Plum rain season photovoltaic output prediction method and system based on small sample learning

Abstract

The invention discloses a plum rain season photovoltaic output prediction method and system based on small sample learning. The method comprises the following steps: constructing a multi-source feature system; clustering and dividing meta-learning tasks based on a meteorological mode; the method comprises the following steps: constructing a TCN-Transform hybrid neural network model, and embedding a learnable meteorological adapter; training a hybrid neural network model by adopting an MAML meta-learning framework, and realizing meta-learning task training of different meteorological models in combination with a physical constraint regularization loss function; model parameters are optimized in three stages through a course learning strategy, and a training set is enhanced by utilizing Mixup time sequence data; and deploying the trained model to carry out output prediction, and outputting a prediction result with a confidence interval. According to the method, high-precision photovoltaic output prediction in a small-sample and high-uncertainty plum rain season scene is realized, and reliable support is provided for power grid dispatching.

Description

Technical Field

[0001] The present invention belongs to the technical field of intersection of artificial intelligence and power systems, and specifically relates to a method for predicting photovoltaic output during the rainy season based on small sample learning. Background Art

[0002] As a key component of clean energy, photovoltaic power generation's output forecast accuracy directly impacts grid dispatch efficiency and operational safety. However, existing forecasting technologies face multiple bottlenecks during unique meteorological scenarios, such as the plum rain season. Due to the short annual occurrence of plum rains and the limited availability of historical data, traditional data-driven models are prone to overfitting due to insufficient training samples.

[0003] The unique alternation of short bursts of heavy rain and intermittent clear skies during the plum rain season causes the output curve to exhibit pulsating and dramatic fluctuations. Physical models based on numerical weather forecasts are limited by their hourly update frequency, making it difficult to capture the sudden changes in meteorological conditions even at the minute level. Statistical learning methods, however, lack the ability to accurately analyze the complex coupling between cloud thickness, humidity gradients, and output due to their limited nonlinear modeling capabilities. More seriously, current mainstream forecasting models often ignore the physical constraints of photovoltaic systems, frequently resulting in risky conditions where predicted values ​​exceed the theoretical maximum output of panels during continuous rainy and overcast conditions. They even produce counter-physical logic where predicted output decreases as sunlight increases, posing a serious threat to power grid security.

[0004] Although transfer learning technology can alleviate the problem of data scarcity, it has not built a structured task system for cloudy and rainy weather patterns; although data enhancement methods can expand the sample size, traditional time series interpolation will destroy the physical continuity of the output curve; although physical information neural networks attempt to integrate prior knowledge, they bring high computational costs due to solving partial differential equations, making it difficult to meet the real-time prediction needs of power plants.

[0005] There is also a patent document such as Chinese patent CN202311764842.X, which is publicly named "Stacking Plum Rainy Season Photovoltaic Power Prediction Method and System Based on BiGRU-Attention". It mainly models photovoltaic time series data through BiGRU and Attention mechanisms, emphasizes network structure optimization and data interpolation methods, and is suitable for prediction tasks under normal data volume conditions.

[0006] These technical shortcomings collectively point to a core problem that has not yet been overcome in the current field - how to achieve dual guarantees of sensitivity to meteorological mutations and consistency with physical laws under limited samples. This is the technical problem currently faced. Summary of the Invention

[0007] In order to solve the above problems, the present invention proposes a method and system for predicting photovoltaic output in the rainy season based on small sample learning.

[0008] The technical solution of the present invention is: a method for predicting photovoltaic output in the rainy season based on small sample learning, comprising the following steps:

[0009] S1. Obtain historical output data, real-time meteorological data, and equipment status data of photovoltaic power plants to build a multi-source feature system;

[0010] S2. Based on the multi-source feature system, the meta-learning tasks are divided based on meteorological pattern clustering;

[0011] S3. Build a TCN-Transformer hybrid neural network model and embed a learnable meteorological adapter;

[0012] S4. Based on S2 and S3, the MAML meta-learning framework is used to train the hybrid neural network model, and combined with the physical constraint regularization loss function to realize the meta-learning task training of different meteorological models;

[0013] S5. Optimize model parameters in three stages through a curriculum learning strategy and enhance the training set using Mixup time series data.

[0014] S6. Deploy the trained model to perform power output prediction and output the prediction results with confidence intervals. Apply the model to other regions for power output prediction through transfer learning.

[0015] The multi-source feature system includes meteorological features, equipment features and time series features, among which,

[0016] The meteorological characteristics include cloud cover and its time difference value, shortwave radiation and its time difference value, relative humidity and its time difference value;

[0017] The equipment characteristics include panel temperature, aging coefficient, and theoretical maximum output value;

[0018] The time series characteristics include the sliding mean of output in the previous three hours and the output quantiles in the same period in history.

[0019] In step S2, a dynamic time warping algorithm is used to cluster historical meteorological data. The clustering results correspond to different meteorological patterns, including continuous rainfall, intermittent clearing, short-term heavy rain, rapid clearing, and slow rainfall. Each meteorological pattern constitutes a meta-learning task, which includes a support set and a query set.

[0020] In step S3, the weather adapter structure can be learned as:

[0021] The input meteorological features are passed through the gated neural network to generate the adjustment coefficient α, which is calculated as follows:

[0022] α=σ(W2·Re LU(W1·x weather +b1)+b2)

[0023] Where σ is the Sigmoid function, W1 and W2 are the neural network weight parameters, and x weather is the meteorological parameter, b1 and b2 are the neural network bias parameters, and ReLU is the neural network activation function.

[0024] In step S4, the physical constraint regularization loss function is:

[0025] L=L MSE +λ1∑max(0,P pred -P theory ) 2 +λ2∑I(ΔR<0)·|ΔP|

[0026] Where L is the physical constraint regularization loss function, L MSE is the sample loss function value, λ1 is the output physical constraint weight coefficient, λ2 is the output meteorological monotonic relationship constraint weight coefficient, P pred is the predicted value, P theory is the theoretical maximum value, ΔR is the change in light intensity, ΔP is the change in output, and I is the indicative function.

[0027] In step S5, the course learning strategy includes three stages:

[0028] In the first stage, only fair weather data is used for model pre-training;

[0029] In the second phase, cloudy alternating weather data is added to activate the weather adapter parameter update;

[0030] In the third stage, continuous rainfall data is introduced to freeze the underlying parameters of the TCN network.

[0031] In step S5, the Mixup time series data enhancement is specifically as follows:

[0032] For two sample sequences (X i ,Y i ) and (X j ,Y j ) performs linear interpolation to generate new samples:

[0033] X new =λX i +(1-λ)X j

[0034] Y new =λY i +(1-λ)Y j

[0035] Where, X new is the new sample feature vector, Y new is the new sample label vector, λ is the interpolation coefficient, X i is the eigenvector of sample sequence i, X j is the feature vector of sample sequence j, Y i is the label vector of sample sequence i, Y j is the label vector of sample sequence j.

[0036] In step S6, transfer learning method:

[0037] Preserve weather adapter and Transformer encoder parameters;

[0038] Replace the TCN decoder layer of the target area and perform fine-tuning using local small sample data.

[0039] In step S6, the confidence interval is generated by:

[0040] Enable Monte Carlo Dropout during model inference, repeatedly sample the prediction results N times, calculate the mean μ and standard deviation σ, and the confidence interval is defined as [μ-1.96σ,μ+1.96σ].

[0041] A photovoltaic output prediction system for the rainy season based on small sample learning, comprising:

[0042] A construction module is used to obtain historical output data, real-time meteorological data, and equipment status data of photovoltaic power plants to build a multi-source feature system;

[0043] A partitioning module is used to partition meta-learning tasks based on meteorological pattern clustering on the basis of a multi-source feature system;

[0044] The network model module is used to build a TCN-Transformer hybrid neural network model and embed a learnable meteorological adapter. The training module is used to train the hybrid neural network model using the MAML meta-learning framework and combine it with a physical constraint regularization loss function to implement meta-learning task training for different meteorological models.

[0045] The optimization module is used to optimize model parameters in three stages through a curriculum learning strategy and enhance the training set using Mixup time series data;

[0046] The prediction module deploys the trained model for power output forecasting, outputs prediction results with confidence intervals, and applies the model to other regions for power output forecasting through transfer learning.

[0047] By introducing a small-sample meta-learning method based on MAML, the present invention significantly improves the rapid adaptability and generalization performance of the photovoltaic output prediction model under the changeable meteorological conditions of the rainy season, and realizes high-precision photovoltaic output prediction under the small-sample and high-uncertainty rainy season scenario, providing reliable support for power grid scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0049] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0050] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0051] To address the technical difficulties in photovoltaic output forecasting during the rainy season, such as data scarcity, complex meteorological changes, and easy violation of physical laws, this paper proposes a small sample prediction method that integrates meta-learning and physical constraints, and achieves high-precision and high-robustness output forecasting through multi-dimensional technological innovation.

[0052] like Figure 1 As shown in FIG, a method for predicting photovoltaic output in the rainy season based on small sample learning includes the following steps:

[0053] S1. Acquire historical output data, real-time meteorological data, and equipment status data from PV power plants to construct a multi-source feature system. This data source includes not only traditional meteorological factors such as solar irradiance, temperature, and humidity, but also incorporates PV power plant equipment status data, such as component temperature and power output. By integrating this multi-dimensional feature data, we can more comprehensively capture the various factors affecting PV power output, providing richer input data for subsequent model training. Unlike conventional methods that rely on a single data source, this comprehensive processing of multi-source data lays the foundation for improving model prediction accuracy and robustness.

[0054] S2. Meta-learning tasks are divided based on meteorological pattern clustering, with each task containing a support set and a query set. The dynamic time warping (DTW) algorithm is used to cluster historical meteorological data, dividing meteorological patterns into several typical types, such as continuous rain, intermittent clear weather, or sudden rainstorms. Each clustering result represents a meteorological pattern, and each pattern constitutes a meta-learning task, which includes a support set and a query set. The support set is used to train the model, while the query set is used to evaluate and test the model. This clustering method based on meteorological patterns is different from the conventional method of training all data equally. It can establish independent prediction tasks for different meteorological conditions, thereby effectively improving training efficiency and accuracy.

[0055] S3. Build a TCN-Transformer hybrid neural network model and embed a learnable weather adapter.

[0056] The TCN-Transformer hybrid neural network model used in this paper consists of five main parts: an input layer, a temporal convolution module (TCN module), a Transformer encoder module, a meteorological adapter module, and an output layer. The input layer is used to receive a multi-source feature sequence that integrates meteorological factors, device status, and time information, and performs linear mapping and normalization on it. The TCN module is composed of a multi-layer one-dimensional dilated convolutional network, combined with residual connections to extract short-term dynamic features within a local time window.

[0057] The Transformer encoder module consists of a multi-head self-attention mechanism and a feedforward neural network, which is combined with positional encoding to model long-range temporal dependencies.

[0058] The weather adapter module dynamically adjusts the weight response of the Transformer output according to the current weather type through a learnable modulation factor, improving the model's adaptability to different weather scenarios;

[0059] The output layer uses a fully connected network to map the fused spatiotemporal features and output the predicted PV output for the corresponding time step. This network structure balances local trend perception with global dependency modeling, demonstrating excellent nonlinear fitting capabilities and meteorological generalization. Compared to conventional single-model systems, this hybrid architecture more effectively integrates local and global features, improving adaptability to complex weather conditions.

[0060] S4. The model is trained using the Model-Agnostic Meta-Learning (MAML) meta-learning framework, combined with a physical constraint regularization loss function. MAML allows the model to quickly adapt to new tasks with a small number of samples and can achieve efficient transfer learning across different meteorological tasks. At the same time, to ensure the physical rationality of the prediction results, a physical constraint regularization loss function is incorporated. Physical constraints include the maximum output power upper limit of the photovoltaic power station and the monotonicity constraint of light-power, ensuring that the model's prediction results are physically feasible and do not exceed the actual working capacity of the photovoltaic panels. This is different from conventional methods that rely solely on data-driven approaches. The addition of physical constraints effectively improves the reliability and accuracy of the model.

[0061] S5. We optimize model parameters in stages through a curriculum learning strategy and enhance the training set with Mixup time series data. Initially, we pre-train the model using relatively simple sunny data, gradually adding more complex meteorological pattern data, and finally training with rainfall data. The introduction of a curriculum learning strategy prevents the model from being disturbed by complex data in the early stages of training, improving training stability and convergence speed. Furthermore, we utilize Mixup time series data augmentation technology to expand the training set, generating more samples and further improving the model's generalization capabilities. Unlike traditional methods that directly use the entire training dataset for training, curriculum learning and data augmentation effectively improve the model's robustness in the presence of small samples.

[0062] S6. Deploy the trained model to perform output forecasts, outputting prediction results with confidence intervals. Compared to traditional single-prediction methods, this can better reflect the reliability of prediction results in practical applications, especially in complex meteorological environments, effectively reducing decision-making risks.

[0063] The present invention specifically comprises the following steps:

[0064] S1. Construction of multi-source feature system

[0065] This method integrates meteorological data, equipment status, and output time series characteristics to construct a high-information-density, multi-source feature system. Meteorological features include cloud cover dynamics, shortwave radiation intensity, and relative humidity gradients; equipment features include panel temperature decay curves and theoretical maximum output values; and time series features extract the output sliding mean of the previous three hours and the historical quantiles for the same period.

[0066] The dynamic time warping (DTW) algorithm is used to align the timestamps of multi-source data, and the isolation forest algorithm is introduced to clean the abnormal zero-value data caused by heavy rain to form a normalized feature matrix with a dimension of 12, providing high-quality input for subsequent modeling.

[0067] S2. Meta-learning task division based on meteorological patterns

[0068] To address the sudden changes in weather during the plum rain season, a meta-learning task generation method driven by meteorological patterns is proposed. First, the similarity of five consecutive days of meteorological sequences is measured using the DTW distance, and the Ward hierarchical clustering algorithm is used to classify meteorological patterns into five types: continuous rainfall, intermittent clearing, short-term heavy rain, rapid clearing, and slow rainfall.

[0069] Continuous rainfall type: refers to the continuous precipitation process over a long period of time, with uninterrupted rainfall and the overall process being stable and continuous.

[0070] Intermittent clearing type: refers to one or more short-term pauses in precipitation during the rainfall process, followed by another period of rainfall.

[0071] Short-term rainstorm type: refers to a precipitation process with high intensity in a relatively short period of time, with concentrated precipitation and rapid rain.

[0072] Quick clearing type: refers to the situation where after the rainfall ends, the clouds quickly dissipate and the weather changes from cloudy and rainy to sunny in a short period of time.

[0073] Slow rainfall type: refers to the gradual development of the rainfall process, the slow increase in rain intensity, the duration of precipitation is long, and it is not obvious in the early stage.

[0074] For each type of pattern, a meta-learning task is constructed, which includes a support set (20 samples) and a query set (50 samples). The support set focuses on the short-term fluctuation patterns of similar weather conditions, while the query set verifies the generalization ability across time periods, forming a "task cluster" structure to enhance the model's adaptability to cloudy and rainy scenarios.

[0075] S3, TCN-Transformer hybrid neural network model construction

[0076] A hybrid neural network model architecture was designed for collaborative extraction of spatiotemporal features: the TCN module employs four layers of causal convolution, capturing the local decay pattern of the output curve through exponentially increasing expansion coefficients (1, 2, 4, and 8). The Transformer encoder is equipped with a learnable position embedding and utilizes a multi-head attention mechanism to model the long-range, hourly dependency of cloud cover and radiation. The core innovation lies in the embedding of a learnable weather adapter, which maps real-time weather features into channel adjustment coefficients through a gating network, dynamically modulating the TCN intermediate layer features, and achieving feature adaptation that is sensitive to sudden weather changes.

[0077] α=σ(W2·ReL U(W1·x weather +b1)+b2)

[0078] Where σ is the Sigmoid function, which multiplies the adjustment coefficient by the intermediate feature representation of the network channel by channel. W1 and W2 are the neural network weight parameters, x weatheris the meteorological parameter, b1 and b2 are the neural network bias parameters, and ReLU is the neural network activation function.

[0079] S4. Meta-learning model training with physical regularization constraints

[0080] Integrating dual physical constraints in the MAML meta-learning framework: Theoretical output hard constraints penalize the predicted value beyond the maximum theoretical value L of the panel through regularization terms theory =∑max(0,P pred -P theory ) 2 , where P pred is the predicted value, P theory is the theoretical maximum value. The light-power monotonicity soft constraint imposes a linear penalty on the case where the radiation is enhanced but the predicted output decreases (L mono =∑I(ΔR<0)·|ΔP|), where ΔR is the change in light intensity, ΔP is the change in output, and I is the indicator function. The meta-training process consists of an inner and outer loop: an inner loop rapidly adapts to task characteristics on the support set (three-step gradient updates, lr=0.01), while an outer loop optimizes shared parameters across tasks (AdamW optimizer, lr=0.001), ensuring that the model is both task-adaptive and physically consistent.

[0081] The physical constraint regularization loss function is:

[0082] L=L MSE +λ1∑max(0,P pred -P theory ) 2 +λ2∑I(ΔR<0)·|ΔP|

[0083] Where L is the physical constraint regularization loss function, L MSE is the sample loss function value, λ1 is the output physical constraint weight coefficient, λ2 is the output meteorological monotonic relationship constraint weight coefficient, P pred is the predicted value, P theory is the theoretical maximum value, ΔR is the change in light intensity, ΔP is the change in output, and I is the indicative function.

[0084] S5. Optimize the course learning strategy in stages

[0085] Use progressive training strategies to improve model robustness:

[0086] Phase 1 (10 rounds): Pre-train the basic network using only fair weather data, fix the weather adapter parameters, and learn a stable output mode;

[0087] Phase 2 (20 rounds): Introducing cloudy alternating weather data, unlocking weather adapter parameters, and activating the ability to dynamically adjust meteorological characteristics;

[0088] Phase 3 (10 rounds): Add continuous rainfall data and freeze the underlying TCN parameters, focusing on fine-tuning the Transformer and weather adapter.

[0089] Combined with Mixup time series data enhancement, linear interpolation (λ~Beta(0.3,0.3)) is performed on similar weather samples to expand the small sample size while maintaining the physical continuity of the output curve, ultimately achieving a smooth transition from simple to complex scenarios.

[0090] Mixup time series data enhancement is specifically as follows:

[0091] For two sample sequences (X i ,Y i ) and (X j ,Y j ) performs linear interpolation, where X is the meteorological sequence and Y is the power sequence, to generate new samples:

[0092] X new =λX i +(1-λ)X j

[0093] Y new =λY i +(1-λ)Y j

[0094] Where, X new is the new sample feature vector, Y new is the new sample label vector, λ is the interpolation coefficient, X i is the eigenvector of sample sequence i, X j is the feature vector of sample sequence j, Y i is the label vector of sample sequence i, Y j is the label vector of sample sequence j, λ~Beta(α,α), α∈[0.2,0.4].

[0095] S6. Model cross-region migration method:

[0096] Preserve weather adapter and Transformer encoder parameters;

[0097] Replace the TCN decoder layer of the target area and perform fine-tuning using local small sample data.

[0098] The confidence interval is generated as follows:

[0099] Enable Monte Carlo Dropout during model inference, repeatedly sample the prediction results N times, calculate the mean μ and standard deviation σ, and the confidence interval is defined as [μ-1.96σ,μ+1.96σ].

[0100] Aiming at the small sample and high volatility meteorological characteristics of the plum rain season, this paper proposes a complete prediction system based on small sample learning. The core introduces the MAML meta-learning framework to enable the model to quickly adapt to sudden weather scenarios. By clustering weather patterns to divide meta-learning tasks, designing a TCN-Transformer hybrid neural network, and embedding a learnable meteorological adapter, the system realizes dynamic adjustment of features under meteorological disturbances. In combination with physical constraint regularization terms, the system ensures that the prediction results conform to the actual physical laws of the photovoltaic system. The system further improves the training stability and generalization ability through course learning and Mixup enhancement.

[0101] A photovoltaic output prediction system for the rainy season based on small sample learning, comprising:

[0102] A construction module is used to obtain historical output data, real-time meteorological data, and equipment status data of photovoltaic power plants to build a multi-source feature system;

[0103] A partitioning module is used to partition meta-learning tasks based on meteorological pattern clustering on the basis of a multi-source feature system;

[0104] The network model module is used to build a TCN-Transformer hybrid neural network model and embed a learnable meteorological adapter. The training module is used to train the hybrid neural network model using the MAML meta-learning framework and combine it with a physical constraint regularization loss function to implement meta-learning task training for different meteorological models.

[0105] The optimization module is used to optimize model parameters in three stages through a curriculum learning strategy and enhance the training set using Mixup time series data;

[0106] The prediction module deploys the trained model for power output forecasting, outputs prediction results with confidence intervals, and applies the model to other regions for power output forecasting through transfer learning.

[0107] The present invention has the following technical effects:

[0108] 1) Significantly improve prediction accuracy and generalization ability in small sample scenarios

[0109] By dividing different weather forms into independent prediction tasks and introducing the MAML meta-learning mechanism, the model can quickly adapt to sudden weather changes under the condition of limited historical samples, solving the problems of traditional deep models' strong data dependence and poor generalization ability.

[0110] 2) Enhance the model’s robustness to complex weather disturbances

[0111] By constructing a learnable meteorological adapter, the model can dynamically adjust its attention to different input features when facing sudden meteorological changes (such as sudden heavy rain and continuous rain), thereby improving its stability under strong nonlinear conditions during the rainy season.

[0112] 3) Ensure the physical consistency and interpretability of prediction results

[0113] The output upper limit constraint and light-power monotonicity regularization term based on the operating mechanism of photovoltaic power stations are introduced to prevent the neural network from generating predicted values ​​that deviate significantly from the actual operating rules, thereby enhancing the credibility and physical rationality of the results.

[0114] 4) Alleviate overfitting and improve training stability through data enhancement and curriculum training

[0115] The Mixup data augmentation strategy expands the distribution range of small sample data, and the course-based unlocking training process improves the parameter convergence speed and robustness, avoiding the problems of unstable model training or decreased generalization ability in complex scenarios.

[0116] 5) Have good cross-scenario migration capabilities

[0117] Through the structural design that separates the TCN decoder from the model, the model only needs a small number of target domain samples to complete high-precision migration prediction when it is deployed at a new site or promoted in new scenarios, reducing deployment costs and improving engineering practicality.

[0118] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for predicting photovoltaic output in the rainy season based on small sample learning, characterized in that: The following steps are involved: S1. Obtain historical output data, real-time meteorological data, and equipment status data of photovoltaic power plants to build a multi-source feature system; S2. Based on the multi-source feature system, the meta-learning tasks are divided based on meteorological pattern clustering; S3. Build a TCN-Transformer hybrid neural network model and embed a learnable meteorological adapter; S4. Based on S2 and S3, the MAML meta-learning framework is used to train the hybrid neural network model, and combined with the physical constraint regularization loss function to realize the meta-learning task training of different meteorological models; S5. Optimize model parameters in three stages through a curriculum learning strategy and enhance the training set using Mixup time series data. S6. Deploy the trained model to perform power output prediction and output the prediction results with confidence intervals. Apply the model to other regions for power output prediction through transfer learning.

2. The method according to claim 1, characterized in that The multi-source feature system includes meteorological features, equipment features and time series features, among which, The meteorological characteristics include cloud cover and its time difference value, shortwave radiation and its time difference value, relative humidity and its time difference value; The equipment characteristics include panel temperature, aging coefficient, and theoretical maximum output value; The time series characteristics include the sliding mean of output in the previous three hours and the output quantiles in the same period in history.

3. The method according to claim 1, characterized in that In step S2, The dynamic time warping algorithm is used to cluster historical meteorological data. The clustering results correspond to different meteorological patterns, including continuous rainfall, intermittent clearing, short-term heavy rain, rapid clearing, and slow rainfall. Each weather pattern constitutes a meta-learning task, which consists of a support set and a query set.

4. The method according to claim 1, wherein In step S3, the weather adapter structure can be learned as: The input meteorological features are passed through the gated neural network to generate the adjustment coefficient α, which is calculated as follows: α=σ(W2·ReLU(W1·x weather +b1)+b2) Where σ is the Sigmoid function, W1 and W2 are the neural network weight parameters, and x weather is the meteorological parameter, b1 and b2 are the neural network bias parameters, and ReLU is the neural network activation function.

5. The method according to claim 1, characterized in that In step S4, the physical constraint regularization loss function is: Where L is the physical constraint regularization loss function, L MSE is the sample loss function value, λ1 is the output physical constraint weight coefficient, λ2 is the output meteorological monotonic relationship constraint weight coefficient, P pred is the predicted value, P theory is the theoretical maximum value, ΔR is the change in light intensity, ΔP is the change in output, and I is the indicative function.

6. The method according to claim 1, characterized in that In step S5, the course learning strategy includes three stages: In the first stage, only fair weather data is used for model pre-training; In the second phase, cloudy alternating weather data is added to activate the weather adapter parameter update; In the third stage, continuous rainfall data is introduced to freeze the underlying parameters of the TCN network.

7. The method according to claim 1, characterized in that In step S5, the Mixup time series data enhancement is specifically as follows: For two sample sequences (X i ,Y i ) and (X j ,Y j ) performs linear interpolation to generate new samples: X new =λX i +(1-λ)X j AND new =λY i +(1-λ)Y j Where, X new is the new sample feature vector, Y new is the new sample label vector, λ is the interpolation coefficient, X i is the eigenvector of sample sequence i, X j is the feature vector of sample sequence j, Y i is the label vector of sample sequence i, Y j is the label vector of sample sequence j.

8. The method according to claim 1, characterized in that In step S6, transfer learning method: Preserve weather adapter and Transformer encoder parameters; Replace the TCN decoder layer of the target area and perform fine-tuning using local small sample data.

9. The method according to claim 1, characterized in that In step S6, the confidence interval is generated by: Enable Monte Carlo Dropout during model inference, repeatedly sample the prediction results N times, calculate the mean μ and standard deviation σ, and the confidence interval is defined as [μ-1.96σ,μ+1.96σ].

10. A rainy season photovoltaic output prediction system based on small sample learning, characterized by: include: A construction module is used to obtain historical output data, real-time meteorological data, and equipment status data of photovoltaic power plants to build a multi-source feature system; A partitioning module is used to partition meta-learning tasks based on meteorological pattern clustering on the basis of a multi-source feature system; The network model module is used to build a TCN-Transformer hybrid neural network model and embed a learnable meteorological adapter; The training module is used to train the hybrid neural network model using the MAML meta-learning framework and combine it with the physical constraint regularization loss function to implement meta-learning task training for different meteorological models; The optimization module is used to optimize model parameters in three stages through a curriculum learning strategy and enhance the training set using Mixup time series data; The prediction module deploys the trained model for power output forecasting, outputs prediction results with confidence intervals, and applies the model to other regions for power output forecasting through transfer learning.

Images