Active power distribution network multi-area photovoltaic power prediction method and system

Abstract

The invention discloses a multi-region photovoltaic power prediction method and system for an active power distribution network, and relates to the technical field of photovoltaic power generation prediction and artificial intelligence, and the method comprises the steps: carrying out the weather scene division of the historical numerical weather forecast data of a source region through K-means clustering, and carrying out the feature processing through Z-score standardization; then, a scene-divided Seq2Seq prediction model is constructed, a bidirectional GRU structure is adopted by a model encoder, and a GRU structure is adopted by a decoder; then, source domain model parameters are migrated to a target domain in a layered mode through a migration learning strategy, bottom-layer feature freezing, middle-layer fine adjustment and top-layer adaptation are included, and gradual unfreezing optimization is adopted; and finally, multi-region photovoltaic power accurate prediction is realized through dynamic scene identification and distance weighting integration. According to the method, the problem of low prediction precision of a data scarce region is effectively solved, and the generalization ability and the expandability of a prediction model are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of photovoltaic power generation prediction and artificial intelligence technology, specifically to a method and system for predicting photovoltaic power in multiple regions of an active power distribution network. Background Technology

[0002] With my country's goal of "peaking carbon and achieving carbon neutrality" and the continuous advancement of its energy transition strategy, new energy sources, primarily photovoltaic and wind power, are developing rapidly, with installed capacity increasing year by year. As distributed photovoltaic power generation is widely applied in active distribution networks, accurate prediction of photovoltaic power output has become a key issue in improving grid operating efficiency and reliability. However, due to the significant spatial heterogeneity of sunlight intensity and meteorological conditions in different geographical locations, traditional prediction methods often exhibit limitations in adaptability and accuracy when dealing with cross-regional, distributed photovoltaic systems, increasing the difficulty of distribution network regulation.

[0003] Due to differences in operation and maintenance, regional meteorological conditions, and other factors, the photovoltaic output characteristics of multiple renewable energy power plants often vary significantly. Using a uniform prediction model makes it difficult to accurately characterize the output differences between different plants, thus limiting prediction accuracy. While modeling each plant individually can improve prediction accuracy, it significantly increases computational and operational costs. Currently, output prediction modeling methods for distributed photovoltaic power generation remain relatively limited, with traditional methods generally facing challenges such as computational complexity, weak generalization ability, and insufficient training samples.

[0004] For photovoltaic (PV) power generation prediction, current mainstream research approaches both domestically and internationally can be summarized into two main categories: data-driven statistical learning methods and mechanism-based physical modeling methods. Data-driven statistical learning methods are based on historical data, analyzing the statistical relationship between historical power output of power plants and corresponding meteorological data to construct machine learning or deep learning models for prediction. Their advantage lies in their ability to automatically learn complex patterns from large amounts of data, but they are highly dependent on the quality and quantity of data. Mechanism-based physical modeling methods start from the physical principles of PV power generation, establishing mathematical models to describe the deterministic relationship between physical parameters such as solar radiation, ambient temperature, and module characteristics and the final output power.

[0005] The shortcomings of existing technologies are as follows: 1. Data-driven statistical learning methods are highly dependent on the quality and quantity of data. However, distributed photovoltaic (PV) generally suffers from a data shortage, which limits the conditions for high-level power prediction of distributed PV. 2. Although mechanism-based physical modeling methods have clear physical meaning, they may have limitations in characterizing actual operating losses such as equipment degradation and dust blockage. 3. Traditional methods have weak generalization ability and cannot meet the prediction accuracy requirements of different new energy power plants. Summary of the Invention

[0006] This invention aims to address at least one of the technical problems existing in the prior art. To this end, one objective of this invention is to propose a multi-region photovoltaic power prediction method for active power distribution networks. This method, by constructing a unified meteorological scenario framework and realizing scenario-based knowledge transfer, aims to solve the problems of low prediction accuracy and modeling difficulties in photovoltaic power prediction in data-scarce areas.

[0007] To address the above problems, this invention provides a method for predicting photovoltaic power in multiple regions of an active power distribution network, comprising the following steps: S1. Weather Scene Segmentation: Based on the K-means clustering algorithm, the historical numerical weather forecast data of the source region is segmented into weather prediction scenes. The historical numerical weather forecast data includes multiple meteorological factors, and feature engineering is performed, including standardizing the data using the Z-score standardization formula. S2. Source Domain Model Construction: In the source region, a dedicated sequence-to-sequence Seq2Seq prediction model is constructed for each weather scenario and trained using the Seq2Seq model architecture. The Seq2Seq model includes an encoder and a decoder. The encoder adopts a bidirectional gated recurrent unit (BiGRU), and the decoder adopts a GRU. S3. Target Domain Transfer: For the target region, local data is used to complete scene matching. Based on the scene division criteria of the source region, a transfer learning strategy is used to transfer the pre-trained model parameters of the source domain to the target domain. The transfer learning strategy includes hierarchical transfer: bottom-level transfer, middle-level transfer, and top-level fine-tuning, and uses a progressive unfreezing strategy for fine-tuning. The bottom-level transfer freezes the first layer BiGRU parameters of the encoder, the middle-level transfer fine-tunes the last layer of the encoder and the first layer of the decoder, and the top-level fine-tuning reinitializes the second layer of the encoder and the output layer. S4. Prediction Execution: In the multi-region prediction system, the weather scene to which the input meteorological data belongs is dynamically identified. The scene is classified by calculating the Euclidean distance from the center of the scene cluster, and the corresponding fine-tuned Seq2Seq model is called to predict photovoltaic power. Finally, a distance-weighted multi-model integration prediction mechanism is introduced to handle the ambiguity of scene boundaries, and the model is updated regularly with new data to optimize prediction performance.

[0008] Preferably, the Z-score standardization formula in S1 is: in The original data values, The mean of all sample data for this feature. The standard deviation of all sample data for this feature is . This represents the new value after Z-score standardization; The optimal number of clusters K is determined by the silhouette coefficient and the elbow rule, and K-means clustering is performed to divide the weather scene into K types.

[0009] Preferably, the input feature sequence constructed in S2 includes a meteorological feature sequence and a time feature sequence; The meteorological characteristic sequence is as follows: in, Represents a meteorological characteristic sequence. express Global horizontal irradiance at any given time express Direct irradiance at any given time express Temperature at any moment express Real-time numerical weather forecast - humidity express Wind speed at any moment express Real-time numerical weather forecast - wind direction express Atmospheric pressure at any given moment; The time feature sequence is: in Represents a time-feature sequence. for Hour code of time for Daily cycle encoding of time, for Seasonal coding of time.

[0010] Preferably, the specific methods for determining the optimal number of clusters K in S1 include: The preset range of K values ​​is from 2 to 8. For each K value, the K-means algorithm is run to calculate the profile coefficient and the inertia value of the elbow rule. The optimal K value is determined by combining the maximum point of the profile coefficient and the inflection point of the elbow rule.

[0011] Preferably, when performing K-means clustering in S1, multiple initializations are used to avoid local optima, and the cluster center coordinates are output to form a weather scene discrimination rule base.

[0012] Preferably, the training process of the Seq2Seq model in S2 includes: The encoder adopts a two-layer BiGRU structure. The first layer BiGRU takes meteorological feature sequences and historical power sequences as inputs, and the second layer BiGRU outputs the hidden state. The decoder takes numerical weather forecast data for future periods as inputs, and after fusing the encoder state, outputs the prediction result through GRU units and a fully connected layer.

[0013] Preferably, the specific process of the gradual thawing strategy in S3 includes: The first 10 training rounds only train the last layer of the decoder. In rounds 11-20, all layers of the decoder are unfrozen. In rounds 21-30, the last layer of the encoder is unfrozen. After round 31, all layers are fine-tuned, and an early stopping strategy and L2 regularization are used to prevent overfitting.

[0014] Preferably, in S4, the prediction performance evaluation uses root mean square error and mean absolute error as indices, and evaluates the model's generalization ability by scenario.

[0015] A system used in a multi-region photovoltaic power prediction method for active power distribution networks includes: The weather scene segmentation module is configured to perform the weather scene segmentation function of S1, including data standardization, K value determination, and K-means clustering; The source domain model building module is configured to perform the source domain model building function of S2, including feature sequence construction and Seq2Seq model training; The transfer learning module is configured to perform S3's target domain transfer functions, including scene matching, hierarchical transfer, and fine-tuning. The prediction execution module is configured to perform S4's prediction functions, including dynamic scene recognition, model invocation, and ensemble prediction. The system optimization module is configured to periodically update the model and evaluate performance to achieve continuous learning. The weather scene segmentation module further includes a feature engineering submodule, which is used to process numerical weather forecast data and apply the Z-score normalization formula; The source domain model building module further includes a sequence feature building submodule, which is used to generate meteorological feature sequences and time feature sequences as model inputs; The prediction execution module further integrates a distance-weighted multi-model mechanism to handle scenario uncertainties.

[0016] The advantages of this invention compared to the prior art are: 1. This invention effectively alleviates the data dependency problem. By using a transfer learning framework, the learned general "weather-power" mapping knowledge is transferred to new distributed photovoltaic sites, significantly reducing the dependence on the quantity and quality of historical data for the target area.

[0017] 2. Enhance the model's predictive ability for complex weather conditions through refined scenario decoupling. Unsupervised clustering is employed to classify typical weather scenarios, and dedicated prediction models are built for each scenario. This overcomes the problem that a single model cannot accurately fit the differentiated effects of different weather patterns, thereby improving the overall prediction accuracy of the model under various weather conditions such as sunny, cloudy, and rainy weather.

[0018] 3. A transferable and universal framework was established, significantly improving the efficiency and consistency of batch modeling across multiple sites. The standardized weather scenario system and model transfer process formed a reusable forecasting method. When deploying models for a large number of distributed photovoltaic sites, it avoided repetitive learning of underlying rules and achieved efficient replication and large-scale application of forecasting knowledge. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram illustrating the transfer learning principle of the present invention. Figure 2 This is a structural diagram of the ultra-short-term prediction model based on Seq2Seq in this invention. Detailed Implementation

[0021] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0022] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "joining" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0023] The present invention will now be described in further detail with reference to the accompanying drawings.

[0024] This invention proposes a multi-region photovoltaic power prediction method based on active power distribution networks. This method, based on meteorological resources and transfer learning, constructs a unified meteorological scenario framework and achieves scenario-based knowledge transfer, aiming to solve the problems of low prediction accuracy and modeling difficulties in photovoltaic power prediction in data-scarce areas. The scheme mainly includes three parts: weather prediction scenario segmentation based on k-means clustering algorithm, multi-region photovoltaic power prediction method based on transfer learning, and scenario-based model transfer. The steps of this invention are as follows: (1) Defining weather forecast scenarios based on k-means clustering algorithm Photovoltaic power generation is closely coupled with meteorological conditions, and the impact mechanisms and sensitivities of various meteorological factors on power differ significantly under different weather types. By classifying weather scenarios, a more targeted "meteorological-power" mapping model can be constructed within the same scenario category, thereby improving prediction accuracy. Furthermore, since the physical power generation patterns of distributed photovoltaic systems are universal, a unified weather scenario classification framework can be used to avoid repetitive learning of basic meteorological models and improve modeling efficiency when modeling massive numbers of sites. This invention employs the K-means clustering algorithm to classify historical weather prediction data into scenarios, obtaining sample subsets representing different weather types, and training prediction models based on each subset to more precisely characterize the photovoltaic power variation patterns under specific weather conditions.

[0025] K-means clustering is a classic unsupervised partitioning method based on distance-based iterative optimization. Its core idea is to divide data into K predetermined categories by minimizing the sum of squared distances from samples to cluster centers. When dealing with potentially complex distributions of weather data, such as meteorological features, the K-means algorithm can efficiently aggregate samples with similar numerical characteristics into the same cluster, effectively identifying different typical weather patterns. Compared to other clustering methods, K-means has a simple structure and fast convergence speed, making it particularly suitable for handling large-scale datasets. This provides a computational efficiency advantage for subsequent batch modeling of massive distributed photovoltaic systems. To ensure the representativeness and stability of the clustering results, this study comprehensively determines the optimal number of clusters K using indicators such as the silhouette coefficient, and employs multiple initializations to avoid local optima problems that may arise from initial centroid selection, thereby ensuring that the partitioned weather scenes have clear physical meaning and statistical representativeness.

[0026] The specific implementation steps of k-means-based cluster analysis for distributed photovoltaic power plants are as follows: Step 1: Data Preparation and Feature Engineering Historical numerical weather prediction (NWP) data for the same period were collected. Key meteorological factors included global irradiance, direct irradiance, temperature, humidity, wind speed, wind direction, and air pressure. The NWP data were processed into feature vectors corresponding to the prediction period. Due to the different dimensions and magnitudes of the various meteorological variables, the feature data had to be standardized. This study used Z-score standardization. The Z-score formula is: (1) in The original data values, The mean of all sample data for this feature. The standard deviation of all sample data for this feature is . This represents the new value after Z-score standardization.

[0027] Step 2: Determine the optimal number of clusters K 1) Preset K value range: Based on business experience and preliminary analysis, a reasonable K value search range is set. In this study, the K value is set to 2 to 8.

[0028] 2) Iterative calculation and evaluation: For each K value in the range, run the K-means algorithm and calculate the corresponding internal evaluation index, including the silhouette coefficient and the elbow rule.

[0029] Silhouette coefficient: measures how closely a sample is clustered with itself and how well it separates from other clusters. Its value ranges from -1 to 1; the higher the average value, the better the clustering effect.

[0030] Elbow rule: Calculate the sum of squared distances from all samples to their cluster centers for different K values. Plot the "K-inertia" curve; the K value corresponding to the inflection point of the curve is generally considered a better choice.

[0031] 3) Determine the K value comprehensively: Combining the maximum point of the contour coefficient, the inflection point of the elbow rule, and considering the interpretability and management complexity of weather type resolution in actual business, the optimal K value is finally determined to be 5.

[0032] Step 3: Perform K-means clustering 1) Initialization and iteration: Randomly select K initial cluster centers and iterate until convergence.

[0033] Assignment phase: Calculate the Euclidean distance from each sample point to all centroids, and assign it to the cluster containing the nearest centroid.

[0034] Update phase: Recalculate the mean of all sample points in each cluster and use this mean as the new cluster center.

[0035] 2) Addressing randomness: Since random selection of the initial centroid may affect the final result, the algorithm is usually run multiple times, and the clustering result with the least inertia is selected as the final solution to ensure the stability of the result.

[0036] Step 4: Clustering Result Analysis and Scenario Definition Scene feature analysis: Analyze the cluster center vector of each final cluster. Through denormalization, interpret the typical meteorological conditions represented by its values.

[0037] Sample attribution: Based on the clustering results, all historical data are labeled with "scene tags". All samples belonging to the same cluster constitute a subset of the dataset.

[0038] Visualization and Verification: Principal component analysis can be used to reduce high-dimensional features to 2-3 dimensions for visualization, allowing for intuitive observation of the separation of each scenario. Simultaneously, the historical distribution of actual photovoltaic output under each scenario is statistically analyzed to verify that there are indeed significant differences in power curves under different scenarios, thus confirming the physical meaning of the clustering results.

[0039] Step 5: Output and Application 1) Generate a scene rule base: Output the final K cluster center coordinates to form weather scene discrimination rules.

[0040] 2) Constructing a predictive workflow Training phase: Using labeled data, a dedicated photovoltaic power prediction model is trained for each subset of data for each weather scenario.

[0041] Prediction phase: When new NWP data is obtained, it is first standardized, then its distance from each scene center (rule base) is calculated, it is assigned to the nearest weather scene, and finally the dedicated model corresponding to that scene is called to perform power prediction.

[0042] (2) Multi-region photovoltaic power prediction method based on transfer learning The principle of transfer learning is as follows Figure 1 As shown, transfer learning can be applied to source domains with high-quality data. The power prediction model features under similar meteorological resources are extracted from photovoltaic power plants in China. First, based on source domain photovoltaic power plants with excellent data quality, the prediction model is then derived from the source domain dataset. Information is obtained from, among which Represents the characteristics of the spatial data space of the source domain. Indicate the corresponding label; then proceed with the source task. In scenarios with similar meteorological resources, key features of the power prediction model are extracted, and the data distribution of the source domain is learned. The label space representing the source task. Represents features in the source domain With tags The joint probability distribution, i.e., the statistical regularity of the source data; then, based on these acquired feature information, a mapping function is used. The function performs knowledge transfer, enabling the distribution of data learned from the source domain to be transferred. Feature transfer to the target domain; and thus, preliminary construction of the target domain. The basic framework of the photovoltaic power plant prediction model, the prediction model uses the target domain dataset Build, in Representing the characteristics of the target domain, The target domain label is represented; the information obtained by the model through transfer learning assists the target task. ,in It represents the data distribution of the target domain and fine-tunes the model parameters accordingly, optimizing the objective function by transferring knowledge from the source domain.

[0043] Through this series of operations, the effective features of the source domain model are ultimately transferred and reused to the target domain, helping to improve the power prediction performance of photovoltaic power plants in the target domain. The specific implementation steps of the multi-region photovoltaic power prediction method based on transfer learning are as follows: Step 1: Source Domain Modeling 1) Select the source domain and target domain Based on the photovoltaic power generation data from 10 photovoltaic stations, the source regions were selected as photovoltaic station 0, photovoltaic station 1, photovoltaic station 2, photovoltaic station 4, and photovoltaic station 7, according to their geographical distribution. The target regions were selected as photovoltaic station 3, photovoltaic station 5, photovoltaic station 6, photovoltaic station 8, and photovoltaic station 9.

[0044] NWP data and historical power data of each photovoltaic site are acquired, and a sliding window is used to construct the input feature sequence and output power sequence. A mutation detector is used to detect outliers, the degree of missing values ​​of variables is calculated, and independent variables with a missing value greater than 60% are removed; the remaining feature variables are used as the selected model input variables, and linear interpolation is used to fill in the missing values.

[0045] 2) Feature Engineering Meteorological characteristics are constructed, including total daily radiation, temperature, humidity, wind speed, and cloud cover.

[0046] (2) in, Represents a meteorological characteristic sequence. express Global horizontal irradiance at any given time express Direct irradiance at any given time express Temperature at any moment express Real-time numerical weather forecast - humidity express Wind speed at any moment express Real-time numerical weather forecast - wind direction express Atmospheric pressure at any given moment.

[0047] Time characteristics: hourly, daily, and seasonal periodic coding.

[0048] (3) in Represents a time-feature sequence. for Hour code of time for Daily cycle encoding of time, for Seasonal coding of time.

[0049] Sequence features: Constructing a time window for power features .

[0050] 3) Weather scene segmentation based on K-means One-hot encoding is used to segment weather forecast scenarios based on the k-means clustering algorithm. To construct input features suitable for machine learning models, after segmenting weather scenarios using the k-means algorithm, the resulting discrete scenario category labels need to be converted into numerical form. This scheme employs one-hot encoding: for a clustering result containing K clusters, the scenario label for each sample is converted into a K-dimensional binary vector. In this vector, only the dimension corresponding to the scenario category to which the sample belongs has a value of 1, while the other dimensions are 0. This encoding method can eliminate the model's misunderstanding of the potential order relationship between categories, transforming non-numerical scenario labels into a sparse, orthogonal feature representation, thereby facilitating the subsequent prediction model to effectively identify and utilize the differentiated power generation patterns under different weather scenarios.

[0051] 4) Construction and training of scenario-specific Seq2Seq models The photovoltaic power output prediction of this invention is constructed using a Seq2Seq model. The structure of the ultra-short-term prediction model based on Seq2Seq is as follows: Figure 2 As shown.

[0052] The prediction model employs a Seq2Seq architecture to perform feature extraction and prediction tasks on time series data. In the encoder module design, a two-layer bidirectional gated recurrent unit (BiGRU) structure is used for feature extraction. The first-layer BiGRU takes numerical weather prediction and historical power sequences as inputs, and through a bidirectional cyclic computation mechanism, effectively captures the bidirectional dependency between meteorological data and historical power, achieving in-depth mining of multi-source time series features. The feature sequence output from the first-layer BiGRU is passed to the second-layer BiGRU. The hidden state output by the second-layer BiGRU and the output state of the first-layer BiGRU together constitute the final state representation of the encoder, serving as the initial input to the decoder.

[0053] The decoder section uses numerical weather prediction data for future time periods as input sequences. It first receives and fuses the output states of the first-layer encoder, achieving preliminary integration of historical features and future meteorological information. The fused feature sequence is then fed into a second-layer GRU unit, which simultaneously receives the state information from the second-layer encoder. Through a gating mechanism, the fused features are dynamically filtered and enhanced, achieving secondary abstraction and deep fusion of the features. The feature vectors processed by the two decoding units are then mapped to temporal features through a fully connected layer, ultimately outputting the photovoltaic power prediction results for future time periods. This structural design fully leverages the bidirectional feature extraction capabilities of BiGRU and the gating mechanism advantages of GRU, enabling effective modeling and accurate prediction of multi-source time-series data.

[0054] Step 2: Design of Transfer Learning Strategy 1) Target area processing In the target region data preparation and scenario matching phase, the core task is to align the target region's data with the general weather scenario system of the source region using limited local data. First, short-term numerical weather forecast data and corresponding measured photovoltaic power data are collected from the target region to form the initial sample set for transfer learning. Then, scenario alignment is performed. Scenario alignment uses the cluster centers of the source region as initial centroids, and iteratively fine-tunes and optimizes the positions of the cluster centers using a small amount of sample data from the target region. Ultimately, each target region sample is assigned a specific weather scenario label, forming a data subset with a clear scenario structure consistent with the source region's architecture, providing a foundation for subsequent scenario-based model transfer.

[0055] 2) Transfer learning architecture design This invention adopts a hierarchical transfer strategy, dividing the model into three levels and designing transfer strategies for different levels: bottom-level transfer, middle-level transfer, and top-level fine-tuning.

[0056] The underlying feature extractor leverages transfer learning to extract features from meteorological time-series patterns. These underlying features include basic patterns of meteorological variables changing over time, such as the periodicity and trends of irradiance, temperature, and wind speed. Because the underlying feature extractor can capture universal features applicable to meteorological sequences in different regions, the parameters of the first BiGRU layer of the source region model encoder are frozen and directly used for the target region.

[0057] The middle layer transfers the mapping from meteorological sequences to power sequences. The later layers of the encoder and the earlier layers of the decoder are fine-tuned. These layers encode the complex nonlinear relationship of how meteorological variables affect the output power of photovoltaic panels, and are responsible for mapping the extracted temporal features to a contextual representation of power prediction. Therefore, the weights of the last layer of the encoder and the weights of the first layer of the decoder are transferred and fine-tuned using a moderate learning rate to adapt to the specific transformation relationship of the target region.

[0058] Top-level fine-tuning adapts to site-specific factors in the target area. The last few layers of the model reflect factors such as photovoltaic module type and efficiency, local shading and pollution levels, and regional microclimate characteristics. Significant differences in equipment characteristics and local environments among different photovoltaic sites can lead to deviations in output power even under similar weather conditions. Top-level fine-tuning corrects these deviations. Therefore, the encoder's second layer and output layer are completely reinitialized to adapt to the power dimensions of the target area.

[0059] This invention employs a progressive unfreezing strategy: in the first 10 rounds, only the last layer of the decoder is trained; in rounds 11-20, all layers of the decoder are unfrozen; in rounds 21-30, the last layer of the encoder is unfrozen; and after round 31, all layers are fine-tuned.

[0060] 3) Scenario-based model transfer Scene-based transfer: Based on the labels obtained after scene alignment of the target area data, a pre-trained scene-specific model in the source domain is loaded for each identified weather scene. This step ensures that the target area model can directly inherit the robust "weather-power" core mapping relationship that has been validated under similar meteorological conditions, laying a high-starting-point model foundation for rapid adaptation.

[0061] Adaptive fine-tuning phase: Targeted parameter adjustments are made to the loaded pre-trained model. To prevent overfitting due to limited data and to retain general knowledge learned from the source domain, an early stopping strategy is adopted, terminating training immediately when performance on the target domain validation set no longer improves. Additionally, L2 weight regularization is introduced to constrain model complexity, enabling the model to seek an optimal balance between prior knowledge from the source domain and new evidence from the target domain.

[0062] (3) Multi-regional prediction system integration Dynamic Scene Recognition and Model Selection: Upon obtaining new numerical weather forecast data for the target area, the system first performs standardized preprocessing identical to that used during the training phase. Then, the scene recognition module calculates the Euclidean distance between the standardized feature vector and the pre-stored cluster centroids of the K-means dataset. The system selects the scene with the smallest distance as the dominant weather type for the current forecast period. After scene classification, the system automatically calls the dedicated prediction model bound to that scene, which has undergone transfer learning and fine-tuning, inputs the processed meteorological sequence, and performs photovoltaic power prediction calculations. This process ensures that the system can dynamically select the most suitable physical law model based on real-time meteorological conditions, thereby improving the relevance and accuracy of the predictions.

[0063] System optimization and evaluation: Performance metrics primarily use root mean square error and mean absolute error, and the model's generalization ability and stability are evaluated for different scenarios. The system has a built-in progressive learning engine that can periodically (e.g., monthly or quarterly) incrementally train the fine-tuned model using new data to adapt to equipment aging, seasonal changes, or local climate drift.

[0064] Finally, any aspects not fully described in this invention utilize existing mature products and technologies.

[0065] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for predicting photovoltaic power in multiple regions of an active power distribution network, characterized in that, Includes the following steps: S1. Weather Scene Segmentation: Based on the K-means clustering algorithm, the historical numerical weather forecast data of the source region is segmented into weather prediction scenes. The historical numerical weather forecast data includes multiple meteorological factors and is subject to feature engineering, including standardization of the data using the Z-score standardization formula. S2. Source Domain Model Construction: In the source region, a dedicated sequence-to-sequence Seq2Seq prediction model is constructed for each weather scenario, and trained using a Seq2Seq model architecture. The Seq2Seq model includes an encoder and a decoder. The encoder adopts a bidirectional gated recurrent unit (BiGRU), and the decoder adopts a GRU. S3. Target Domain Transfer: For the target region, local data is used to complete scene matching. Based on the scene division criteria of the source region, a transfer learning strategy is used to transfer the pre-trained model parameters of the source domain to the target domain. The transfer learning strategy is a hierarchical transfer, including: bottom layer transfer, middle layer transfer and top layer fine-tuning, and a progressive unfreezing strategy is used for fine-tuning. The bottom layer transfer freezes the first layer BiGRU parameters of the encoder, the middle layer transfer fine-tunes the last layer of the encoder and the front layer of the decoder, and the top layer fine-tuning reinitializes the second layer of the encoder and the output layer. S4. Prediction Execution: In the multi-region prediction system, the weather scene to which the input meteorological data belongs is dynamically identified. The scene is classified by calculating the Euclidean distance from the center of the scene cluster, and the corresponding fine-tuned Seq2Seq model is called to predict photovoltaic power. Finally, a distance-weighted multi-model integration prediction mechanism is introduced to handle the ambiguity of scene boundaries, and the model is updated regularly with new data to optimize prediction performance.

2. The method for predicting photovoltaic power in multiple regions of an active distribution network according to claim 1, characterized in that: The Z-score standardization formula in S1 is: ； in The original data values, The mean of all sample data for this feature. The standard deviation of all sample data for this feature is . This represents the new value after Z-score standardization; The optimal number of clusters K is determined by the silhouette coefficient and the elbow rule, and K-means clustering is performed to divide the weather scene into K types.

3. The method for predicting photovoltaic power in multiple regions of an active distribution network according to claim 1, characterized in that: The input feature sequence constructed in S2 includes meteorological feature sequence and time feature sequence; The meteorological feature sequence is as follows: ； in, Represents a meteorological characteristic sequence. express Global horizontal irradiance at any given time express Direct irradiance at any given time express Temperature at any moment express Real-time numerical weather forecast - humidity express Wind speed at any moment express Real-time numerical weather forecast - wind direction express Atmospheric pressure at any given moment; The time feature sequence is: ； in Represents a time-feature sequence. for Hour code of time for Daily cycle encoding of time, for Seasonal coding of time.

4. The method for predicting photovoltaic power in multiple regions of an active distribution network according to claim 1, characterized in that: The specific methods for determining the optimal number of clusters K in S1 include: The preset range of K values ​​is from 2 to 8. For each K value, the K-means algorithm is run to calculate the profile coefficient and the inertia value of the elbow rule. The optimal K value is determined by combining the maximum point of the profile coefficient and the inflection point of the elbow rule.

5. The method for predicting photovoltaic power in multiple regions of an active distribution network according to claim 1, characterized in that: When performing K-means clustering in S1, multiple initializations are used to avoid local optima, and the cluster center coordinates are output to form a weather scene discrimination rule base.

6. The method for predicting photovoltaic power in multiple regions of an active distribution network according to claim 1, characterized in that: The training process of the Seq2Seq model in S2 includes: The encoder adopts a two-layer BiGRU structure. The first layer BiGRU takes meteorological feature sequences and historical power sequences as inputs, and the second layer BiGRU outputs the hidden state. The decoder takes numerical weather forecast data for future periods as inputs, and after fusing the encoder state, outputs the prediction result through GRU units and a fully connected layer.

7. The method for predicting photovoltaic power in multiple regions of an active distribution network according to claim 1, characterized in that: The specific process of the gradual thawing strategy in S3 includes: The first 10 training rounds only train the last layer of the decoder. In rounds 11-20, all layers of the decoder are unfrozen. In rounds 21-30, the last layer of the encoder is unfrozen. After round 31, all layers are fine-tuned, and an early stopping strategy and L2 regularization are used to prevent overfitting.

8. The method for predicting photovoltaic power in multiple regions of an active distribution network according to claim 1, characterized in that: In S4, the prediction performance evaluation uses root mean square error and mean absolute error as indices, and evaluates the model's generalization ability in different scenarios.

9. A system used in the multi-region photovoltaic power prediction method for active distribution networks as described in claim 1, characterized in that, include: The weather scene segmentation module is configured to perform the weather scene segmentation function of S1, including data standardization, K value determination, and K-means clustering; The source domain model building module is configured to perform the source domain model building function of S2, including feature sequence construction and Seq2Seq model training; The transfer learning module is configured to perform S3's target domain transfer functions, including scene matching, hierarchical transfer, and fine-tuning. The prediction execution module is configured to perform S4's prediction functions, including dynamic scene recognition, model invocation, and ensemble prediction. The system optimization module is configured to periodically update the model and evaluate performance to achieve continuous learning. The weather scene segmentation module further includes a feature engineering submodule, which is used to process numerical weather forecast data and apply the Z-score normalization formula. The source domain model construction module further includes a sequence feature construction submodule, which is used to generate meteorological feature sequences and time feature sequences as model inputs; The prediction execution module further integrates a distance-weighted multi-model mechanism to handle scenario uncertainties.

Images