Photovoltaic power prediction method and system based on generative large model

Abstract

The application provides a photovoltaic power prediction method and system based on a generative large model, and relates to the technical field of power prediction.The method comprises the following steps: performing multi-task field pre-training on a general generative large model based on massive historical operation data of photovoltaic power stations to obtain a photovoltaic generative large model; and receiving static parameters uploaded by a target power station, including power station geographic coordinates, installed capacity, component type, support inclination angle and azimuth angle parameters.The photovoltaic generative large model is obtained through multi-task pre-training, zero-sample prediction is realized with the aid of an enhanced physical prior prompt word, and the prediction result is verified and optimized through triple physical post-processing, so that the prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of power prediction technology, and in particular to a photovoltaic power prediction method and system based on a generative large model. Background Technology

[0002] To alleviate the cold start problem of newly built power plants, the industry often adopts a transfer learning strategy. This involves pre-training a basic model on data from multiple source power plants and then fine-tuning it using a small amount of historical data from the target power plant. However, this approach may have some shortcomings in practical application. On the one hand, different power plants have significant differences in geographical environment, component types, and installation conditions, and fine-tuning each power plant individually would consume additional computational and human resources. On the other hand, when the available historical data for the target power plant is very limited (e.g., less than two weeks), the model may fit local noise or abnormal events to some extent during the fine-tuning process, resulting in insignificant improvement in prediction accuracy under normal operating conditions, or even a tendency for a decline in generalization ability.

[0003] For example, a newly built 1.5MW rooftop distributed photovoltaic power station in an industrial park uses monocrystalline silicon modules and has a fixed 15-degree tilt angle, facing due south. After grid connection, the local weather was predominantly cloudy for the first 10 days. One morning, a rapid cloud event lasting approximately 40 minutes occurred, causing the irradiance to plummet from 620W / m² to 180W / m² within 15 minutes, resulting in drastic fluctuations in measured power. If a traditional transfer learning approach is used, fine-tuning the pre-trained model using only these 10 days of historical data could lead to overlearning of the instantaneous changes in this rapid cloud event, misinterpreting it as a common spring power output pattern in the region. When subsequent periods of sunny weather occur, the fine-tuned model might incorrectly predict multiple unnecessary power drops during the midday period, causing significant deviations between the predicted and actual power outputs. This could negatively impact the power station's participation in the provincial electricity spot market during the midday trading session. Meanwhile, the power station has not yet been equipped with local meteorological sensors, and the small number of outliers mixed in the data further increases the risk of overfitting during the fine-tuning process. Summary of the Invention

[0004] This invention provides a photovoltaic power prediction method and system based on a generative large model, which avoids misjudging short-term occasional meteorological disturbances as normal power output patterns and ensures that the prediction sequence closely matches the real physical characteristics of photovoltaics.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a photovoltaic power prediction method based on a generative large model, the method comprising: Step 1: Pre-train the general generative large model on a large amount of historical operation data of photovoltaic power plants for multiple tasks to obtain the photovoltaic generative large model. Step 2: Receive static parameters uploaded by the target power station, including the power station's geographical coordinates, installed capacity, component type, and support tilt and azimuth parameters; Step 3: Based on static parameters and gridded data of future time-period numerical weather forecasts obtained from the meteorological service interface, extract prior information on cloud field disturbance quantification around the target power station and construct enhanced physical prior cue words; Step 4: After splicing the enhanced physical prior cue words with the meteorological forecast data of the target power plant in the future period according to the preset time resolution, input the data into the photovoltaic generative large model to perform zero-sample power prediction inference and obtain the preliminary predicted power sequence of the target power plant in the future period. Step 5: Perform triple physical post-processing verification on the preliminary predicted power sequence, including nighttime zero-value forced constraint correction based on solar altitude angle calculation, output upper and lower limit truncation processing based on installed capacity threshold, and moving average smoothing processing on the section where the power change slope exceeds the preset abrupt change threshold within the continuous time window, to obtain the final predicted power sequence.

[0006] Secondly, a photovoltaic power prediction system based on a generative large model includes: The model training module is used to pre-train a general generative large model on a massive amount of historical operation data of photovoltaic power plants in multiple tasks and domains to obtain a photovoltaic generative large model. The data receiving module is used to receive static parameters uploaded by the target power station, including the power station's geographical coordinates, installed capacity, component type, and support tilt and azimuth parameters. The prompt word construction module is used to extract quantitative prior information on cloud field disturbances around the target power station based on static parameters and gridded data of future time-period numerical weather forecasts obtained from the meteorological service interface, and to construct enhanced physical prior prompt words. The model prediction module is used to stitch together the enhanced physical prior cue words with the meteorological forecast data of the target power plant in the future period according to the preset time resolution, and then input the data into the photovoltaic generative large model to perform zero-sample power prediction inference to obtain the preliminary predicted power sequence of the target power plant in the future period. The processing and verification module is used to perform triple physical post-processing verification on the preliminary predicted power sequence, including nighttime zero-value forced constraint correction based on solar altitude angle calculation, output upper and lower limit truncation processing based on installed capacity threshold, and moving average smoothing processing for sections where the power change slope exceeds the preset abrupt change threshold within a continuous time window, to obtain the final predicted power sequence.

[0007] Thirdly, a computing device, comprising: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0008] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0009] The above-described solution of the present invention has at least the following beneficial effects: This invention pre-trains a general generative large-scale model for photovoltaic (PV) multi-task domain applications, enabling the model to learn meteorological correlations with PV power output, solar position mechanisms, and cloud field disturbance characteristics. This endows the model with PV industry-specific knowledge, effectively solving the problem that traditional physical models struggle to adapt to complex operating conditions such as module dust accumulation, aging, and partial shading. It overcomes the limitations of traditional models in terms of accuracy, fundamentally improving the adaptability and accuracy of power prediction under various operating conditions. Enhanced physical priors are constructed by integrating static parameters of the power plant with meteorological grid data, quantifying the spatial distribution and dynamic disturbance characteristics of the cloud field. Compared to traditional single-time-series data modeling methods, the addition of spatial meteorological information effectively perceives cloud field movement and spatial differences in irradiance, guiding the model to generate prediction results based on physical mechanisms. This overcomes the limitations of simple data fitting and improves the reliability of prediction results under complex meteorological conditions. This approach achieves zero-sample prediction without local data from target power plants, addressing the cold-start problem for newly built, sensorless distributed photovoltaic power plants. It abandons the traditional station-by-station fine-tuning model of transfer learning, avoiding overfitting and poor generalization caused by small-sample fine-tuning, while also saving significant computing power and manual maintenance costs. A triple-progressive physical post-processing verification mechanism is employed to physically correct and constrain the model's prediction results. Through nighttime zero-value constraints, installed capacity boundary truncation, and power surge smoothing, it successively eliminates issues such as false nighttime output, abnormal power exceeding thresholds, and non-physical power surge noise. This prevents short-term, occasional meteorological disturbances from being misjudged as normal output patterns, ensuring that the predicted sequence closely matches the true physical characteristics of photovoltaic power plants, effectively supporting power plant scheduling and power trading decisions. Attached Figure Description

[0010] Figure 1 This is a schematic flowchart of a photovoltaic power prediction method based on a generative large model provided in an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of a photovoltaic power prediction system based on a generative large model provided in an embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] like Figure 1 As shown, embodiments of the present invention propose a photovoltaic power prediction method based on a generative large model, the method comprising the following steps: Step 1: Pre-train the general generative large model on a large amount of historical operation data of photovoltaic power plants for multiple tasks to obtain the photovoltaic generative large model. Step 2: Receive static parameters uploaded by the target power station, including the power station's geographical coordinates, installed capacity, component type, and support tilt and azimuth parameters; Step 3: Based on static parameters and gridded data of future time-period numerical weather forecasts obtained from the meteorological service interface, extract prior information on cloud field disturbance quantification around the target power station and construct enhanced physical prior cue words; Step 4: After splicing the enhanced physical prior cue words with the meteorological forecast data of the target power plant in the future period according to the preset time resolution, input the data into the photovoltaic generative large model to perform zero-sample power prediction inference and obtain the preliminary predicted power sequence of the target power plant in the future period. Step 5: Perform triple physical post-processing verification on the preliminary predicted power sequence, including nighttime zero-value forced constraint correction based on solar altitude angle calculation, output upper and lower limit truncation processing based on installed capacity threshold, and moving average smoothing processing on the section where the power change slope exceeds the preset abrupt change threshold within the continuous time window, to obtain the final predicted power sequence.

[0014] In this embodiment of the invention, multi-task domain pre-training enables the large model to master the general mapping law between photovoltaic power output and meteorological elements. During the inference stage, no historical data from the target power plant is required; power prediction can be completed solely through static parameters and numerical weather prediction, solving the cold start problem caused by data scarcity in newly built power plants. A zero-sample inference paradigm is adopted, keeping model parameters constant. Prediction behavior is entirely guided by physical prior cues, avoiding overlearning of anomalous events in limited local data and ensuring prediction stability under different weather types. By deploying light field perturbation sentinel probes in the upwind and crosswind directions, the spatial variation coefficient and spatial gradient sequence of irradiance are extracted from the gridded data of numerical weather prediction. The impact of cloud field movement on irradiance is quantified as physical prior information and embedded into the model input, enhancing the ability to characterize rapid power fluctuations under cloudy weather conditions. When cloud cover exceeds a threshold, a closed sampling domain is automatically constructed; when the prevailing wind direction deflects, the probe position is automatically adjusted and switched to a diagonal sampling path, ensuring that the physical prior information always matches real-time meteorological conditions. By sequentially constraining the solar altitude angle to zero at night, truncating the installed capacity threshold, and smoothing out excessive power change slope, potential physically unreasonable outputs in model inference are eliminated. All meteorological inputs are taken from numerical weather prediction grid data, eliminating the need to deploy local meteorological sensors at the target power plant, thus reducing hardware investment and operational burden. Using the area of ​​a spherical triangle as the spatial weighting basis, the physical differences in cloud field dwell time at different locations due to Earth's curvature are considered, improving the accuracy of the representation of prior information on cloud field disturbance quantification.

[0015] In a preferred embodiment of the present invention, step 1 includes: Step 100: Collect historical operation datasets from multiple photovoltaic power plants. The historical operation datasets for each photovoltaic power plant include meteorological element sequences and corresponding actual power generation sequences for each time step within a historical period, specifically including: Historical operational datasets from multiple photovoltaic (PV) power plants were collected, broadly selecting grid-connected PV plants with varying climates, topography, installed capacity, PV module types, and support system installation parameters as data acquisition targets. This ensured that the collected samples covered all actual operating conditions, including stable output on sunny days, low-light output on cloudy days, fluctuating output during cloudy periods, and rapid cloud disturbances. A uniform 15-minute fixed time resolution was used to collect historical operational data, dividing each day into 96 standard time-series steps, fully covering the entire day and night period, conforming to industry time-series accuracy standards for short-term PV power forecasting. During the collection process, only data from normally grid-connected operations was retained, discarding invalid data from periods of equipment maintenance, shutdown for debugging, or grid disconnection due to faults.

[0016] For each standard time step, meteorological element sequences and actual generated power sequences from the same power station and at the same timestamp are collected synchronously. The meteorological element sequence includes five parameters: total horizontal irradiance, ambient temperature, near-surface 10-meter wind speed, near-surface 10-meter wind direction, and total cloud cover. All parameters use 15-minute average data. The actual generated power sequence is the 15-minute average of the total active power generated by the entire power station, collected and aggregated by the power station's inverters, representing the actual power generation output status at each time step.

[0017] Using a unified timestamp as a benchmark, the meteorological element sequence and the actual generated power sequence are time-series aligned and bound one-to-one, and standardized data cleaning is performed. Specifically, cleaning involves removing invalid samples with incorrect timestamps, null values, or transmission anomalies; and removing abnormal samples exceeding physical ranges, including those with irradiance outside the range of 0 to 1360 W / m², ambient temperature outside the range of -40℃ to 60℃, actual generated power less than 0 kW, or exceeding the rated installed capacity of the corresponding power plant. The rated installed capacity of the power plant is uniformly limited to no more than 100 MW per plant, and this value serves as a fixed threshold for power limit screening. After cleaning, all samples under normal operating conditions and effective disturbance conditions are fully retained, and all valid time-series data from all power plants are aggregated to form a multi-PV power plant historical operation dataset.

[0018] Step 101: Label the solar position parameters for each time step in the historical operation dataset, including the solar altitude angle and solar azimuth angle; label the irradiance anomaly fluctuation sections in the historical operation dataset with cloud field disturbance labels, which are used to characterize the degree of influence of cloud field movement on irradiance in the section, thus obtaining the labeled historical operation dataset, specifically including: For each time-series sample in the historical operational dataset, based on the sample collection time, date, and corresponding geographical latitude and longitude of the photovoltaic power station, the solar declination angle, solar hour angle, solar altitude angle, and solar azimuth angle are calculated step-by-step to complete the annotation of solar position parameters. Specifically, the solar declination angle is calculated to characterize the latitude offset of the Earth's direct sunlight point, i.e. .

[0019] In the formula, The solar declination angle, in degrees (°). The number of days within the year is determined based on the year, month, and day of sample collection, and its value ranges from 1 to 365. The constant 284 is an astronomical calibration constant, representing the cumulative number of days shift from January 1st to the autumnal equinox of the current year, used to correct the annual phase deviation of the solar declination angle.

[0020] The solar hour angle is used to characterize the angular deviation of the sun relative to local noon, i.e. .

[0021] In the formula, Solar hour angle, in degrees; It is local true solar time, in hours. The hour angle at noon is 0°, the hour angle in the morning is negative, and the hour angle in the afternoon is positive.

[0022] The solar altitude angle is solved using spherical trigonometry, i.e. .

[0023] In the formula, Solar altitude angle, in degrees (°). The latitude of the station is expressed in degrees (°). The solar altitude angle is calculated using arcsine calculation, with a value ranging from -90° to 90°.

[0024] The solar azimuth angle is used to characterize the angle of deviation of the sun's horizontal projection relative to due south. In the formula, The solar azimuth angle is expressed in degrees. After calculation, the final solar azimuth angle is obtained by combining the positive and negative correction quadrants of the hour angle. The solar altitude angle calculated step-by-step is then bound to the corresponding time-series samples to complete the annotation of solar position parameters for the entire sample.

[0025] Based on this, cloud field disturbance quantification and labeling were performed on the irradiance sequence of the dataset. A fixed sliding window of 1 hour was used to traverse the time series data. Each window contained four consecutive 15-minute time steps. The statistical characteristics of irradiance (mean irradiance, standard deviation irradiance, and irradiance fluctuation variability) were calculated for each window, where the irradiance fluctuation variability was calculated by dividing the standard deviation of irradiance within the window by the mean irradiance within the window. The intensity of cloud field disturbance was determined by the fluctuation index.

[0026] Set a fixed threshold for cloud field disturbance detection. The threshold was determined using the maximum inter-class variance algorithm. This involved collecting time-series photovoltaic irradiance samples from multiple regions and seasons, and calculating the irradiance fluctuation variability for each sample. Samples under stable operating conditions (sunny and cloudy days with no significant fluctuations and no cloud cover) were designated as the stable operating condition sample set. Samples under disturbed operating conditions (short-term rapid increases and decreases in irradiance due to cloud movement and shading) were designated as the disturbed operating condition sample set. The variability distribution characteristics of the two sample sets were statistically analyzed. Different candidate variability thresholds were iterated, and the inter-class variance corresponding to each candidate value was calculated. The variability value corresponding to the maximum inter-class variance was determined as the optimal operating condition distinction threshold. Finally, a cloud field disturbance judgment threshold was assumed. .when When the current irradiance status is determined to be stable, it is marked as an undisturbed level; when At that time, the current period was determined to be a segment of abnormal irradiance fluctuation, indicating the presence of cloud field movement disturbances. The abnormal fluctuation segment was then classified and labeled using a three-level quantitative classification system. This indicates a slight cloud field disturbance. This represents a moderate cloud field disturbance. 0.50 represents severe cloud field disturbance. After labeling the entire dataset using the above quantization method, a labeled historical dataset is obtained, carrying both solar position parameters and multi-level cloud field disturbance labels.

[0027] Step 102: Based on the labeled historical running dataset, construct multi-task joint training samples. These samples include power prediction main task samples, solar position sensing auxiliary task samples, and cloud field disturbance identification auxiliary task samples, specifically including: Based on the annotated historical running dataset obtained in step 101, and taking the time-series sample with complete annotation information in a single strip as the smallest unit, three types of matching task samples are constructed: power prediction main task sample, solar position perception auxiliary task sample, and cloud field disturbance identification auxiliary task sample. Finally, they are merged to form a multi-task joint training sample. The three types of tasks share the same input feature space to ensure the consistency of multi-task learning features.

[0028] When constructing the power prediction master task samples, all meteorological element sequences at each time step are used as model input features, and the actual photovoltaic power generation value corresponding to the same time step is used as the true label for the model regression prediction, forming the input-output paired samples for the power prediction master task. The power prediction master task is a continuous value regression task, and the mean squared error loss function is used to quantify the prediction bias during the training process. This master task is used to learn the core time-series mapping relationship between meteorological parameters and photovoltaic output, and is the core supervision task for model pre-training.

[0029] When constructing the solar position sensing auxiliary task samples, the meteorological element input features are completely consistent with those of the main task. The solar altitude angle and solar azimuth angle values ​​obtained in step 101 through time-step quantization calculations using the solar declination angle, solar hour angle, and spherical trigonometric astronomical formula are retrieved to form the solar position sensing auxiliary task samples. The solar position sensing auxiliary task is also a continuous value regression task, employing the mean squared error loss function. This auxiliary task guides the model to actively learn the fundamental astrophysical laws corresponding to photovoltaic power output, strengthening the model's physical understanding of diurnal variations and seasonal changes in illumination.

[0030] When constructing the cloud field disturbance identification auxiliary task samples, the same meteorological element input features are reused. The four levels of cloud field disturbance level labels (no disturbance, light, moderate, and severe) quantified and labeled at the corresponding time steps are used as the true labels for the auxiliary classification task, thus forming the cloud field disturbance identification auxiliary task samples. The cloud field disturbance identification auxiliary task is a multi-classification task, and the cross-entropy loss function is used to quantify the classification error. This auxiliary task is used to guide the model to identify the irradiance disturbance characteristics caused by cloud field movement, improving the model's adaptability to drastically fluctuating conditions.

[0031] Three types of single-task samples generated at the same time step are bound and fused, so that the three types of task samples share the same set of meteorological input features, with only the task supervision labels being independent of each other, forming a single set of multi-task joint training samples. All time-series samples in the dataset are traversed to generate a sufficient number of multi-task joint training samples with balanced operating conditions in batches, and the samples are summarized to form a complete multi-task joint training sample set.

[0032] Step 103: Use multi-task joint training samples to pre-train the general generative large model in a multi-task domain to obtain the photovoltaic generative large model, specifically including: A pre-trained general generative large model with general time series generation capabilities was selected as the base model. This general generative large model is an autoregressive time series generation model based on a Transformer decoder-only architecture, containing L layers of Transformer decoding layers, each layer containing a multi-head self-attention mechanism and a feedforward neural network, with a hidden layer dimension of D. The input and output layers of the base model were adapted: the 32-layer Transformer decoding layer backbone structure of the base model was retained, with each layer containing a multi-head self-attention mechanism with 32 attention heads and a feedforward neural network with a hidden layer dimension of 11008; the input embedding layer was replaced with a linear projection layer adapted to photovoltaic meteorological time series characteristics, and three parallel output projection heads were added, corresponding to power prediction, solar position sensing, and cloud field disturbance identification tasks, respectively, enabling the model to adapt to photovoltaic time series meteorological input characteristics and support multi-task parallel output. The modified model uses a single common input layer to receive meteorological features and multiple independent output heads, corresponding to the main task output of power prediction, the auxiliary task output of solar position perception, and the auxiliary task output of cloud field disturbance identification. The network parameters of each branch are updated independently and do not interfere with each other. The newly added common input layer maps meteorological feature vectors to a 4096-dimensional hidden layer space; the three independent output heads are all single-layer fully connected networks, mapping the hidden layer features to power prediction values ​​(1-dimensional), solar position parameters (2-dimensional, including solar altitude angle and solar azimuth angle), and cloud field disturbance category probabilities (4-dimensional), respectively.

[0033] A weighted multi-task joint loss function is constructed, where the overall loss is the weighted sum of the main task loss and the losses of the two types of auxiliary tasks. The weight of the power prediction main task loss is set to... The loss weight for the solar position perception auxiliary task is set to... The loss weight for the cloud field disturbance identification auxiliary task is set to... The total weight of the three is 1. The high weight of the main task ensures the core predictive ability of the model, while the low weight of the auxiliary task implements physical prior constraints to avoid the model overfitting data noise and local outliers.

[0034] Multi-task joint loss function ; In the formula, This represents the total training loss for a single iteration of the model. For power prediction, the main task loss is calculated. Loss of solar position sensing auxiliary mission; This method aids in identifying cloud field disturbances and reduces task losses. Through this weighted fusion approach, the model can be simultaneously constrained to learn the temporal patterns of photovoltaic power output, astrophysical laws, and cloud field disturbance characteristics, achieving multi-task collaborative optimization.

[0035] The entire historical operational dataset was divided into training and validation sets based on the stations. 80% of the station data was randomly selected as the training set, and the remaining 20% ​​as the validation set. Within each station, the first 80% of the time period was used for training, and the last 20% for testing. Training samples were batch-extracted from the multi-task joint training sample set and input into the modified general generative model. The model reads meteorological element features through a common input layer and performs unified normalization. The Min-Max normalization method is used to linearly map the original values ​​of each meteorological element to the [0,1] interval. The normalized features are input into three independent task output heads to complete forward inference calculations. The power prediction output head and the solar position sensing output head obtain continuous prediction values ​​through linear operations of the network weight matrix and mapping with a nonlinear activation function. The general forward inference calculation formula is: ; In the formula, The forecast output is continuous, corresponding to forecast power, forecast solar altitude angle, and forecast solar azimuth angle. This is the network weight matrix; This is a bias parameter, and its initial value is uniformly set to 0; Softmax is a non-linear activation function used to fit non-linear mapping relationships. After feature transformation, the cloud field disturbance recognition output head is fed into the Softmax activation function for probability normalization. Predicted probabilities are calculated for each category. The Softmax calculation formula is: ; In the formula, The first output of the network The original score for the perturbation level is obtained linearly from the output head network layer, i.e. ; , The corresponding output headers for each category are: Task-specific weight matrices and bias parameters, This is a single-category index marker, referring to any single category among the four cloud field disturbance levels. Values ​​are 1, 2, 3, and 4, corresponding to no disturbance, slight cloud field disturbance, moderate cloud field disturbance, and severe cloud field disturbance, respectively. It is a natural constant; It represents the total number of disturbance level categories, indicating the total number of all categories of cloud field disturbances. For the sample prediction, the first The probability of cloud-like field disturbance level. The final predicted level is determined based on the maximum probability criterion. ; In the formula, To ultimately predict the level of cloud field disturbance, It is the first The probability that a sample is predicted to be without perturbation. , , These represent the probabilities of predicting mild, moderate, and severe cloud field disturbances, respectively. After the model simultaneously outputs the predicted power value, predicted solar position parameters, and predicted cloud field disturbance level, the prediction results of each branch and the true labels of the samples are substituted into the corresponding loss functions to calculate the power prediction loss, solar position perception loss, and cloud field disturbance recognition loss in sequence. Then, using a fixed weight, these losses are weighted and summed using the total loss calculation formula to obtain the overall training loss of the model for the current batch.

[0036] The backpropagation algorithm is executed based on the overall training loss. The stochastic gradient descent algorithm is used to iteratively update all network weights and bias parameters of the general generative large model layer by layer. The parameter iterative update formula is as follows: , ; In the formula, , For the first The current weight matrix and bias parameters for each iteration; , For the first The weight matrix and bias parameters updated after each iteration; The learning rate is fixed and is set to 0.001. , These are the gradients of the total loss with respect to the current weights and biases, respectively. The network parameters are continuously adjusted through gradient backpropagation, constantly optimizing the model's feature extraction and temporal generation capabilities. The training process involves repeatedly executing batch sample input, forward inference, loss calculation, and parameter iterative updates, resulting in multiple rounds of iterative training.

[0037] The model convergence criteria are set up using two indicators: the decrease in iterative loss and the prediction accuracy. The first indicator, the loss convergence indicator, is that the relative decrease in the total training loss of each iteration over 20 consecutive iterations is less than 0.1% compared to the total loss of the previous iteration. The second indicator, the accuracy convergence indicator, is that the mean absolute error (MAE) of the power prediction in the validation set does not reach a new low over 10 consecutive iterations, and the fluctuation of the MAE in the most recent 5 iterations does not exceed 5%. When both conditions are met, the model is considered to have converged, iterative training is stopped, and all network parameters and multi-task inference structures of the model are solidified. The dedicated generative model after completing the multi-task pre-training in the photovoltaic field is the photovoltaic generative large model. This model fully learns the meteorological mapping law, astrophysical law, and cloud field perturbation evolution law of photovoltaic power output, and has the time series generation and feature perception capabilities specific to the photovoltaic field.

[0038] This embodiment involves the complete and time-series alignment of raw data from multiple geographically diverse power stations, covering various actual operating conditions. This avoids the scenario bias problem caused by a single data source and addresses the shortcomings of traditional single-source data training models in terms of generalization ability. Standardized solar position parameter numerical calculations and annotations are completed uniformly across all time steps, and cloud field disturbance level labels are automatically assigned based on irradiance fluctuation quantification indicators, providing the model with accurate astronomical and cloud disturbance supervision information. A three-task-bound joint training sample is constructed, with power prediction as the main task and two types of physical attribute auxiliary tasks for collaborative supervision training. This guides the model to simultaneously learn the underlying astrophysical laws of photovoltaics and the characteristics of cloud field disturbance fluctuations, overcoming the problem that general-purpose large models can only learn surface data correlations and lack physical understanding of the photovoltaic field. It also compensates for the shortcomings of traditional pure data-driven models that cannot adapt to complex operating conditions such as component dust accumulation, degradation, and local shading. The domain pre-training is completed by using weighted multi-task loss joint optimization. The main and auxiliary tasks mutually constrain each other to correct the model parameters, reducing the tendency of the model to overfit short-term local extreme cloud cover anomalies. The pre-trained photovoltaic generative large model can be adapted to various newly built photovoltaic power plants without fine-tuning the target plant's local historical data, thus alleviating the technical problems of overfitting and prediction accuracy distortion in the cold start stage of newly built power plants from the source.

[0039] In a preferred embodiment of the present invention, steps 2 and 3 include: Step 200: Receive static parameters uploaded by the target power plant, including the power plant's geographical coordinates, installed capacity, module type, and support tilt and azimuth parameters, specifically including: The system receives and verifies the fixed static parameters uploaded by the target photovoltaic power station in real time. These parameters are inherent static attributes of the power station after its construction is finalized, remaining constant throughout its entire lifecycle and requiring no real-time updates. All parameters adopt the standard values ​​set in the power station's grid connection filing archive. First, the geographical coordinates of the power station: the longitude of the target photovoltaic power station's center is set to range from 73°E to 135°E, and the latitude to range from 18°N to 54°N, in degrees, used to lock the spatial reference of the power station and match the location of numerical weather forecast grid points. Second, the installed capacity: assuming the rated installed capacity of a single station ranges from 0.1MW to 100MW, used to limit the maximum theoretical power generation output of the power station, serving as the benchmark for subsequent abnormal power sample screening. Third, photovoltaic module type parameters: the peak power of the modules ranges from 350W to 700W, and the photoelectric conversion efficiency ranges from... The range of the three parameters is as follows: 18% to 23%; temperature coefficient range of -0.32% / ℃ to -0.42% / ℃; all parameters are fixed calibration values ​​from the module's factory. Fourth, the photovoltaic support tilt angle, a fixed angle between the module's installation plane and the horizontal ground, in degrees, ranging from 0° to 60°, is determined by the power station construction filing and directly determines the effective angle for receiving solar radiation. Fifth, the photovoltaic support azimuth angle, a fixed angle between the module's normal direction and due north, in degrees, ranging from 0° to 360°, with a standard due south orientation of 180°, used to characterize the module's fixed orientation. All collected and verified static parameters are uniformly entered into the system database for permanent storage, and the parameter values ​​remain fixed throughout the process.

[0040] Step 300: Obtain historical long-term meteorological statistics for the target power station area, and extract the prevailing wind direction for the target power station area in the corresponding season or month from the historical long-term meteorological statistics, specifically including:

[0041] Long-term historical meteorological statistics for the climate zone and administrative region where the target photovoltaic power station is located are retrieved. The dataset spans at least ten years and includes hourly wind direction and speed time-series observation records, covering the complete meteorological variation patterns of the 12 months and four seasons of the year, avoiding statistical bias caused by short-term meteorological fluctuations. A seasonal month matching rule is used to clearly define the corresponding season and month: spring (March, April, May), summer (June, July, August), autumn (September, October, November), and winter (December, January, February). Based on the specific calendar month of the time to be predicted, the corresponding season is matched, and all wind direction time-series samples from the historical data of the same month and season as the time to be predicted are selected to exclude interference from out-of-contemporary meteorological samples. The meteorological wind direction is divided into 16 wind direction sectors, each with a standard wind direction sector of 22.5°, and the frequency and cumulative duration of samples in each sector are statistically analyzed. The duration percentage of each wind sector was calculated separately. The single wind sector with the largest cumulative duration percentage and the highest frequency of occurrence within the same month and season was determined as the statistically dominant wind direction in the area where the target power station is located. This statistically dominant wind direction represents the steady-state characteristics of the regional climate and characterizes the mainstream direction of normalized cloud movement and atmospheric disturbance transmission. It serves as the deployment benchmark for initial optical field perturbation sentinel detectors, ensuring that detector deployment conforms to the long-term cloud field movement patterns in the region.

[0042] Step 301: Using the grid point where the target power station's geographical coordinates are located as a reference, deploy a first optical field perturbation sentinel on the upwind side along the prevailing wind direction and a second optical field perturbation sentinel on the crosswind side; the first, second, and third optical field perturbation sentinels correspond to different grid point locations in the numerical weather prediction grid data, respectively. Extract the irradiance forecast value corresponding to each sentinel's grid point from the numerical weather prediction grid data, specifically including: The reference grid point of the numerical weather prediction dataset is matched with the geographic coordinates of the target photovoltaic power station. Specifically, the numerical weather prediction grid point with the closest Euclidean spatial distance to the center of the target photovoltaic power station area is selected as the reference grid point to lock the spatial origin of the entire probe deployment. A uniform 0.25°×0.25° latitude and longitude numerical weather prediction grid resolution is adopted. All probe deployment, translation, and distance calculation are performed based on this unified grid system, and the grid coordinate system, projection method, and spatial resolution remain consistent throughout. Combining the regional statistical dominant wind direction obtained in step 300, the standardized deployment of three light field perturbation sentinel probes is completed according to the spatial differences of upwind and crosswind directions. Each probe independently corresponds to a different forecast grid point, with no grid overlap or position intersection. Specifically, the deployment involves placing a second optical field perturbation sentinel detector along the directly upwind direction of the statistically dominant wind direction, starting from the power station's baseline grid point. A second optical field perturbation sentinel detector is then placed at intervals of one standard grid point, and a first optical field perturbation sentinel detector is placed at intervals of two standard grid points. This first optical field perturbation sentinel detector is positioned one standard grid point away from the power station's baseline grid point compared to the second, forming a gradient forward monitoring structure. This structure is used to layer and proactively capture mainstream cloud field disturbances approaching the power station from a distance. A third optical field perturbation sentinel detector is placed one standard grid point away from the power station's baseline grid point, using the vertical lateral direction of the statistically dominant wind direction as the lateral wind direction. This third detector is used to monitor irradiance disturbances caused by unconventional lateral cloud intrusions, compensating for the spatial blind spots of single upwind monitoring. These three detectors form a spatial deployment structure with a two-layer gradient monitoring system in the upwind direction and auxiliary monitoring in the lateral wind direction, comprehensively covering the high-probability cloud field disturbance input paths around the power station.

[0043] The first, second, and third light field perturbation sentinel probes uniquely correspond to three independent, spatially non-overlapping grid points in the numerical weather prediction grid data. These three probe grid points belong to the same numerical weather prediction grid system as the power station's baseline grid point, with completely unified spatial resolution and projected coordinate system, ensuring no systematic bias in spatial location matching. The preset forecast timescale is fixed at 0h to 24h for short-term and ultra-short-term forecasts. This value is determined according to the industry standard for ultra-short-term photovoltaic power prediction and is suitable for the characteristics of rapid intraday cloud field evolution and frequent short-term disturbances. After the probes are deployed, the system reads the numerical weather prediction grid data for future periods in batches at 15-minute time steps according to the preset forecast timescale from 0h to 24h. The time-by-time total horizontal irradiance forecast values ​​for the grid points corresponding to the three sentinel probes are extracted, with the unit uniformly set to W / m². The collected multi-point, multi-time series irradiance data constitute the original observation sequence of spatial irradiance around the power station.

[0044] Step 302: Obtain the forecast cloud cover value for the future period from the numerical weather prediction grid data. When the forecast cloud cover value exceeds the preset cloud cover criterion, connect the first, second, and third light field perturbation sentinel elements sequentially to form a closed sampling domain. Obtain the irradiance forecast value of each grid point corresponding to the first, second, and third light field perturbation sentinel elements within the closed sampling domain, as well as the geographic latitude and longitude coordinates of each grid point in the numerical weather prediction grid data. Based on the geographic latitude and longitude coordinates of the grid points corresponding to each perturbation sentinel element, determine the spherical triangle formed by the three grid points on the Earth's surface, and calculate the area of ​​the spherical triangle as the spherical area of ​​the closed sampling domain. The area of ​​the spherical triangle is... The product characterizes the spatial extent of the closed sampling domain on the Earth's surface, and this spatial extent determines the time scale required for the cloud field to traverse the closed sampling domain. The irradiance forecast values ​​at corresponding grid points of the first, second, and third light field perturbation sentinel elements are arithmetically averaged to obtain the spatial mean of the irradiance forecast values ​​within the closed sampling domain. Based on the deviation between the spatial mean and the irradiance forecast values ​​at corresponding grid points of each sentinel element, a statistic characterizing the dispersion of the irradiance forecast values ​​at each grid point relative to the spatial mean is calculated. The ratio of this dispersion statistic to the spatial mean is used as the spatial variation coefficient of irradiance. This spatial variation coefficient characterizes the degree of non-uniformity in the spatial distribution of irradiance within the closed sampling domain caused by cloud field movement, specifically including: Cloud cover values ​​for the entire region for a predetermined time period are read from numerical weather prediction grid data. The predetermined time period is uniformly taken as the ultra-short-term forecast period from 0h to 24h, with a time sampling step of 15 minutes. The predetermined cloud cover threshold is set to 0.3, which is determined according to the general statistical standard for cloud disturbance identification in the photovoltaic industry. The normalized cloud cover range is 0 to 1. Cloud cover less than or equal to 0.3 is judged as clear sky with few clouds, with no significant spatial disturbance of irradiance; cloud cover greater than 0.3 is judged as cloudy with cloud transit, indicating that there is a significant risk of cloud field activity and irradiance disturbance in the target area, and the closed sampling domain construction process is initiated. The three grid positions corresponding to the first, second, and third light field perturbation sentinel probes are connected end to end to form a closed triangular sampling domain. This sampling domain covers the core cloud field disturbance-affected area around the target power station.

[0045] Real-time irradiance forecasts were collected for grid points corresponding to three sentinel elements within a closed sampling domain. Simultaneously, the geographic longitude and latitude coordinates of each grid point were acquired. Based on the Earth's spherical geometry, the spherical triangle formed by the three grid points on the Earth's surface was calculated using their coordinates. The spherical area of ​​the closed sampling domain was then calculated using the formula for calculating the area of ​​a spherical triangle, assuming the Earth's average radius is a fixed value. =6371km, the spherical coordinates of the three probe grid points are respectively ,in Latitude The longitude is used. The lengths of the three spherical sides are calculated using the spherical distance formula. Then, calculate the area of ​​the spherical triangle using Heron's formula for spherical triangles. The calculation formula is: ; In the formula, The semi-perimeter of the spherical triangle The spherical area of ​​the closed sampling domain is expressed in square kilometers. The size of the spherical area directly represents the spatial extent of the closed sampling domain. The larger the spatial extent, the longer the cloud field travels through the region, and the longer its influence on the irradiance within the region lasts.

[0046] The spatial mean of irradiance in the closed sampling domain is obtained by arithmetically averaging the predicted irradiance values ​​at three grid points: ; In the formula, This is the spatial mean of irradiance, in W / m². 2 ; These represent the predicted irradiance values ​​corresponding to the three detector grid points. Based on the deviation between the spatial mean and the predicted irradiance values ​​at each grid point, the dispersion statistics of the irradiance data are calculated, using the standard deviation as the statistical index of dispersion. The calculation formula is as follows: ; In the formula, For the first The predicted irradiance value corresponding to each detector grid point The standard deviation of the irradiance forecast values ​​characterizes the dispersion of data at each grid point relative to the spatial mean; a larger value indicates a more significant difference in the spatial irradiance distribution. The ratio of the dispersion statistic to the spatial mean is defined as the irradiance spatial variation coefficient. The irradiance spatial variation coefficient is used to quantify the degree of non-uniformity in the spatial distribution of irradiance within a closed sampling domain caused by cloud field movement. A larger coefficient of variation indicates a more non-uniform spatial distribution of the cloud field and a higher risk of disturbance to photovoltaic power output.

[0047] Step 303: Obtain the prevailing wind direction forecast value for the future period from the numerical weather prediction grid data. When the deflection angle of the prevailing wind direction forecast value relative to the prevailing wind direction on which the sentinel was deployed exceeds a preset angle criterion, the third light field perturbation sentinel is translated along the deflected new wind direction using the spatial resolution of the numerical weather prediction grid data as the translation step size; obtain the grid position of the translated third light field perturbation sentinel in the numerical weather prediction grid data; calculate the grid position of the translated third light field perturbation sentinel and the first light field perturbation sentinel respectively. The first distance between the grid points of the sentinel and the second distance between the grid points of the second light field perturbation sentinel; the light field perturbation sentinel corresponding to the smaller of the first and second distances is selected as the connection object; the translated third light field perturbation sentinel and the corresponding grid point of the connection object in the numerical weather prediction grid data are connected by a straight line, and the sequence of all grid points along the straight line is used as the oblique sampling path; the irradiance forecast values ​​of each grid point along the oblique sampling path are extracted to form an irradiance spatial gradient sequence, specifically including: The system dynamically corrects the placement of sentinel probes based on intraday ultra-short-term forecast data, adapting to real-time wind and cloud path offset conditions. Data reading timeframes are uniformly set from 0h to 24h, with a fixed sampling time step of 15 minutes. The system reads the real-time forecast prevailing wind direction from the numerical weather prediction grid data hourly, compares this forecast prevailing wind direction with the seasonal statistical prevailing wind direction obtained in step 300, and calculates the absolute deflection angle of the two wind directions using the wind direction deflection angle calculation formula. The statistical prevailing wind direction angle is set as... The real-time forecast of the prevailing wind direction angle is All units are degrees, ranging from 0° to 360°, representing the absolute deflection angle of the wind direction. The calculation formula is: ; In the formula, The final absolute wind direction deflection angle, ranging from 0° to 180°, is used to characterize the true deflection angle between two wind directions, avoiding angle calculation errors caused by 360° circular wind direction jumps. The preset wind direction deflection angle criterion threshold is fixed at 22.5°, based on the meteorological standard interval of wind direction sector 16. When the absolute wind direction deflection angle... When the angle is greater than 22.5°, it indicates a substantial deflection of the prevailing wind field in the region, with a significant shift in the cloud field's movement path, necessitating adaptive translation adjustment of the detector element; when the absolute wind direction deflection angle is greater than 22.5°, it indicates a substantial deflection of the prevailing wind field in the region, with a significant shift in the cloud field's movement path, requiring adaptive translation adjustment of the detector element. When the angle is less than or equal to 22.5°, it is determined that there is no significant shift in the wind field and the cloud field intrusion path is stable. The original probe deployment position is retained, and no position adjustment operation is performed.

[0048] The probe translation is performed according to the numerical weather prediction grid system. The spatial resolution of the prediction grid points is fixed at 0.25° × 0.25° latitude and longitude, and the single-grid spacing is used as the unique translation step size. When it is determined that the position needs to be adjusted, the entire third light field perturbation sentinel probe is translated along the new wind direction after deflection by one standard grid point step to obtain the new latitude and longitude grid point position corresponding to the translated third light field perturbation sentinel probe. Based on the Earth's spherical geometric distance formula, the spherical straight-line distance between the translated third light field perturbation sentinel probe and the first light field perturbation sentinel probe is calculated and denoted as the first distance D1; the spherical straight-line distance between the translated third light field perturbation sentinel probe and the second light field perturbation sentinel probe is calculated and denoted as the second distance D2. The spherical distance calculation uniformly adopts the Earth's average radius. =6371km.

[0049] Numerical comparisons were performed between the first distance D1 and the second distance D2. The optical field perturbation sentinel probe with the smaller distance was selected as the sole connection object, ensuring that the sampling path closely matched the shortest intrusion path of the cloud field and had the strongest spatial correlation. The grid points of the translated third optical field perturbation sentinel probe were then connected to the grid points of the selected connection object with a straight line. All numerical weather forecast grid points along this straight line path were extracted and arranged in order of wind direction and distance from farthest to nearest, forming a standardized oblique sampling path. The total horizontal irradiance forecast value for the same forecast time was extracted grid point by grid point along the oblique sampling path, with the unit uniformly set to W / m². These values ​​were arranged sequentially according to the path arrangement, forming a continuous and ordered spatial gradient sequence of irradiance, characterizing the spatial gradual change of irradiance during cloud field movement.

[0050] Step 304: The spatial variation coefficient of irradiance and the spatial gradient sequence of irradiance are used as prior information for cloud field perturbation quantization. These are fused with static parameters to obtain enhanced physical prior cue words. If any of these terms is not calculated, the corresponding term is considered non-existent, and only existing terms are used as prior information for cloud field perturbation quantization. If neither is calculated, the prior information for cloud field perturbation quantization is empty. Specifically, this includes: The spatial variation coefficient of irradiance obtained by quantification in step 302 through arithmetic mean and standard deviation mean normalization, and the spatial gradient sequence of irradiance obtained by continuous sampling along the cloud field deflection path in step 303 are uniformly defined as prior information for cloud field disturbance quantification. The spatial variation coefficient of irradiance is obtained by quantifying the ratio of the spatial dispersion of irradiance at multiple points to the spatial mean. Its numerical value quantitatively characterizes the spatial non-uniformity of irradiance within the closed sampling domain; a larger coefficient of variation indicates more significant differences in local cloud shading and a more disordered spatial distribution of irradiance. The spatial gradient sequence of irradiance is constructed by continuous sampling of irradiance along the dynamic movement path of the cloud field. The continuous rise and fall of the sequence values ​​characterizes the continuous spatial gradual change and disturbance evolution of irradiance brought about by the movement and intrusion of the cloud field, reflecting the movement direction, coverage area, and disturbance strength of the cloud field. The two types of cloud field disturbance quantification prior information are dimensionally fused with the fixed static parameters of the target photovoltaic power station collected in step 200. The inherent physical properties of the power station and the spatiotemporal characteristics of real-time cloud field disturbance are integrated to construct an enhanced physical prior cue word that is adapted to multiple meteorological conditions.

[0051] A fixed and unique adaptive fusion judgment rule is adopted, which is compatible with all meteorological scenarios such as clear sky, few clouds, cloudy, and sudden wind direction changes. The rule boundaries are clear. When only the cloud cover criterion is met and the spatial variation coefficient of irradiance is calculated, but the wind direction deflection adjustment logic is not triggered and there is no spatial gradient sequence of irradiance, only the spatial variation coefficient of irradiance is used as the prior information for effective cloud field disturbance quantification in the fusion. When only the wind direction deflection adjustment logic is triggered and the spatial gradient sequence of irradiance is generated, but the cloud cover criterion is not met and the spatial variation coefficient of irradiance is not calculated, only the spatial gradient sequence of irradiance is used as the prior information for effective cloud field disturbance quantification in the fusion. When both the cloud cover criterion and the wind direction deflection criterion are met, the spatial variation coefficient of irradiance and the spatial gradient sequence of irradiance are used together in the fusion. When neither the cloud cover criterion nor the wind direction deflection adjustment logic is met, it is determined that there are no cloud field disturbance characteristics in the current area, the prior information for cloud field disturbance quantification is empty, and only basic physical prior prompts are generated based on the static parameters of the power station to achieve prior information output that is compatible with all operating conditions.

[0052] This embodiment combines the static inherent parameters of the power station with the regional meteorological and spatiotemporal characteristics. Based on a dual-dimensional deployment of historical statistical wind direction and real-time forecast wind direction, it adaptively adjusts the light field perturbation sentinel probes. This overcomes the technical shortcomings of traditional fixed monitoring points that cannot adapt to wind direction deflection and cloud field path shifts, improving the targeting and accuracy of cloud field disturbance monitoring around photovoltaic power stations. By constructing a spherical triangular closed sampling domain, the spatial area of ​​the sampling domain is accurately calculated based on spherical geometry principles. The spatial mean of irradiance is calculated using an arithmetic mean, closely conforming to the real geographical spatial characteristics of the Earth's sphere and avoiding spatial errors caused by planar calculations. Simultaneously, the coefficient of variation is used to quantify the spatial non-uniformity of irradiance, quantitatively characterizing the spatial distribution disturbance characteristics of the cloud field. This solves the problem that traditional technologies can only monitor single-point irradiance and cannot reflect overall spatial disturbance differences. The design incorporates an adaptive probe translation mechanism for wind direction deflection and a method for constructing oblique sampling paths. This allows for dynamic adjustment of the monitoring range and sampling path based on real-time wind direction changes, effectively capturing continuous spatial gradient changes in irradiance after cloud field deflection. It fully reconstructs the dynamic evolution of irradiance during cloud field movement, overcoming the limitations of traditional static sampling methods that cannot adapt to dynamic meteorological changes, thus improving the completeness and timeliness of data acquisition. An adaptive prior information fusion strategy is employed to dynamically filter effective cloud field disturbance prior information and fuse it with the power plant's static parameters. The constructed enhanced physical prior cue words possess both the inherent physical characteristics of the power plant and the spatiotemporal features of real-time cloud field disturbances. This provides photovoltaic power prediction models with adaptable physical prior constraints for various operating conditions, effectively improving the accuracy of photovoltaic power prediction under complex weather conditions.

[0053] In a preferred embodiment of the present invention, step 4 includes: Step 400: Using the enhanced physical prior cue words as a prefix context, the meteorological forecast data for the target power plant's future time period, divided according to a preset time resolution, is concatenated segment by segment into the prefix context to form a structured input sequence; wherein, the meteorological forecast data includes irradiance, temperature, wind speed, cloud cover, and wind direction forecast elements for each time step, specifically including: The enhanced physical prior cue words generated through fusion are used as prefix context information for model inference. This prefix context is fully embedded with the inherent static parameters of the target photovoltaic power station and the quantified cloud field perturbation prior information. The inherent static parameters include the geographical coordinates of the power station, installed capacity, module type, support tilt angle, and support azimuth angle. The quantified cloud field perturbation prior information includes the spatial variation coefficient of irradiance and the spatial gradient sequence of irradiance. A standardized ultra-short-term forecast configuration is uniformly adopted, with the future forecast period preset to be fixed from 0h to 24h and the time resolution preset to be fixed at 15min. The complete 24-hour forecast duration is evenly divided into 96 continuous and equally spaced time series steps to achieve fine-grained intraday time series sampling.

[0054] Numerical weather forecast data for the entire region is extracted step-by-step. Each time step contains five core meteorological forecast elements, all of which are average values ​​within their respective 15-minute time steps: total horizontal irradiance (W / m²), ambient temperature (°C), near-surface 10-meter wind speed, total cloud cover (15-minute average, ranging from 0 to 1), and near-surface 10-meter wind direction (degrees, ranging from 0° to 360°). Following a temporal progression rule, the time-step-normalized meteorological forecast data is sequentially concatenated to the prefix context of the enhanced physical prior cue words, ultimately forming a structured input sequence with physical information at the front, temporal features at the back, dimensional alignment, and orderly arrangement. This structured input sequence achieves a deep fusion of the inherent physical properties of the power plant, cloud field spatial disturbance characteristics, and intraday dynamic meteorological temporal characteristics, providing a standardized and physically interpretable input basis for the autoregressive inference of the photovoltaic generative large-scale model.

[0055] Step 401: Input the structured input sequence into the photovoltaic generative large model. Based on the static parameters embedded in the prefix context and the prior information quantified by cloud field perturbation, the photovoltaic generative large model performs autoregressive inference on the meteorological forecast data at each time step, and outputs the corresponding power prediction value for each time step. Arrange the power prediction values ​​output at each time step in chronological order to obtain the preliminary predicted power sequence, which specifically includes: The complete structured input sequence constructed in step 400 is then input into the photovoltaic generative large-scale model that has completed training and convergence. The photovoltaic generative large-scale model relies on multi-dimensional physical prior information embedded in the pre-prefix context of the structured input sequence, such as the power plant's static parameters, irradiance spatial variation coefficient, and irradiance spatial gradient sequence, to solidify the intrinsic physical mapping relationship between photovoltaic power output and astronomical location, spatial cloud field disturbances, meteorological dynamics, and the inherent characteristics of the power plant. The structured input sequence is defined as... , ,in The physical prior feature vector corresponding to the enhanced physical prior cue words is used to fix the global physical constraint parameters; This is a time-step meteorological forecast feature vector used to carry the intraday dynamic meteorological evolution characteristics. The model first performs unified dimensional mapping and feature fusion calculation on the input features to complete the pre-feature purification and provide standardized feature input for subsequent power numerical calculation.

[0056] The model employs a time-step autoregressive inference mechanism with physical prior constraints to solve for power numerical values. It follows an iterative deduction logic of prior feature constraints on time-series meteorological features and preceding power output constraints on subsequent power output. The single-time-step power prediction value is obtained through a two-layer operation of feature weighted fusion and nonlinear activation mapping. First, the fusion calculation of physical prior features and time-series meteorological features is performed. The feature fusion calculation formula is as follows: ; In the formula, For the t-th time step, the hidden layer features are fused. This is the physical prior feature weight matrix; This is the meteorological time series feature weight matrix; For the first Weather forecast feature vector at the time step; Here are the hidden layer bias parameters. Based on this, a temporal iteration constraint is introduced, and the power nonlinear inference for the current time step is completed by combining the predicted output from the previous time step. The complete single-step power prediction calculation formula is: ; In the formula, Here is the predicted photovoltaic power at time step t of the model, in MW. This is the nonlinear activation function for the model output layer, used to fit the nonlinear relationship between photovoltaic power output and meteorological and physical characteristics; This is the output layer feature weight matrix; This is a temporal correlation weight matrix, used to inherit the evolution pattern of output from previous time steps; These are the output layer bias parameters; This represents the power prediction value for the preceding adjacent time step. The model iterates frame by frame starting from the first time-series time step. The first time step has no preceding time-series output, and the default time-series constraint term is 0. The first power prediction value is calculated solely based on prior physical features and initial meteorological features. Subsequent time steps iterate by overlaying the preceding output time-series features, outputting the corresponding photovoltaic power prediction value for each time step. After all 96 time-series time steps have been inferred, the power prediction values ​​generated for each time step are sequentially arranged and integrated according to the original time-series order from 0h to 24h, forming a complete and continuous intraday preliminary predicted power sequence. This sequence contains the original inference results for the entire time period without physical rule corrections.

[0057] This embodiment constructs a structured input sequence that deeply integrates physical prior information with dynamic meteorological time-series characteristics, breaking the limitation of traditional photovoltaic prediction models that rely solely on single meteorological data. By embedding the static inherent attributes of the power station and the quantitative characteristics of cloud field spatial disturbances as pre-constraints into the model input, it provides clear physical mechanism support for model inference, enabling the model to perceive the personalized power generation characteristics of different power stations and complex cloud field disturbance conditions. This improves the model's adaptability and prediction accuracy from the input source, effectively reducing prediction bias under complex meteorological scenarios. Employing a photovoltaic generative large-scale model autoregressive time-series inference method, it relies on multi-dimensional physical prior constraints to achieve refined time-step output extrapolation, fully aligning with the time-series characteristics of rapid intraday cloud field evolution and dynamic fluctuations in irradiance. Compared to traditional fitting prediction and shallow machine learning prediction methods, this inference method can capture short-term, small-amplitude dynamic changes in photovoltaic output, improving the time-series fit of ultra-short-term photovoltaic power prediction.

[0058] In a preferred embodiment of the present invention, step 5 includes: Step 500, First Verification: Calculate the corresponding solar altitude angle for each time step in the preliminary predicted power sequence, and set the predicted power value for time steps with a solar altitude angle less than or equal to zero to zero, thus obtaining the first corrected sequence. This specifically includes: The preliminary predicted power sequence output in step 401 undergoes a first-level physical mechanism constraint verification and correction. Based on the objective physical laws of solar astronomical operation, false power generation prediction values ​​during invalid sunshine periods are eliminated to ensure that the power output results conform to the basic characteristics of day and night power generation. For each time step of the preliminary predicted power sequence, the corresponding forecast time's year, month, day, hour, minute, and time series information, as well as the fixed geographical latitude and longitude coordinates of the target power station, are matched. The real-time solar altitude angle is calculated step by step. The solar altitude angle is uniformly solved using the spherical trigonometric astronomical calculation formula, as follows: ; In the formula, This is the real-time solar altitude angle, in degrees. The target photovoltaic power station has a fixed geographical latitude, in degrees. The real-time solar declination angle at the corresponding moment, in degrees; The value is the real-time solar hour angle at the corresponding moment, in degrees.

[0059] Based on astrophysical characteristics, a fixed judgment rule is set. When the solar altitude angle is less than or equal to 0°, the sun is at or below the horizon at the corresponding time, and there is no effective solar shortwave radiation incident on the Earth's surface. The photovoltaic modules cannot complete photoelectric conversion, and there is no effective power generation output, which is an absolutely ineffective power generation period. The entire time step of the preliminary predicted power sequence is iterated one by one, and the solar altitude angle at each time step is judged. The power prediction value for the corresponding time step with a solar altitude angle less than or equal to 0° is uniformly corrected to 0. For effective sunshine periods with a solar altitude angle greater than 0°, the corresponding power prediction value remains unchanged from the original inference value. After all time steps are verified, the first corrected sequence is obtained, eliminating false power output at night and during dawn and dusk.

[0060] Step 501, the second verification, involves truncating the power values ​​at each time step in the first correction sequence. Power values ​​greater than the upper limit of the installed capacity threshold are truncated to the upper limit, and power values ​​less than zero are truncated to zero, resulting in the second correction sequence, which specifically includes:

[0061] A second round of numerical boundary compliance verification and correction is performed on the first correction sequence obtained in step 500. This involves constraining the photovoltaic output within a reasonable numerical range based on the inherent fixed parameters of the target power plant, eliminating abnormal power values ​​such as exceeding upper limits or negative values ​​generated by the model inference, which are inconsistent with the actual operation of photovoltaic projects. This method limits the effective range of photovoltaic power prediction values ​​to 0MW to Pn, where Pn is the rated installed capacity of the target photovoltaic power plant, in MW. The rated installed capacity Pn of the target photovoltaic power plant adapted to this scheme has a fixed range of 0.1MW to 100MW. This range is determined based on the physical characteristics of photovoltaic power generation and grid connection operation specifications. The actual power generation output of the photovoltaic power plant cannot be lower than 0 or exceed the rated installed capacity; the range boundary possesses unique physical determinism and engineering compliance. The fixed upper limit threshold is the rated installed capacity Pn of the target power station for grid connection filing. The threshold value is a fixed value for grid connection filing during the construction phase of the power station. It belongs to the static inherent parameter of the power station and remains constant throughout the process. The lower limit threshold is fixed at 0MW. It is determined based on the physical mechanism of photoelectric conversion. The photovoltaic output is zero at the lowest level under no-sunlight and no-power-generating conditions. There is no negative output condition.

[0062] The system iterates through all power values ​​in the first correction sequence step by step, performing standardized boundary truncation. Specifically, the corrections are as follows: when the predicted power value at a given time step exceeds the upper limit threshold Pn of the power plant's rated installed capacity (ranging from 0.1MW to 100MW), the value is determined to be an abnormal output exceeding the upper limit, violating the power plant's maximum power generation capacity constraint. The power value at that moment is then truncated and corrected to the rated installed capacity threshold Pn. When the predicted power value at a given time step is less than 0MW, the value is determined to be an abnormal negative output value violating the physical logic of photoelectric conversion, and the power value at that moment is truncated and corrected to 0MW. When the predicted power value is within the valid range of 0MW to Pn, it is determined to be a compliant and normal output value, perfectly matching the actual power generation capacity of the power plant, and the original inferred value remains unchanged. After all time steps have been iterated and corrected, abnormal data for all time periods are removed, resulting in a second correction sequence with fully compliant numerical boundaries that conforms to engineering operation standards.

[0063] Step 502, the third verification, involves scanning the second corrected sequence through time windows, calculating the power change slope between adjacent time steps. When the power change slope between adjacent time steps within a certain time window exceeds a preset abrupt change threshold, the power values ​​of each time step within the corresponding time window are smoothed using a moving average to obtain the final predicted power sequence. Specifically, this includes:

[0064] A third time-series continuity check and correction is performed on the second corrected sequence obtained in step 501 to address the problems of pulse-like abrupt changes and short-term abnormal jumps that are prone to occur during model inference, thus aligning with the actual physical operation characteristics of photovoltaic power output, which varies continuously and gradually with weather conditions. Using a single 15-minute time step as the basic calculation unit, the power change rate is calculated for each adjacent time window. The degree of power output fluctuation is quantified by the power change slope. In the formula, For the first The slope of power change corresponding to each time step, in MW / min; The power value before correction at the current time step, in MW; This represents the power value of the previous adjacent time step, in MW. The time step is fixed and uniformly set to 15 minutes.

[0065] The preset power surge threshold is a relative threshold, defined as 0.02 × PnMW / min, where Pn is the rated installed capacity of the target photovoltaic power station (in MW). This threshold is obtained through statistical calibration after normalization of measured photovoltaic power samples. Specifically, historical intraday measured power time-series data of large-scale photovoltaic power stations with multiple regions and installed capacities are collected, and the power change slope kt is calculated every 15 minutes. Then, each slope is divided by the rated installed capacity of the corresponding power station to obtain the relative slope. Valid relative slope samples caused by the natural passage of clouds under all sunny and cloudy weather conditions are statistically analyzed, and the relative slope value corresponding to the 99th percentile of the normal samples is taken as a fixed proportion threshold (approximately 0.02 / min). This is used to quantitatively distinguish between compliant normal power output fluctuations caused by actual cloud movement and non-physical abnormal surges without physical basis generated by model iterative inference.

[0066] The slope of power change across the entire sequence is calculated by iterating through each time window. For time windows with an absolute slope value greater than 0.02 × PnMW / min, non-physical power abrupt noise is identified, triggering a smoothing correction mechanism. For abrupt change periods, a central moving average algorithm with a window length of 3 is used for smoothing correction to eliminate instantaneous pulse anomalies. The moving average correction calculation formula is as follows: ; In the formula, For the revised version Power value per time step; , , These represent the original power values ​​for the preceding, current, and subsequent time steps of the abrupt change period. For time series windows where the absolute value of the power change slope is less than or equal to 0.02 × PnMW / min, it is determined to be a normal power output fluctuation, and the power value for the corresponding period remains unchanged. After full sequence traversal correction, while fully preserving the effective power output fluctuation characteristics caused by cloud field disturbances, abnormal abrupt change noise in the model is filtered out, resulting in a final predicted power sequence that is time-continuous, numerically compliant, physically logically complete, and closely matches the actual operating patterns of the power plant.

[0067] This embodiment designs a three-tiered, multi-dimensional physical verification and correction system, with each layer constraining and optimizing the prediction results step by step, achieving comprehensive compliance correction from the perspectives of physical mechanisms, numerical boundaries, and temporal continuity. The first layer is the zero-value correction of the solar altitude angle, which follows astronomical laws and eliminates false power generation predictions during nighttime and twilight periods, ensuring that the prediction results conform to the basic physical logic of photovoltaic day and night power generation. The second layer is the installed capacity boundary truncation correction, which relies on the inherent rated parameters of the power station to constrain the output range, eliminating abnormal engineering values ​​such as overcapacity and negative output, and ensuring the engineering practicality of the prediction results. The third layer is the temporal smoothing correction, which distinguishes between effective cloud field fluctuations and model inference noise, eliminating non-physical pulse abrupt changes while retaining the characteristics of real output disturbances, and conforming to the continuous and gradual operating characteristics of photovoltaic output.

[0068] like Figure 2 As shown, embodiments of the present invention also provide a photovoltaic power prediction system based on a generative large model, comprising: The model training module is used to pre-train a general generative large model on a massive amount of historical operation data of photovoltaic power plants in multiple tasks and domains to obtain a photovoltaic generative large model. The data receiving module is used to receive static parameters uploaded by the target power station, including the power station's geographical coordinates, installed capacity, component type, and support tilt and azimuth parameters. The prompt word construction module is used to extract quantitative prior information on cloud field disturbances around the target power station based on static parameters and gridded data of future time-period numerical weather forecasts obtained from the meteorological service interface, and to construct enhanced physical prior prompt words. The model prediction module is used to stitch together the enhanced physical prior cue words with the meteorological forecast data of the target power plant in the future period according to the preset time resolution, and then input the data into the photovoltaic generative large model to perform zero-sample power prediction inference to obtain the preliminary predicted power sequence of the target power plant in the future period. The processing and verification module is used to perform triple physical post-processing verification on the preliminary predicted power sequence, including nighttime zero-value forced constraint correction based on solar altitude angle calculation, output upper and lower limit truncation processing based on installed capacity threshold, and moving average smoothing processing for sections where the power change slope exceeds the preset abrupt change threshold within a continuous time window, to obtain the final predicted power sequence.

[0069] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0070] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0071] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0072] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A photovoltaic power prediction method based on a generative large model, characterized in that, The method includes: Step 1: Pre-train the general generative large model on massive historical operation data of photovoltaic power plants in multiple tasks. The multi-task joint training samples include power prediction main task samples, solar position perception auxiliary task samples and cloud field disturbance identification auxiliary task samples to obtain the photovoltaic generative large model. Step 2: Receive static parameters uploaded by the target power station, including the power station's geographical coordinates, installed capacity, component type, and support tilt and azimuth parameters; Step 3: Based on static parameters and gridded data from future time-period numerical weather forecasts obtained from the meteorological service interface, extract prior information on cloud field perturbation quantification around the target power station, and construct enhanced physical prior cue words. This includes: obtaining historical long-term meteorological statistics for the area where the target power station is located, extracting the statistical dominant wind direction for the corresponding season or month from the historical long-term meteorological statistics; and deploying the first optical field perturbation sentinel detector and the second optical field perturbation sentinel detector along the prevailing wind direction on the upwind side, using the gridded point where the geographical coordinates of the target power station are located as the reference. A third light field perturbation sentinel is deployed on the crosswind side. The first, second, and third light field perturbation sentinels correspond to different grid points in the numerical weather prediction grid data. The irradiance forecast value for each sentinel corresponding to the grid point is extracted from the numerical weather prediction grid data. The forecast cloud cover value for the future period is obtained from the numerical weather prediction grid data. When the forecast cloud cover value exceeds the preset cloud cover criterion, the first, second, and third light field perturbation sentinels are linked together. Sentinel elements are sequentially connected to form a closed sampling domain; the spatial variation coefficient of irradiance is calculated based on the irradiance forecast values ​​of the corresponding grid points of each sentinel element within the closed sampling domain; the prevailing wind direction forecast value for the future period is obtained from the numerical weather prediction grid data; when the deflection angle of the prevailing wind direction forecast value relative to the prevailing wind direction on which the sentinel elements were deployed exceeds a preset angle criterion, the third light field perturbation sentinel element is translated along the deflected new wind direction using the spatial resolution of the numerical weather prediction grid data as the translation step size; the translated third light field perturbation sentinel element is then compared with... The first or second light field perturbation sentinel is connected to form a slanted sampling path; the irradiance forecast values ​​of each grid point along the slanted sampling path are extracted to form an irradiance spatial gradient sequence; the irradiance spatial variation coefficient and the irradiance spatial gradient sequence are used as prior information for cloud field perturbation quantification, and fused with static parameters to obtain enhanced physical prior cue words. When one of them is not calculated, the corresponding term is considered non-existent, and only the existing term is used as the prior information for cloud field perturbation quantification; if neither is calculated, the prior information for cloud field perturbation quantification is empty; Step 4: After splicing the enhanced physical prior cue words with the meteorological forecast data of the target power plant in the future period according to the preset time resolution, input the data into the photovoltaic generative large model to perform zero-sample power prediction inference and obtain the preliminary predicted power sequence of the target power plant in the future period. Step 5: Perform triple physical post-processing verification on the preliminary predicted power sequence, including nighttime zero-value forced constraint correction based on solar altitude angle calculation, output upper and lower limit truncation processing based on installed capacity threshold, and moving average smoothing processing on the section where the power change slope exceeds the preset abrupt change threshold within the continuous time window, to obtain the final predicted power sequence.

2. The photovoltaic power prediction method based on a generative large model according to claim 1, characterized in that, Step 1 includes: Collect historical operation datasets from multiple photovoltaic power stations. The historical operation datasets for each photovoltaic power station include meteorological element sequences and corresponding actual power generation sequences for each time step within a historical period. The solar position parameters for each time step in the historical operation dataset are labeled with the solar position parameters, including the solar altitude angle and the solar azimuth angle. Cloud field disturbance labels are labeled for the abnormal irradiance fluctuation sections in the historical operation dataset. The cloud field disturbance labels are used to characterize the degree of influence of cloud field movement on the irradiance in the section, thus obtaining the labeled historical operation dataset. Based on the labeled historical running dataset, construct multi-task joint training samples; A multi-task domain pre-training method is used to train a general generative large model using multi-task joint training samples, resulting in a photovoltaic generative large model.

3. The photovoltaic power prediction method based on a generative large model according to claim 2, characterized in that, The spatial variation coefficient of irradiance is calculated based on the irradiance forecast values ​​of each sentinel element corresponding to the grid points within the closed sampling domain, including: Obtain the irradiance forecast values ​​of the corresponding grid points of the first, second, and third light field perturbation sentinel elements within the closed sampling domain, as well as the geographic latitude and longitude coordinates of each grid point in the numerical weather prediction grid point data; Based on the geographic latitude and longitude coordinates of the grid points corresponding to each perturbation sentinel, the spherical triangle formed by the three corresponding grid points on the Earth's surface is determined, and the area of ​​the spherical triangle is calculated as the spherical area of ​​the closed sampling domain; wherein, the area of ​​the spherical triangle represents the spatial distribution range of the closed sampling domain on the Earth's surface, and the spatial distribution range determines the time scale required for the cloud field to cross the closed sampling domain. The irradiance prediction values ​​of the corresponding grid points of the first, second, and third light field perturbation sentinel elements are arithmetically averaged to obtain the spatial mean of the irradiance prediction values ​​within the closed sampling domain. Based on the deviation between the spatial mean and the irradiance forecast values ​​of each sentinel element at corresponding grid points, a statistic characterizing the degree of dispersion of the irradiance forecast values ​​of each grid point relative to the spatial mean is calculated. The ratio of the statistic of dispersion to the spatial mean is used as the spatial variation coefficient of irradiance, which is used to characterize the degree of non-uniformity of the spatial distribution of irradiance within a closed sampling domain caused by cloud field movement.

4. The photovoltaic power prediction method based on a generative large model according to claim 3, characterized in that, Connecting the translated third optical field perturbation sentinel to either the first or second optical field perturbation sentinel to form a diagonal sampling path includes: Obtain the grid position of the third light field perturbation sentinel probe in the numerical weather prediction grid data after translation; Calculate the first distance between the grid point position of the third optical field perturbation sentinel after translation and the grid point position of the first optical field perturbation sentinel, and the second distance between the grid point position of the third optical field perturbation sentinel and the grid point position of the second optical field perturbation sentinel. The light field perturbation sentinel corresponding to the smaller of the first and second distances is selected as the connection object; the translated third light field perturbation sentinel is connected to the corresponding grid point in the numerical weather prediction grid data along a straight line, and the sequence of all grid points along the straight line is used as the diagonal sampling path.

5. The photovoltaic power prediction method based on a generative large model according to claim 1, characterized in that, Step 4 includes: Using enhanced physical prior cue words as a prefix context, meteorological forecast data for the target power plant's future time period, divided according to a preset time resolution, are concatenated segment by segment into the prefix context to form a structured input sequence; wherein, the meteorological forecast data includes irradiance, temperature, wind speed, cloud cover, and wind direction forecast elements for each time step; The structured input sequence is input into the photovoltaic generative big model. The photovoltaic generative big model performs autoregressive inference on the meteorological forecast data at each time step based on the static parameters embedded in the prefix context and the prior information quantified by cloud field perturbation, and outputs the corresponding power prediction value for each time step. Arrange the power prediction values ​​output at each time step in chronological order to obtain a preliminary predicted power sequence.

6. The photovoltaic power prediction method based on a generative large model according to claim 5, characterized in that, Step 5 includes: The first verification involves calculating the solar elevation angle for each time step in the preliminary power prediction sequence, setting the power prediction value corresponding to the time step with a solar elevation angle less than or equal to zero to zero, thus obtaining the first correction sequence. The second verification involves truncating the power values ​​at each time step in the first correction sequence. Power values ​​greater than the upper limit of the installed capacity threshold are truncated to the upper limit of the installed capacity threshold, and power values ​​less than zero are truncated to zero, thus obtaining the second correction sequence. The third verification involves scanning the second corrected sequence through time windows and calculating the power change slope between adjacent time steps. When the power change slope between adjacent time steps within a certain time window exceeds a preset mutation threshold, the power values ​​of each time step within the corresponding time window are smoothed by moving average to obtain the final predicted power sequence.

7. A photovoltaic power prediction system based on a generative large model, wherein the system implements the method as described in any one of claims 1 to 6, characterized in that, include: The model training module is used to pre-train a general generative large model on a massive amount of historical operation data of photovoltaic power plants in multiple tasks and domains to obtain a photovoltaic generative large model. The data receiving module is used to receive static parameters uploaded by the target power station, including the power station's geographical coordinates, installed capacity, component type, and support tilt and azimuth parameters. The prompt word construction module is used to extract quantitative prior information on cloud field disturbances around the target power station based on static parameters and gridded data of future time-period numerical weather forecasts obtained from the meteorological service interface, and to construct enhanced physical prior prompt words. The model prediction module is used to stitch together the enhanced physical prior cue words with the meteorological forecast data of the target power plant in the future period according to the preset time resolution, and then input the data into the photovoltaic generative large model to perform zero-sample power prediction inference to obtain the preliminary predicted power sequence of the target power plant in the future period. The processing and verification module is used to perform triple physical post-processing verification on the preliminary predicted power sequence, including nighttime zero-value forced constraint correction based on solar altitude angle calculation, output upper and lower limit truncation processing based on installed capacity threshold, and moving average smoothing processing for sections where the power change slope exceeds the preset abrupt change threshold within a continuous time window, to obtain the final predicted power sequence.

8. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images