Method, device, equipment and medium for physically bootstrapped photovoltaic power prediction

Abstract

The present application relates to the field of power prediction, in particular to a photovoltaic power prediction method and device based on physical guidance generation, equipment and medium. The present application extracts visual features from time series sky images and matches historical entries aligned across modalities, which can make full use of intuitive information such as cloud changes and light trends, effectively making up for the limitations of single time series data; then based on the historical power curve, combined with the current working condition, physical correction is carried out, so that the power reference is more in line with the actual operating state, and the model deviation is reduced. On this basis, by predicting the fluctuation residual and superimposing it on the power curve, the short-term power fluctuation characteristics can be accurately captured, and the fitting ability under sudden weather can be improved. The overall scheme combines data-driven and physical constraints, which not only enhances the prediction robustness under extreme and variable weather, but also improves the problems of insufficient fluctuation fitting and large deviation, and improves the photovoltaic power prediction accuracy and reliability.

Description

Technical Field

[0001] This invention relates to the field of power prediction, specifically to a method, apparatus, equipment, and medium for predicting photovoltaic power generated by physical guidance. Background Technology

[0002] With the rapid development of renewable energy and the large-scale grid connection of photovoltaic power plants globally, accurate prediction of photovoltaic power is crucial for the safe and stable operation of the power grid. Current photovoltaic power prediction methods largely rely on single meteorological data or historical power time-series data, failing to adequately utilize multimodal information such as sky images and weather texts, and exhibiting insufficient cross-modal feature fusion, making it difficult to comprehensively characterize key influencing factors such as cloud movement and changes in sunlight. Furthermore, in complex scenarios such as sudden cloud cover changes and extreme weather, existing models exhibit poor robustness, easily leading to problems such as insufficient fitting of power fluctuations and large prediction biases, thus failing to meet the practical needs of high-proportion grid-connected photovoltaic systems for high-precision and high-reliability power prediction. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide a method, apparatus, device and medium for predicting photovoltaic power through physical guidance, in order to solve the problems of poor robustness of existing models, and the tendency to underfit power fluctuations and large prediction deviations.

[0004] In a first aspect, embodiments of the present invention provide a physically guided photovoltaic power prediction method, the method comprising: Obtain a sequence of sky images within the current time period, wherein the sequence of sky images includes multiple consecutive frames of sky images arranged in chronological order; Identify consecutive multi-frame sky images to obtain target visual features, and use the target visual features to query at least one matching target historical entry from the database. The database includes multiple historical entries, each historical entry including a text vector and a historical power curve corresponding to the text vector. The text vector is obtained by aligning the visual features and semantic feature vectors of the historical sky images. The historical power value is determined based on the historical power curve, and the actual power value is calculated using the historical power value and the current operating parameters. Based on the actual power value, predict the fluctuation residual sequence for the future time period, and construct a prediction curve based on the fluctuation residual sequence and the power curve corresponding to the actual power value.

[0005] Furthermore, the method also includes: Obtain historical sky images; Obtain the structured weather description text corresponding to the historical sky images; Visual features are extracted from the historical sky images, and semantic feature vectors are obtained by extracting the weather description text. The text vector is obtained by performing cross-modal alignment based on the visual features and the semantic feature vector; Based on each of the text vectors, the corresponding weather description text, and the historical power curve of the photovoltaic power station, a corresponding historical entry is constructed and stored in the database based on multiple historical entries.

[0006] Furthermore, the step of performing cross-modal alignment based on the visual features and the semantic feature vector to obtain the text vector includes: The visual features and semantic feature vectors are input into the dual-tower model. The cosine similarity between the visual features and semantic feature vectors that match within a batch is calculated through the dual-tower model. The cosine similarity between the visual features and semantic feature vectors that match but do not match within a batch is also calculated to obtain the loss function value. The model parameters of the dual-tower model are optimized by backpropagation based on the loss function value, so that the semantic feature vector output by the dual-tower model is adjusted to match the visual features. The training is iterated until the dual-tower model converges, and the semantic feature vector output by the dual-tower model is used as a text vector aligned with the visual features.

[0007] Furthermore, the calculation of the actual power value using the historical power value and the current operating parameters includes: Obtain the current irradiance and the average irradiance for the same period in history, and calculate the ratio between the current irradiance and the average irradiance; The temperature coefficient of the photovoltaic power station is obtained, wherein the temperature coefficient is determined based on the current temperature and historical temperature of the photovoltaic power station; The actual power value is calculated based on the historical power value, the ratio, and the temperature coefficient.

[0008] Furthermore, the prediction of the fluctuation residual sequence over a future time period based on the actual power value includes: The future time period is divided into different levels, the time interval and residual prediction weight corresponding to each level are determined, and the actual power value is split according to the time level to form a corresponding time series input sequence. Input the time series input sequences corresponding to each level into the prediction model, and generate the initial residual sequences for each level respectively; The initial residual sequence is fused based on the residual prediction weights of each level to obtain the fluctuating residual sequence for the future time period.

[0009] Furthermore, the method also includes: Extract weather quantitative features from the weather description text included in the target historical entries; The parameter adjustment range is determined based on the aforementioned weather quantitative characteristics; The basic model parameters of the prediction model are adjusted according to the parameter adjustment range to obtain the adjusted prediction model.

[0010] Furthermore, the method also includes: Key weather factors are extracted from the weather description text included in the target historical entries, and power change data is obtained based on the prediction curve; Based on the key weather factors and the power change data, the percentage of each weather factor's impact on power is calculated.

[0011] Secondly, embodiments of the present invention provide a physically guided photovoltaic power prediction device, the device comprising: The acquisition module is used to acquire a sequence of sky images within the current time period, wherein the sequence of sky images includes multiple consecutive frames of sky images arranged in chronological order; The recognition module is used to recognize consecutive multi-frame sky images, obtain target visual features, and use the target visual features to query at least one matching target historical entry from the database. The database includes multiple historical entries, each historical entry includes a text vector and a historical power curve corresponding to the text vector. The text vector is obtained by aligning the visual features and semantic feature vectors of the historical sky images. The calculation module is used to determine the historical power value based on the historical power curve, and to calculate the actual power value using the historical power value and the current operating parameters; The processing module is used to predict the fluctuation residual sequence in the future time period based on the actual power value, and to construct a prediction curve based on the fluctuation residual sequence and the power curve corresponding to the actual power value.

[0012] Thirdly, embodiments of the present invention provide a computer device, including: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the method described in the first aspect or any corresponding embodiment thereof.

[0013] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer instructions that cause a computer to perform the method described in the first aspect or any of its corresponding embodiments.

[0014] This invention extracts visual features from time-series sky images and matches them with cross-modal aligned historical entries. This fully utilizes intuitive information such as cloud changes and illumination trends, effectively compensating for the limitations of single time-series data. Based on historical power curves, physical corrections are performed using current operating conditions, making the power benchmark more closely match actual operating conditions and reducing model bias. Furthermore, by predicting fluctuation residuals and superimposing them onto the power curve, short-term power fluctuation characteristics can be accurately captured, improving the fitting ability under abrupt weather changes. The overall solution combines data-driven approaches with physical constraints, enhancing the prediction robustness under extreme and variable weather conditions while addressing issues of insufficient fluctuation fitting and large biases, thus improving the accuracy and reliability of photovoltaic power prediction. Attached Figure Description

[0015] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is a schematic flowchart of a photovoltaic power prediction method based on physical guidance according to some embodiments of the present invention; Figure 2 This is a flowchart illustrating another physically guided photovoltaic power prediction method according to some embodiments of the present invention; Figure 3 This is a structural block diagram of a photovoltaic power prediction device generated by physical guidance according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] According to embodiments of the present invention, a method, apparatus, device, and medium for predicting photovoltaic power through physical guidance are provided. It should be noted that the steps shown in the flowcharts in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowcharts, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0019] This embodiment provides a physically guided photovoltaic power prediction method. Figure 1 This is a flowchart of a physically guided photovoltaic power prediction method according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps: Step S101: Obtain the sky image sequence within the current time period, wherein the sky image sequence includes multiple consecutive sky images arranged in chronological order.

[0020] In this embodiment, firstly, the time range for image acquisition needs to be determined (consistent with T during the model training phase, such as 1 hour), and the sampling interval Δt (such as 10 seconds) needs to be determined. This time range must be strictly aligned with the sampling times of sensors such as the power plant's irradiance and component temperature to ensure no deviation in the time dimension. During actual acquisition, automated shooting is performed using an all-sky imager (TSI) deployed above the power plant. The imager needs to be pre-calibrated (including white balance, exposure parameters, and geometric distortion correction) to avoid image feature distortion due to equipment parameter deviations. During acquisition, the system triggers shooting according to a preset Δt, automatically adding a precise timestamp to each frame and storing it as a continuous sequence in chronological order.

[0021] At the same time, a real-time data verification mechanism needs to be added: First, check the clarity of each frame (by calculating the image entropy value or edge gradient, filtering out blurry, overexposed / underexposed invalid frames); second, verify the integrity of the sequence. If no image is captured at a certain time point (such as a brief equipment failure), the temporal structure needs to be preserved by interpolation or marking missing bits to ensure that the sequence length conforms to the theoretical value of T / Δt (such as 360 frames corresponding to 1 hour and 10-second intervals).

[0022] Finally, the verified consecutive image frames are arranged in ascending order by timestamp to form a structured sky image sequence, which is stored in a tensor format that can be directly input into the visual model (e.g., [360,224,224,3], corresponding to the number of frames, height, width, and number of channels).

[0023] Step S102: Identify consecutive multi-frame sky images, obtain target visual features, and use the target visual features to query at least one matching target historical entry from the database. The database includes multiple historical entries, each of which includes a text vector and a historical power curve corresponding to the text vector. The text vector is obtained by aligning the visual features and semantic feature vectors of the historical sky images.

[0024] In this embodiment, the sequence of sky images is input into a pre-trained dual-tower model visual encoder (such as 3DCNN or VisionTransformer). The encoder first extracts spatial features (such as texture, cloud shape, and brightness distribution) from a single frame image, and then captures dynamic change features between frames (such as cloud movement speed and morphological evolution) through a temporal convolution / attention mechanism. Finally, it outputs a target visual feature vector V_now with uniform dimensions (such as 512 dimensions).

[0025] Next, the text vector T of all historical entries in the database is retrieved (this vector is the result of aligning visual features and text semantic features through InfoNCE loss in the historical stage). The similarity value between V_now and each T is calculated one by one using the cosine similarity algorithm. The closer the cosine similarity is to 1, the higher the matching degree between the real-time visual features and the weather pattern corresponding to the historical text vector.

[0026] Finally, in the target item retrieval stage: a similarity threshold (e.g., 0.8) is set and items are sorted from high to low similarity to select the Top-K (e.g., Top-3) historical items that meet the criteria. Each item contains information such as a text vector, the corresponding historical power curve, and the original weather description text. At the same time, to ensure the reliability of the retrieval, anomaly filtering logic needs to be added. If the similarity of the Top-K items is all below the threshold, a fallback mechanism is triggered (e.g., selecting the item with the second highest similarity and marking it as "low matching degree"). Finally, at least one target historical item is output.

[0027] Step S103: Determine the historical power value based on the historical power curve, and calculate the actual power value using the historical power value and the current operating parameters.

[0028] In this embodiment, the dimensional characteristics of the historical power curve are determined. This curve uses time and core operating parameters (such as speed, load rate, and ambient temperature) as the horizontal axis and power values ​​as the vertical axis. First, the operating dimension of the curve needs to be calibrated to establish a one-to-one correspondence index between "operating parameter combinations and power values." Next, based on the target operating range (such as a specific speed range or load range), data matching is performed on the preprocessed curve: if it is a discrete operating point, the power value corresponding to that operating parameter in the curve is directly found; if it is a continuous operating range, the average power value within that range is calculated using integration or weighted average methods, or the peak and valley power values ​​within the range are extracted as characteristic historical power values.

[0029] In this embodiment of the application, the actual power value is calculated using historical power values ​​and current operating parameters, including: obtaining the current irradiance and the average irradiance of the same period in history, and calculating the ratio between the current irradiance and the average irradiance; obtaining the temperature coefficient of the photovoltaic power station, wherein the temperature coefficient is determined based on the current temperature and historical temperature of the photovoltaic power station; and calculating the actual power value based on the historical power value, the ratio, and the temperature coefficient.

[0030] As an example, the formula is as follows: P_corrected = P_retrieved * (I_now / I_hist) * f(T_now, T_hist) where P_corrected is the corrected power, P_retrieved is the retrieved historical power value, I_now / I_hist is the ratio of the current measured irradiance to the historical average irradiance for the same period, and f(T_now, T_hist) is a function of the current module temperature and the historical module temperature.

[0031] Step S104: Predict the fluctuation residual sequence for the future time period based on the actual power value, and construct the prediction curve based on the fluctuation residual sequence and the power curve corresponding to the actual power value.

[0032] In this application embodiment, predicting the fluctuation residual sequence over a future time period based on actual power values ​​includes: Step A1 involves dividing the future time period into different levels, determining the time interval and residual prediction weight corresponding to each level, and splitting the actual power value according to the time level to form the corresponding time series input sequence.

[0033] The time period to be predicted is broken down into a multi-level structure, such as three core levels: hourly, daily, and weekly. The hourly time interval is set as a unit of 1 hour, the daily interval is a natural day, and the weekly interval is a 7-day interval. Then, the residual prediction weight is determined according to the degree of influence of different levels on the power residual. For example, the hourly level directly reflects short-term power fluctuations, so the weight is set to 0.5, the daily level to 0.3, and the weekly level to 0.2.

[0034] Subsequently, the actual historical power data of the object to be predicted is collected and split according to the divided time levels. For example, continuous power data is sliced ​​by hour to form an hourly time series input sequence, daily power average / extreme values ​​are integrated into a daily sequence, and weekly power change trends are summarized into a weekly sequence. During the splitting process, it is necessary to ensure that the time axis of each level sequence is aligned, and missing data is filled in by interpolation. Finally, a time series input sequence with unified dimensions that corresponds one-to-one with each time level is formed.

[0035] In this embodiment, the residual prediction weights for each level can be determined as follows: Weather quantification features such as cloud cover, irradiance, and temperature are extracted from weather description text and sky images to obtain standardized weather feature vectors. These standardized weather feature vectors are then input into a lightweight adaptive network to obtain the original weight outputs for each time level. The original weight outputs are then normalized to obtain the optimal residual prediction weights for each time level that sum to 1.

[0036] Specifically, the process involves retrieving weather description text and real-time sky images from historical entries of the target data. The text descriptions of cloud cover, irradiance, and temperature are extracted, and corresponding visual features are extracted from the sky images using image recognition algorithms. Secondly, standardized quantization rules are established to convert the text descriptions and visual features into numerical values. For example, cloud cover is quantized from 0 to 10, irradiance is converted to actual measured values, and temperature is quantized based on deviations from standard operating conditions. Then, all quantized values ​​are normalized to unify the dimensions to the 0-1 range, eliminating the impact of numerical differences. Finally, the normalized weather factor values ​​are integrated to form a standardized weather feature vector with fixed dimensions and a unified format, providing standardized input for subsequent weight learning.

[0037] A lightweight adaptive network was constructed, employing a two-layer fully connected structure. The input layer dimension matches the dimension of the standardized weather feature vector, while the output layer dimension matches the number of time levels. A ReLU activation function was introduced to enhance non-linear learning capabilities and prevent overfitting due to excessive model complexity. Next, the standardized weather feature vector obtained in step 1 was input into the network's input layer and transformed into an intermediate feature vector through feature mapping in the first fully connected layer. This intermediate feature vector was then processed by the ReLU activation function and input into the second fully connected layer, where weight matrix operations were used to learn the association between features and the weights of each time level. The network's output layer outputs the original weight values ​​corresponding to each time level. These values ​​initially reflect the importance of the residuals at each level under different weather features, laying the foundation for subsequent optimization.

[0038] Obtain the original weight outputs for each time level, sum all original weights, and determine if the original weights meet the requirement that the sum is 1. If not, initiate normalization processing. Use the L2 normalization algorithm to divide the original weight of each time level by the sum of all original weights to calculate the normalized weight for each level. Then, verify the normalized weights to confirm that all weight values ​​are within the range of 0-1 and the sum is strictly equal to 1, avoiding unreasonable weight allocation. Finally, use the verified normalized weights as the optimal residual prediction weights for each time level.

[0039] Step A2: Input the time series input sequences corresponding to each level into the prediction model, and generate the initial residual sequences for each level.

[0040] The sequences are standardized to map power values ​​to the [0,1] interval, eliminating dimensional differences. Then, hourly, daily, and weekly time-series input sequences are fed into prediction models with the same architecture (such as LSTM, Transformer, etc.). The models are trained and predicted independently for each level of sequence: for example, the hourly sequence learns the pattern of power variation with hours and outputs hourly power predictions; similarly, the daily and weekly models output power predictions for their respective dimensions. Afterwards, the difference between the predicted values ​​and historical actual power values ​​at each level is calculated to obtain the residual value for each time unit. These residual values ​​are arranged chronologically to form the initial residual sequences for the hourly, daily, and weekly levels.

[0041] Meanwhile, this step also requires embedding a parameter tuning process for weather features: first, extract quantitative features such as wind force, temperature, and weather type from the weather description text of the target historical entry (such as "sunny turning cloudy, wind force level 3, temperature 25℃") through text segmentation and feature mapping; then, based on the preset feature-parameter mapping rules, determine the adjustment range of the model's basic parameters (such as learning rate and number of hidden layer neurons).

[0042] In this embodiment of the application, the method further includes: extracting weather quantitative features from the weather description text included in the target historical entry; determining the parameter adjustment range based on the weather quantitative features; adjusting the basic model parameters of the prediction model according to the parameter adjustment range to obtain the adjusted prediction model.

[0043] Understandably, the process involves several steps. First, the structured weather description text in the target historical entries is segmented and semantically analyzed to extract core weather elements (such as cloud cover, cloud movement speed, irradiance variation rate, and visibility). Second, quantification rules are set for each weather element; for example, "heavy cloud cover" is mapped to a numerical range of 8-10, and "rapid cloud movement" is quantified as a speed value ≥5 m / s, transforming the text description into a calculable value. Then, the quantified values ​​are standardized to unify the units and value ranges (e.g., normalized to the 0-1 range). Finally, all quantified weather elements are integrated to form a fixed-dimensional weather quantification feature vector, ensuring that the features can be directly used for subsequent parameter adjustment calculations.

[0044] Analyze the optimal values ​​of model parameters corresponding to different weather characteristics in historical data to determine the influence weight of each feature dimension; input the extracted weather quantification feature vector into the correlation model to calculate the basic adjustment value corresponding to each feature dimension; then, combine the current prediction error of the model, historical adjustment effects and other feedback information to correct the basic adjustment value to avoid over-adjustment; finally, obtain the final adjustment range of each model parameter by weighted summation, clarify the specific value / proportion of the parameter to be increased or decreased, and ensure that the adjustment range fits the prediction needs corresponding to the current weather characteristics.

[0045] The basic parameter system of the prediction model is reviewed, and the types of adjustable parameters (such as the learning rate of LSTM, the number of convolutional kernels, attention weights, etc.) and their current values ​​are identified. The corresponding parameters are modified according to the calculated adjustment range, such as increasing or decreasing the learning rate proportionally and adjusting the number of convolutional kernels numerically. Then, the effectiveness of the adjusted parameters is verified to ensure that the parameter values ​​are within a reasonable range for stable model operation. The adjusted parameters are then loaded into the prediction model to complete the model parameter update, resulting in an adjusted prediction model adapted to the current weather characteristics, ensuring that the model can optimize prediction accuracy based on weather changes.

[0046] As an example, weather quantification features are extracted from the weather description text of the target historical entry, "rapidly moving fractal cumulus clouds, drastic irradiance fluctuations, and high component temperatures." The cloud movement speed is recorded as 4.5 m / s, the irradiance fluctuation rate as 25%, and the temperature deviation as 8℃. Next, based on the preset mapping relationship between weather features and parameters, the adjustment range is determined to be an increase of 0.0005 in the learning rate, an increase of 8 neurons in the LSTM hidden layer, and a decrease of 0.05 in the dropout coefficient. Then, the basic parameters of the original prediction model are modified according to this range. Finally, the parameters are updated and the validity is verified to obtain the adjusted prediction model adapted to the current drastically fluctuating weather.

[0047] Step A3: Based on the residual prediction weights of each level, the initial residual sequence is fused to obtain the fluctuation residual sequence for the future time period.

[0048] Retrieve the prediction weights for each level of residuals (e.g., 0.5 for hourly, 0.3 for daily, and 0.2 for weekly), and confirm the alignment of the initial residual sequences at each level in the time dimension. If the hourly sequence is based on 1-hour units, the daily sequence on 24-hour units, and the weekly sequence on 168-hour units, the daily and weekly residual sequences need to be decomposed into hourly granularity first, or the time granularity should be unified through a mean distribution or other methods. Subsequently, for each time unit residual value in the initial residual sequence of each level, multiply it by the corresponding level's weight coefficient. For example, if the hourly initial residual value is 2, the daily value is 1, and the weekly value is 0.5, then the weighted calculation is 2×0.5+1×0.3+0.5×0.2=1.4.

[0049] After completing the weighted calculations for all time units, these weighted residual values ​​are arranged sequentially according to the future time periods to form a continuous fluctuating residual sequence. During the fusion process, an outlier correction step is also required. If a residual value at a certain level deviates from the normal range (e.g., exceeding 3 times the standard deviation), the weight of that level is temporarily adjusted and recalculated to ensure that the fused fluctuating residual sequence accurately reflects the residual fluctuation patterns of power at different future time scales.

[0050] In the embodiments of this application, such as Figure 2 As shown, the method also includes: Step S201: Obtain historical sky images and obtain the structured weather description text corresponding to the historical sky images.

[0051] First, image resources from multiple channels are integrated, including continuous images taken at fixed time intervals (e.g., every 10 minutes) by fixed-point sky cameras deployed at weather stations, regional sky remote sensing images transmitted by meteorological satellites, and labeled sky images in publicly available datasets. Second, the acquired raw images undergo standardized preprocessing: image resolution is standardized (e.g., adjusted to 224×224 pixels), color space is converted to RGB format, and invalid images are removed due to blurriness, occlusion (e.g., the camera being covered by foreign objects), or abnormal shooting angles. Simultaneously, metadata is bound to each valid image, including shooting time (accurate to the minute), geographical location (latitude and longitude), and shooting equipment parameters, ensuring the image has temporal and spatial traceability. Finally, the processed images are stored in an image database (e.g., MongoDB GridFS) categorized by time series and geographical region, and indexed for rapid subsequent retrieval. The entire process must ensure the integrity and consistency of the image data, covering different weather scenarios (sunny, cloudy, rainy, snowy, foggy, etc.).

[0052] First, official meteorological observation records corresponding to the image capture time and geographical location are retrieved. These records contain structured information reviewed by meteorologists, such as weather phenomena (sunny, cloudy, light rain, fog), cloud cover (0-10%), visibility (km), and cloud base height (m). If official records are lacking, unstructured texts such as contemporary meteorological broadcasts and public weather logs are used to supplement them. Second, the acquired text is structured: a unified text template is developed to organize scattered weather information into a fixed format of "time-location-weather phenomenon-cloud cover-visibility," for example, "October 1, 2025, 14:00, a certain district in a certain city, sunny, 10% cloud cover, visibility 15km." Simultaneously, Natural Language Processing (NLP) technology is used to clean the text, removing redundant expressions and correcting typos to ensure the accuracy and standardization of the text description. Finally, a one-to-one mapping relationship between images and text is established, using a unique identifier (such as an image ID) to associate the structured weather description text with the corresponding image and store it in a text database.

[0053] Step S202: Extract visual features from historical sky images and extract weather description text to obtain semantic feature vectors.

[0054] The preprocessed historical sky images are input into the model, and low-level features (such as edges, textures, and color distribution) and high-level features (such as cloud morphology, distribution range, and sky brightness features) are extracted layer by layer through convolutional layers. Then, the fully connected classification layer of the model is removed, and the output tensor of the penultimate layer is retained as the initial visual features. Finally, the initial features are normalized (such as L2 normalization) and converted into fixed-dimensional (such as 2048-dimensional) visual feature vectors, which can accurately represent the core visual attributes of the image. In terms of semantic feature extraction, for structured weather description text, a pre-trained language model (such as BERT, Word2Vec) is used for processing: first, the text is segmented and converted into word embedding vectors that the model can recognize. After being input into the model, the self-attention mechanism is used to capture the relationship between semantic units in the text (such as the relationship between "light rain" and "low visibility"). Then, the output vector at the [CLS] position of the model is extracted as the initial semantic features, and then it is converted into a fixed-dimensional (such as 2048-dimensional) semantic feature vector that is adapted to the visual feature dimension.

[0055] Step S203: Perform cross-modal alignment based on visual features and semantic feature vectors to obtain text vectors.

[0056] In this embodiment of the application, cross-modal alignment is performed based on visual features and semantic feature vectors to obtain text vectors, including: Step B1: Input the visual features and semantic feature vectors into the dual-tower model. Calculate the cosine similarity between the matching visual features and semantic feature vectors within a batch, and calculate the cosine similarity between the matching and non-matching visual features and semantic feature vectors within a batch, to obtain the loss function value.

[0057] A dual-tower model consisting of a visual tower and a text tower is constructed. Preprocessed fixed-dimensional visual feature vectors are input into the visual tower, and semantic feature vectors are input into the text tower. Feature mapping is performed using a symmetric multilayer perceptron to ensure consistent output dimensions. Next, training batches are constructed, each containing several matched visual-semantic feature pairs and randomly combined mismatched pairs, forming positive and negative sample sets. Then, all features within a batch are input into the corresponding tower structure to obtain projected feature vectors, and the cosine similarity of matched and mismatched pairs is calculated for each pair. Finally, the InfoNCE loss function, combined with a temperature coefficient, is used to calculate the loss function value for that batch, quantifying the cross-modal feature alignment effect of the current model.

[0058] Step B2 involves optimizing the model parameters of the dual-tower model through backpropagation based on the loss function value, thereby adjusting the semantic feature vector output by the dual-tower model towards the direction of the matching visual features.

[0059] Using the obtained loss function value as the optimization objective, the gradient of the loss value with respect to all trainable parameters of the dual-tower model is derived in reverse, and its impact on the parameters of the text tower and visual tower is calculated separately. Then, the optimizer updates the parameters according to the direction and magnitude of the gradient, focusing on adjusting the semantic feature mapping weights of the text tower to make the semantic projection vectors converge towards the matching visual projection vector space, while simultaneously fine-tuning the visual tower parameters to help improve the matching degree. Finally, after each batch of parameter updates, it is verified whether the loss value decreases. If it does not decrease, the batch size or learning rate is adjusted to ensure that the model optimizes in the direction of semantic and visual feature alignment.

[0060] Step B3: Iterate the training until the dual-tower model converges, and use the semantic feature vector output by the dual-tower model as the text vector aligned with the visual features.

[0061] A convergence criterion is established, determining the total number of training epochs. Simultaneously, the loss value and matching accuracy on the validation set are monitored. Model convergence is determined when the validation set loss value no longer decreases and the accuracy stabilizes above a preset threshold for multiple consecutive epochs. Next, training set data is input in batches, and similarity calculation, loss calculation, backpropagation, and parameter updates are repeated. Model performance is evaluated using the validation set after each epoch. Then, if the model fails to converge, iteration continues. If overfitting occurs, an early stopping mechanism or a Dropout layer is added to suppress overfitting. Finally, after model convergence, all parameters are fixed. The semantic feature vector output by the dual-tower model's text tower is used as a text vector aligned with visual features for subsequent cross-modal retrieval and matching tasks.

[0062] Step S204: Based on each text vector, the corresponding weather description text, and the historical power curve of the photovoltaic power station, construct the corresponding historical entries and store them in the database based on multiple historical entries.

[0063] In this embodiment, each text vector obtained through cross-modal alignment is used as the core identifier, and its corresponding original structured weather description text (such as "sunny, cloud cover 10%, visibility 15km") and the historical power curve of the photovoltaic power station during the same period (including time-series data such as timestamps and power values) are associated to ensure that the three correspond one-to-one and the time dimension is completely aligned, without any data misalignment issues.

[0064] Secondly, a unified historical data format was determined, encapsulating text vectors (fixed-dimensional feature values), weather description text (structured strings), and historical power curves (time-series arrays) into a standardized dictionary structure, such as {"text_vector":[0.12,0.35,...],"weather_desc":"2025-10-01 14:00 sunny with 10% cloud cover","power_curve":[[14:00,500kW],[14:10,510kW],...]}, ensuring data structure consistency. Then, a database adapted for storing time-series data and feature vectors was built. An index was created for the text vectors to improve subsequent retrieval efficiency, and the storage structure was optimized for the time-series characteristics of the historical power curves. A unique identifier ID was added to each historical entry. Finally, all encapsulated historical entries were batch-written into the database.

[0065] It should be noted that by employing a dual-tower structure to map visual and semantic feature vectors separately, effective comparison of the two modalities can be ensured within a unified space. Furthermore, by constructing positive and negative samples and using the InfoNCE loss function to calculate similarity and loss values, the model's ability to learn image-text matching relationships is significantly enhanced, reducing feature misalignment caused by cross-modal differences. Optimizing model parameters through backpropagation based on the loss function allows the semantic features output by the text tower to gradually converge towards the corresponding visual features, achieving accurate cross-modal alignment. Iterative training to convergence, coupled with strategies such as early stopping and Dropout, improves the stability and generalization ability of text vectors, avoiding overfitting. The resulting text vectors simultaneously carry semantic information and visual features, significantly improving the accuracy and reliability of subsequent historical entry retrieval, fundamentally addressing the issues of insufficient multimodal data fusion and inaccurate weather pattern matching.

[0066] In this embodiment of the application, the method further includes: extracting key weather factors based on the weather description text included in the target historical entry, and obtaining power change data based on the prediction curve; and calculating the proportion of the influence of each weather factor on the power based on the key weather factors and the power change data.

[0067] First, key weather factors such as cloud cover, cloud movement speed, irradiance, and component temperature are extracted from the weather description text of the target historical entries and standardized and quantified. Second, based on the corresponding forecast curves and historical power curves, power change data such as power fluctuation amplitude, rate of change, and peak-to-valley difference are extracted to ensure strict temporal alignment between weather factors and power data. Then, correlation analysis, weighted regression, or feature contribution algorithms are used to calculate the correlation strength between each weather factor and power change, eliminating interference factors and noise data. Finally, the contribution of each weather factor is normalized to obtain the proportion of each weather factor's influence on power change, and the sum of these proportions is 1, intuitively reflecting the importance of different weather factors.

[0068] As an example, three key factors were extracted from the weather description text: "rapid movement of fractulum clouds, large fluctuations in irradiance, and relatively high temperature," which were quantified as cloud movement speed of 4.2 m / s, irradiance fluctuation rate of 22%, and temperature deviation of 6℃. Secondly, the power fluctuation amplitude was found to be 18% and the rate of change was 1.2 kW / s from the power curve. Then, the contribution percentages were calculated: irradiance accounted for 0.55%, cloud movement for 0.30%, and temperature for 0.15%. Finally, after normalization, the percentages of each factor's influence on power were obtained: irradiance 55%, cloud movement 30%, and temperature 15%, providing a basis for model adjustment and predictive interpretation.

[0069] This embodiment also provides a physically guided photovoltaic power prediction device, which is used to implement the above embodiments and preferred embodiments, and will not be repeated as described previously. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0070] This embodiment provides a physically guided photovoltaic power prediction device, such as... Figure 3 As shown, it includes: The acquisition module 301 is used to acquire a sequence of sky images within the current time period, wherein the sequence of sky images includes multiple consecutive frames of sky images arranged in chronological order; The recognition module 302 is used to recognize multiple consecutive sky images, obtain target visual features, and use the target visual features to query at least one matching target historical entry from the database. The database includes multiple historical entries, each historical entry includes a text vector and a historical power curve corresponding to the text vector. The text vector is obtained by aligning the visual features and semantic feature vectors of the historical sky images. The calculation module 303 is used to determine the historical power value based on the historical power curve, and to calculate the actual power value using the historical power value and the current operating parameters; The processing module 304 is used to predict the fluctuation residual sequence in the future time period based on the actual power value, and to construct a prediction curve based on the fluctuation residual sequence and the power curve corresponding to the actual power value.

[0071] In this embodiment of the application, the device further includes: a construction module, used to acquire historical sky images; acquire structured weather description text corresponding to the historical sky images; extract visual features from the historical sky images and extract semantic feature vectors from the weather description text; perform cross-modal alignment based on the visual features and semantic feature vectors to obtain text vectors; construct corresponding historical entries based on each text vector, the corresponding weather description text, and the historical power curve of the photovoltaic power station, and store them in a database based on multiple historical entries.

[0072] In this embodiment, the construction module is used to input visual features and semantic feature vectors into the dual-tower model, calculate the cosine similarity between matching visual features and semantic feature vectors within a batch, and calculate the cosine similarity between matching and mismatched visual features and semantic feature vectors within a batch to obtain a loss function value; optimize the model parameters of the dual-tower model based on the loss function value through backpropagation, so that the semantic feature vector output by the dual-tower model is adjusted towards the direction of matching visual features; iterative training is performed until the dual-tower model converges, and the semantic feature vector output by the dual-tower model is used as a text vector aligned with the visual features.

[0073] In this embodiment of the application, the calculation module 303 is used to obtain the current irradiance and the average irradiance of the same period in history, and to calculate the ratio between the current irradiance and the average irradiance; to obtain the temperature coefficient of the photovoltaic power station, wherein the temperature coefficient is determined based on the current temperature and historical temperature of the photovoltaic power station; and to calculate the actual power value based on the historical power value, the ratio and the temperature coefficient.

[0074] In this embodiment, the processing module 304 is used to divide the future time period into different levels, determine the time interval and residual prediction weight corresponding to each level, and split the actual power value according to the time level to form the corresponding time series input sequence; input the time series input sequence corresponding to each level into the prediction model, and obtain the initial residual sequence of each level; and fuse the initial residual sequence based on the residual prediction weight of each level to obtain the fluctuation residual sequence in the future time period.

[0075] In this embodiment of the application, the apparatus further includes: an adjustment module, configured to extract weather quantitative features from the weather description text included in the target historical entry; determine the parameter adjustment range based on the weather quantitative features; and adjust the basic model parameters of the prediction model according to the parameter adjustment range to obtain the adjusted prediction model.

[0076] In this embodiment of the application, the device further includes: an analysis module, used to extract key weather factors from the weather description text included in the target historical entry, and to obtain power change data according to the prediction curve; and to calculate the proportion of the influence of each weather factor on the power based on the key weather factors and the power change data.

[0077] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 4 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system).

[0078] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0079] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0080] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device as shown by a landing page for an app. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, which can be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0081] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0082] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0083] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0084] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for predicting photovoltaic power generated by physical guidance, characterized in that, The method includes: Obtain a sequence of sky images within the current time period, wherein the sequence of sky images includes multiple consecutive frames of sky images arranged in chronological order; Identify consecutive multi-frame sky images to obtain target visual features, and use the target visual features to query at least one matching target historical entry from the database. The database includes multiple historical entries, each historical entry including a text vector and a historical power curve corresponding to the text vector. The text vector is obtained by aligning the visual features and semantic feature vectors of the historical sky images. The historical power value is determined based on the historical power curve, and the actual power value is calculated using the historical power value and the current operating parameters. Based on the actual power value, predict the fluctuation residual sequence for the future time period, and construct a prediction curve based on the fluctuation residual sequence and the power curve corresponding to the actual power value.

2. The method according to claim 1, characterized in that, The method further includes: Obtain historical sky images; Obtain the structured weather description text corresponding to the historical sky images; Visual features are extracted from the historical sky images, and semantic feature vectors are obtained by extracting the weather description text. The text vector is obtained by performing cross-modal alignment based on the visual features and the semantic feature vector; Based on each of the text vectors, the corresponding weather description text, and the historical power curve of the photovoltaic power station, a corresponding historical entry is constructed and stored in the database based on multiple historical entries.

3. The method according to claim 2, characterized in that, The process of performing cross-modal alignment based on the visual features and the semantic feature vectors to obtain the text vector includes: The visual features and semantic feature vectors are input into the dual-tower model. The cosine similarity between the visual features and semantic feature vectors that match within a batch is calculated through the dual-tower model. The cosine similarity between the visual features and semantic feature vectors that match but do not match within a batch is also calculated to obtain the loss function value. The model parameters of the dual-tower model are optimized by backpropagation based on the loss function value, so that the semantic feature vector output by the dual-tower model is adjusted to match the visual features. The training is iterated until the dual-tower model converges, and the semantic feature vector output by the dual-tower model is used as a text vector aligned with the visual features.

4. The method according to claim 1, characterized in that, The calculation of the actual power value using the historical power value and the current operating parameters includes: Obtain the current irradiance and the average irradiance for the same period in history, and calculate the ratio between the current irradiance and the average irradiance; The temperature coefficient of the photovoltaic power station is obtained, wherein the temperature coefficient is determined based on the current temperature and historical temperature of the photovoltaic power station; The actual power value is calculated based on the historical power value, the ratio, and the temperature coefficient.

5. The method according to claim 1, characterized in that, The prediction of the fluctuation residual sequence over a future time period based on the actual power value includes: The future time period is divided into different levels, the time interval and residual prediction weight corresponding to each level are determined, and the actual power value is split according to the time level to form a corresponding time series input sequence. Input the time series input sequences corresponding to each level into the prediction model, and generate the initial residual sequences for each level respectively; The initial residual sequence is fused based on the residual prediction weights of each level to obtain the fluctuating residual sequence for the future time period.

6. The method according to claim 5, characterized in that, The method further includes: Extract weather quantitative features from the weather description text included in the target historical entries; The parameter adjustment range is determined based on the aforementioned weather quantitative characteristics; The basic model parameters of the prediction model are adjusted according to the parameter adjustment range to obtain the adjusted prediction model.

7. The method according to claim 1, characterized in that, The method further includes: Key weather factors are extracted from the weather description text included in the target historical entries, and power change data is obtained based on the prediction curve; Based on the key weather factors and the power change data, the percentage of each weather factor's impact on power is calculated.

8. A photovoltaic power prediction device generated by physical guidance, characterized in that, The device includes: The acquisition module is used to acquire a sequence of sky images within the current time period, wherein the sequence of sky images includes multiple consecutive frames of sky images arranged in chronological order; The recognition module is used to recognize consecutive multi-frame sky images, obtain target visual features, and use the target visual features to query at least one matching target historical entry from the database. The database includes multiple historical entries, each historical entry includes a text vector and a historical power curve corresponding to the text vector. The text vector is obtained by aligning the visual features and semantic feature vectors of the historical sky images. The calculation module is used to determine the historical power value based on the historical power curve, and to calculate the actual power value using the historical power value and the current operating parameters; The processing module is used to predict the fluctuation residual sequence in the future time period based on the actual power value, and to construct a prediction curve based on the fluctuation residual sequence and the power curve corresponding to the actual power value.

9. A computer device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method of any one of claims 1 to 7.

Images