generation capacity caused by clouds, weather, and system status are mainly represented by the normalized decay state, thus enabling future multi-timestep photovoltaic power reconfiguration.

[0034] Example 2: Photovoltaic Power Multi-Step Prediction System like Figure 2 As shown in this embodiment, a photovoltaic power prediction system based on clear-sky constraints and cloud state perception is used to implement the multi-step photovoltaic power prediction method described in Embodiment 1. The system includes a data input module, a clear-sky physical modeling module, a normalized attenuation state construction module, a cloud state feature construction module, a multi-time-step attenuation state prediction module, a power reconstruction module, and a result output module.

[0035] The data input module is used to acquire historical actual output power, historical meteorological data, predicted meteorological data, sky image data, time information, and photovoltaic power station parameters. Data with time attributes are synchronized according to a unified time base, and corresponding model input data is constructed based on a preset prediction time resolution.

[0036] The clear-sky physics modeling module calculates solar geometric characteristics based on the geographical location parameters and corresponding time information of the photovoltaic power station, and obtains the clear-sky horizontal irradiance at the corresponding time using the clear-sky irradiance model. Furthermore, it combines the installed capacity of the photovoltaic power station and the parameters of the photovoltaic modules to obtain the theoretical power under clear-sky conditions. The theoretical power under clear-sky conditions represents the theoretical power output of the photovoltaic power station under the corresponding solar position and clear-sky conditions, and serves as the physical benchmark for constructing the normalized degradation state and reconfiguring future photovoltaic power.

[0037] The normalized decay state construction module is used to construct normalized decay states based on historical actual output power and corresponding theoretical power under clear skies, to characterize the deviation of the actual power generation capacity of the photovoltaic power plant from the theoretical power generation capacity under clear skies. Simultaneously, valid and invalid samples are determined according to the physical effective domain discrimination rule. The normalized decay states corresponding to valid samples are used as valid training labels, while invalid samples are excluded from model training.

[0038] The cloud state feature construction module is used to construct model input features related to cloud disturbance based on sky image data and historical meteorological observation data. Spatial features are extracted from continuous sky image data to obtain image spatial features characterizing cloud coverage, spatial distribution, and cloud morphology. Cloud motion features are obtained based on changes between adjacent sky images to characterize cloud direction, speed, and the movement of clouds near or away from the sun. Simultaneously, the Clear Sky Index (CSI) is calculated based on the actual total horizontal irradiance from historical meteorological observation data and the corresponding clear sky horizontal irradiance obtained from the clear sky physical modeling module, and further, historical CSI variation features are obtained based on continuous CSI data. Cloud disturbance features, CSI, and their historical variation features together serve as input information for subsequent multi-time-step normalized decay state prediction.

[0039] The multi-time-step decay state prediction module constructs joint input features from historical normalized decay state, cloud disturbance characteristics, Clear Sky Index (CSI) and its historical variation characteristics, predicted meteorological characteristics, solar geometric characteristics, and theoretical clear sky power. Based on effective training labels, it trains a multi-time-step prediction model and uses this model to obtain a normalized decay state prediction sequence for multiple future time steps. In one specific implementation, the prediction time resolution is set to 15 minutes, and the prediction range is set to the next 24 hours. The multi-time-step prediction model then outputs the normalized decay state prediction values ​​for the next 96 time steps at a time.

[0040] The power reconfiguration module combines (multiplies) the normalized degradation state prediction values ​​for each future time step output by the multi-time-step degradation state prediction module with the corresponding future clear-sky theoretical power to obtain the photovoltaic power prediction results for multiple future time steps. For the k-th future time step, the photovoltaic power prediction value is obtained according to the following formula: , in, Indicates the future number Predicted photovoltaic power at each time step This represents the normalized decay state prediction value for the corresponding time step output by the multi-time-step prediction model. This represents the theoretical power of clear skies at a future time.

[0041] The output module is used to output a photovoltaic power prediction sequence for multiple future time steps according to a preset time order, and to provide the prediction results to photovoltaic power plant operation monitoring, energy storage scheduling, photovoltaic-storage-charging microgrid energy management, or new energy grid-connected control systems. In this embodiment, the model training data uses a dataset with a unified time resolution. For a 15-minute prediction resolution, the actual photovoltaic power can be statistically analyzed according to a unified 15-minute time window, ensuring that the actual power, meteorological data, theoretical power under clear skies, and cloud status information use a consistent time base. If the actual power uses the average value of the 15-minute time window as the power label, then the theoretical power under clear skies and the corresponding meteorological and cloud status information are also matched using the same time caliber; in another implementation, the instantaneous power at the sampling moment can also be used as the label, but the data caliber should be kept consistent during the training, verification, and testing phases.

[0042] The system first reads historical actual output power, historical meteorological data, historical weather forecast data, and parameters such as the power station's geographical location and installed capacity, and then aligns the data from different sources according to time. Based on the latitude and longitude of the photovoltaic power station's location and the corresponding time, it calculates the solar zenith angle, solar altitude angle, and solar azimuth angle, and uses a clear-sky irradiance model to calculate the clear-sky horizontal irradiance at the corresponding time. Furthermore, it combines the photovoltaic power station's installed capacity and temperature correction information to calculate the theoretical clear-sky power for each time step.

[0043] Based on the preset physical effective domain discrimination rules, samples that satisfy both a theoretical power of clear sky greater than or equal to a preset power threshold and a cosine value of the solar zenith angle greater than or equal to a preset solar incidence threshold are determined as valid samples. For valid samples, a normalized decay state is constructed based on the proportional relationship between historical actual output power and the corresponding theoretical power of clear sky, and this normalized decay state is used as a valid training label. For invalid samples that do not meet the conditions (such as time points where the theoretical power of clear sky is too low or the solar altitude angle is too low), they are not used as valid normalized decay state training labels and are excluded from model training.

[0044] like Figure 4 As shown, during the construction of cloud state-related features, the system acquires continuous sky image data corresponding to the photovoltaic power data time. Spatial feature extraction and adjacent-time-time image change analysis are performed on these continuous sky images to obtain cloud disturbance features characterizing cloud coverage, spatial distribution, and cloud movement changes. Simultaneously, the Clear Sky Index (CSI) is calculated based on the actual total horizontal irradiance and clear sky horizontal irradiance at the corresponding time, and historical CSI change features are constructed based on the CSI data from consecutive historical times. These cloud disturbance features, along with the CSI and its historical change features, serve as input information for the subsequent multi-time-step normalized decay state prediction model.

[0045] like Figure 3 As shown, during the model training phase, the normalized decay state at the current and historical times, cloud disturbance features extracted from continuous sky images, clear sky index (CSI) and its historical variation features, meteorological prediction features corresponding to the target time, solar geometric features, clear sky theoretical power features, and time period features are constructed as model inputs. The normalized decay state at the next 96 time steps is constructed as the model output. The multi-time-step normalized decay state prediction model is trained based on effective training labels.

[0046] During the prediction phase, the system acquires current and historical state information and sky image information, extracts cloud disturbance features and calculates the current and historical Clear Sky Index (CSI). At the same time, it obtains meteorological prediction data, solar geometric features and clear sky theoretical power corresponding to each target time in the future. Through the completed multi-time step prediction model, the system obtains the normalized decay state prediction results for the next 96 time steps.

[0047] The 96 normalized degradation state prediction values ​​are combined with the corresponding future clear-sky theoretical power (i.e., multiplication and reconstruction) to obtain the photovoltaic power prediction sequence for the next 24 hours. This forms a rolling prediction process of "data input layer → physical modeling layer → cloud state feature construction layer → multi-step normalized degradation state prediction layer → power reconstruction layer → result output layer".

[0048] The above description is merely one embodiment of the present invention and is not intended to limit the present invention. Any minor modifications, equivalent substitutions, and improvements made to the above embodiment based on the technical essence of the present invention should be included within the protection scope of the present invention.

Claims

1. A photovoltaic power prediction method based on clear-sky constraints and cloud state perception, characterized in that, Includes the following steps: S1: Acquire historical actual output power, historical meteorological data, predicted meteorological data, time information, site parameters, and sky image data of photovoltaic power plants; perform time synchronization, anomaly processing, and missing value processing on data with time attributes to construct input features. S2, calculate the solar geometric characteristics based on the site parameters and corresponding time information, and calculate the clear sky horizontal irradiance at the corresponding time based on the clear sky irradiance model. Combine the installed capacity and component parameters in the site parameters to calculate the clear sky theoretical power, which serves as the physical benchmark for prediction. S3. Based on the ratio between the historical actual output power and the theoretical power of clear sky at the corresponding time, a normalized attenuation state is constructed; at the same time, features are extracted from continuous historical sky images to obtain cloud disturbance features, and the clear sky index (CSI) is obtained based on the ratio between the actual irradiance and the horizontal irradiance of clear sky at the corresponding time. S4. Based on the theoretical power of clear sky and the cosine value of solar zenith angle, establish the physical effective domain discrimination rule for the normalized decay state. Determine the samples that simultaneously satisfy the theoretical power of clear sky being greater than or equal to a preset power threshold and the cosine value of solar zenith angle being greater than or equal to a preset solar incidence threshold as valid samples, and use the normalized decay state corresponding to the valid samples as the valid training label for the subsequent multi-time step prediction model; exclude the samples that do not meet the conditions as invalid samples. S5 trains a multi-time-step prediction model based on effective training labels. It takes cloud disturbance features, clear sky index (CSI) and its historical variation features, predicted meteorological features, solar geometric features, clear sky theoretical power and historical normalized decay state as inputs, and obtains the normalized decay state prediction sequence for multiple future time steps through the trained multi-time-step prediction model. S6. Based on the normalized decay state prediction sequence for multiple future time steps and the corresponding clear-sky theoretical power at future moments, the photovoltaic power prediction results for multiple future time steps are reconstructed.

2. The photovoltaic power prediction method based on clear-sky constraints and cloud state perception according to claim 1, characterized in that, The site parameters in S1 include at least geographical location parameters, installed capacity, and photovoltaic module parameters; the solar geometric features in S2 include one or more of the following: solar altitude angle, solar zenith angle, solar azimuth angle, and solar zenith angle cosine value.

3. The photovoltaic power prediction method based on clear-sky constraints and cloud state perception according to claim 1, characterized in that, The clear-sky irradiance model in S2 uses the Ineichen clear-sky model to estimate the global horizontal solar irradiance under clear-sky conditions based on the sun's position, geographical location, and atmospheric conditions at the target time; the theoretical clear-sky power at time t is... The calculation formula is: ， in, The rated installed capacity of the power station for Horizontal irradiance under clear skies at all times The reference irradiance under the standard test conditions (STC) for photovoltaic modules is typically taken as... , for Temperature correction factor at any time.

4. The photovoltaic power prediction method based on clear-sky constraints and cloud state perception according to claim 1, characterized in that, In S3, the normalized decay state is constructed by the ratio of the historical actual output power to the theoretical power under clear skies at the corresponding moment. The calculation formula is expressed as follows: ， For normalized decaying state variables, This represents the actual photovoltaic output power. Theoretical power for clear skies, To prevent extremely small positive numbers with a denominator of zero.

5. The photovoltaic power prediction method based on clear-sky constraints and cloud state perception according to claim 1, characterized in that, In S4, the physical effective domain is determined based on the theoretical power of clear sky and the cosine value of the solar zenith angle. The physical effective domain determination rule is expressed as follows: ， in, for The result of the physical effective domain determination at time t. This indicates that the corresponding time point is a valid sample. This indicates that the corresponding time point is an invalid sample; for Theoretical power of clear sky at any given moment. To preset the theoretical power threshold for clear skies, for The zenith angle of the sun at that moment. To preset the solar incidence threshold; for For samples, their corresponding normalized decay states are not used as valid labels for model training.

6. The photovoltaic power prediction method based on clear-sky constraints and cloud state perception according to claim 1, characterized in that, In S3, cloud disturbance features include image spatial features characterizing cloud cover and spatial distribution, as well as cloud motion features obtained based on optical flow information between consecutive sky images; the Clear Sky Index (CSI) calculation formula is: ， in, for The clear sky index at any given time. for The actual total irradiance of the horizontal surface at any given time. for The horizontal irradiance of a clear sky at any given moment.

7. The photovoltaic power prediction method based on clear-sky constraints and cloud state perception according to claim 1, characterized in that, In S5, the normalized decay state, cloud disturbance characteristics, clear sky index (CSI) and its historical variation characteristics at the current and historical times, as well as the predicted meteorological characteristics, solar geometric characteristics and clear sky theoretical power corresponding to each future target time are constructed as model inputs. The normalized decay state of the next N consecutive time steps is constructed as the model output, establishing a mapping relationship between the input features and the normalized decay state of multiple future time steps.

8. The photovoltaic power prediction method based on clear-sky constraints and cloud state perception according to claim 7, characterized in that, In S6, the predicted photovoltaic power at the k-th time step is: ， in, For the future Predicted photovoltaic power at each time step This represents the normalized decay state predicted at the corresponding time step. This corresponds to the theoretical power of clear skies at a future time.

9. A system used in the photovoltaic power prediction method based on clear-sky constraints and cloud state perception as described in any one of claims 1-8, characterized in that, include: The data processing module acquires and processes historical actual output power, historical meteorological data, forecast meteorological data, time information, site parameters, and sky image data of photovoltaic power plants, providing a data foundation for the system. The clear-sky theoretical power calculation module obtains solar geometric characteristics, clear-sky horizontal irradiance, and clear-sky theoretical power based on station parameters and time information, serving as the physical benchmark for system prediction; The normalized decay state construction module constructs normalized decay states based on historical actual output power and corresponding clear sky theoretical power, and determines valid and invalid samples according to the physical effective domain discrimination rules. The normalized decay states corresponding to the valid samples are used as valid training labels, and invalid samples are excluded from model training. The cloud state feature construction module obtains cloud disturbance features based on continuous sky image data, and obtains the Clear Sky Index (CSI) and its historical variation features based on the actual total horizontal irradiance and clear sky horizontal irradiance. The multi-time-step prediction module trains a multi-time-step prediction model based on effective training labels, and obtains a normalized decay state prediction sequence for multiple future time steps through the trained multi-time-step prediction model based on cloud state characteristics, predicted meteorological characteristics, solar geometric characteristics, clear sky index (CSI), clear sky theoretical power, and historical normalized decay state. The power reconfiguration module obtains photovoltaic power prediction results for multiple future time steps based on the normalized decay state prediction sequence and the theoretical power under clear skies at the corresponding future time.

10. The prediction system according to claim 9, characterized in that, The multi-time-step prediction module uses a gradient boosting tree model, a random forest model, or a neural network model to establish a nonlinear mapping relationship between input features and the normalized decay state of multiple future time steps.

Images