A Method and System for Predicting Photovoltaic Power at New Plateau Power Stations Based on Transfer Learning

Abstract

This invention relates to a method and system for predicting photovoltaic power at newly constructed high-altitude power stations based on transfer learning. The method includes: identifying photovoltaic power stations with abundant historical data that match the station's altitude, solar radiation characteristics, climate conditions, and hardware configuration; constructing a power prediction model, which includes a feature extraction module and a task adaptation module. The feature extraction module comprises multiple sequentially connected convolutional layers and a feature attention module; the task adaptation module comprises at least two sequentially connected convolutional layers and a fully connected layer; pre-training the power prediction model using historically rich photovoltaic power station data, and re-training the task adaptation module using historical data from the newly constructed high-altitude power stations to obtain the final trained power prediction model; and performing photovoltaic power prediction at the newly constructed high-altitude power stations. This invention can better achieve photovoltaic power forecasting under conditions of scarce power station data in high-altitude areas.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power prediction technology, and in particular relates to a method and system for predicting photovoltaic power at newly built plateau power stations based on transfer learning. Background Technology

[0002] Currently, vigorously developing new energy sources such as solar energy has become an inevitable trend. However, photovoltaic power generation is highly volatile and uncertain. Therefore, conducting high-precision photovoltaic power forecasting is an effective means to promote photovoltaic consumption and is of paramount importance. The harsh and complex climate of plateau regions makes the installation and construction of photovoltaic panels difficult. For newly operational photovoltaic power plants in plateau regions, historical power generation data is scarce, making accurate power forecasting challenging.

[0003] Photovoltaic power generation is primarily determined by solar irradiance, but is also affected by temperature. Generally, higher solar irradiance results in greater power generation, and lower temperatures lead to stronger photovoltaic panel output. In high-altitude regions such as the Qinghai-Tibet Plateau, the thin air, low temperatures, and harsh conditions make photovoltaic panel installation difficult. However, the Qinghai-Tibet Plateau is one of the richest regions in my country in terms of solar energy resources, with significant photovoltaic power generation potential. Therefore, it is necessary to conduct specialized research on photovoltaic power forecasting for plateau regions.

[0004] The most significant difference between plateau regions and other areas lies in the comprehensive environmental differences brought about by altitude. High altitude results in thinner air, a thinner atmosphere leading to reduced solar radiation attenuation and significantly increased irradiance; simultaneously, lower temperatures are conducive to improving the conversion efficiency of photovoltaic cells. These unique geographical and climatic conditions present plateau regions with both construction and maintenance challenges in photovoltaic power generation, as well as unparalleled resource advantages.

[0005] The literature He Y, Gao Q, Jin Y, et al. Short-term photovoltaic power forecasting method based on convolutional neural network[J]. Energy Reports, 2022, 8: 54-62. uses historical photovoltaic power data of the power plant to forecast future photovoltaic power, showing good forecast accuracy in scenarios with abundant historical photovoltaic power data. However, existing photovoltaic power forecasting methods only forecast photovoltaic power for power plants with sufficient historical power data, relying on a large amount of historical power generation data. They cannot conduct high-precision photovoltaic power forecasting for newly operating power plants with scarce data. At the same time, existing technologies often target photovoltaic power plants in plains or low-altitude areas, without considering characteristics such as altitude. This makes it difficult to achieve accurate forecasting accuracy in photovoltaic forecasting in extremely high-altitude areas such as the Qinghai-Tibet Plateau, and even more difficult to achieve accurate photovoltaic power forecasting for newly built power plants in plateau areas with scarce data. Summary of the Invention

[0006] To address the power forecasting problem for data-sparse photovoltaic (PV) power stations in plateau regions, this invention aims to construct a PV power forecasting method for data-sparse stations based on deep CNN (Convolutional Neural Network) transfer learning. By transferring and sharing knowledge between PV power stations with abundant historical data and newer stations with scarce historical data, and incorporating features such as altitude and temperature that reflect the unique natural conditions of plateau regions, this invention achieves high-precision PV power forecasting for data-sparse stations in plateau regions.

[0007] Given the unique characteristics of high-altitude regions, it is necessary to select characteristic variables that fully reflect their physical mechanisms when constructing photovoltaic power forecasting models. In addition to historical photovoltaic power data, this application also selected four features: irradiance, clear-sky irradiance, temperature, and altitude. This constitutes a complete physics-driven forecasting framework. Irradiance, as the core input variable, directly quantifies the energy source of the photovoltaic system. Clear-sky irradiance provides a reference benchmark for irradiance under ideal cloudless conditions, characterizing the maximum possible value of solar irradiance in cloudless situations. Clear-sky irradiance is mainly determined by the relative motion of the Earth and the Sun, and is a function of longitude, latitude, and time, serving as important future prior information in photovoltaic power forecasting. Temperature captures the thermal effect loss of photovoltaic cells, which is particularly crucial in high-altitude regions—the efficiency improvement brought by the low-temperature environment can partially offset the additional costs of high-altitude construction and operation and maintenance. Altitude, as a static geographical feature, implicitly includes environmental factors such as air pressure and oxygen content that affect equipment performance. These features are complementary and indispensable.

[0008] The photovoltaic power prediction method for newly built plateau power stations based on transfer learning provided in this application specifically includes the following steps:

[0009] S1. Obtain the historical photovoltaic power, irradiance, clear sky irradiance, temperature, and altitude of the newly built plateau power station to be predicted, as well as the clear sky irradiance of the future time to be predicted. Based on the altitude, solar radiation characteristics, climate conditions, and hardware configuration factors of the newly built plateau power station, determine the photovoltaic power stations with rich historical data that match it through multi-factor quantitative similarity ranking, and obtain the historical data of the photovoltaic power stations with rich historical data.

[0010] S2. Construct a power prediction model, which includes a feature extraction module and a task adaptation module. The feature extraction module contains multiple convolutional layers and a feature attention module connected in sequence. The task adaptation module contains at least two convolutional layers and a fully connected layer connected in sequence. The output of the last convolutional layer in the feature extraction module is connected to the feature attention module to obtain attention weights. The obtained attention weights are multiplied element-wise with the output of the last convolutional layer in the feature extraction module to obtain a new matrix, which is used as the input of the convolutional layer in the task adaptation module.

[0011] S3. Use historical photovoltaic power station data to pre-train the power prediction model, and use historical data from newly built plateau power stations to retrain the task adaptation module in the power prediction model to obtain the final power prediction model after training.

[0012] S4. Use the trained final power prediction model to predict the photovoltaic power of the newly built plateau power station.

[0013] Furthermore, based on the altitude, solar radiation characteristics, climate conditions, and hardware configuration of newly built plateau power stations, the following methods are used to determine which photovoltaic power stations with abundant historical data match them through multi-factor quantitative similarity ranking:

[0014] Altitude is the core dominant factor, with an altitude difference of ≤500m between the source and target domains as the quantitative standard. This ensures that the atmospheric physical characteristics of the source and target domains are consistent, fundamentally guaranteeing the similarity of radiation transmission and temperature environment, while ensuring that the total horizontal radiation deviation is ≤15% and the deviation of the direct radiation ratio is ≤10%.

[0015] Climate conditions are matched with an annual average relative humidity deviation of ≤12% and an annual average number of dusty days deviation of ≤8% to avoid regional distribution mismatch caused by differences in attenuation mechanisms.

[0016] The efficiency deviation of hardware configuration control components is ≤3%, the installed capacity ratio is in the range of 0.5~2, and the system topology is consistent, so as to reduce the interference of equipment differences on the effectiveness of migration.

[0017] Furthermore, when determining which photovoltaic power stations with abundant historical data match the newly built plateau power stations based on their altitude, solar radiation characteristics, climate conditions, and hardware configuration, multi-factor quantitative similarity ranking is also included:

[0018] First, a rigid screening is performed using altitude as the main control factor, retaining only existing stations with an altitude difference of ≤500m from the target station into the candidate set. For candidate stations that pass the altitude screening, the relative difference mapping method is used to convert the original differences between them and the target station in three categories of factors—solar radiation, climate conditions, and hardware configuration—into similarity within a unified interval. Then, an equal-weighted arithmetic mean is used to obtain a comprehensive similarity score, thereby completing the ranking of candidate source domains.

[0019] Furthermore, the power prediction model is pre-trained using historical photovoltaic power plant data, and the task adaptation module in the power prediction model is re-trained using historical data from newly built plateau power plants. The method to obtain the final trained power prediction model is as follows:

[0020] First, the model is pre-trained using photovoltaic power plants with abundant historical data to obtain a pre-trained model. Then, the parameters of the pre-trained model are fine-tuned using a small amount of historical data from newly built power plants. During the fine-tuning process, the feature extraction module of the pre-trained model is frozen, the task adaptation module of the model is fine-tuned, and the convolution kernel and bias term in the task adaptation module are updated through error backpropagation. After retraining the pre-trained model with a small amount of historical data from newly built power plants, a fine-tuned power prediction model is obtained.

[0021] Furthermore, the calculation formula for the convolutional layer in the feature extraction module is as follows:

[0022] ;

[0023] in, These are the input features of the i-th layer. It is the output feature of the j-th layer. It is a weight matrix. It is a bias matrix; symbol It's a convolution operation. This represents the ReLU function, a non-linear activation function.

[0024] Furthermore, the method for predicting photovoltaic power at newly built plateau power stations using the trained final power prediction model is as follows:

[0025] The historical irradiance, historical clear-sky irradiance, historical temperature, historical photovoltaic power, altitude at the prediction time, and clear-sky irradiance at the prediction time of the newly built plateau power station are input into the trained final power prediction model. The network output of the final power prediction model is the photovoltaic power at the prediction time.

[0026] A new photovoltaic power prediction system for high-altitude power stations based on transfer learning is used to implement the above-mentioned methods, including:

[0027] The historical data and photovoltaic power station acquisition module is used to acquire the historical photovoltaic power, historical irradiance, historical clear sky irradiance, historical temperature, altitude, and clear sky irradiance for future moments of the newly built plateau power station to be predicted. Based on the altitude, solar radiation characteristics, climate conditions, and hardware configuration factors of the newly built plateau power station, the module uses multi-factor quantitative similarity ranking to identify photovoltaic power stations with rich historical data that match it, and acquires the historical data of the photovoltaic power stations with rich historical data.

[0028] A power prediction model construction module is used to construct a power prediction model. The prediction model includes a feature extraction module and a task adaptation module. The feature extraction module contains multiple convolutional layers and a feature attention module connected in sequence. The task adaptation module contains at least two convolutional layers and a fully connected layer connected in sequence. The output of the last convolutional layer in the feature extraction module is connected to the feature attention module to obtain attention weights. The obtained attention weights are multiplied element-wise with the output of the last convolutional layer in the feature extraction module to obtain a new matrix, which is used as the input to the convolutional layer in the task adaptation module.

[0029] The power prediction model training module is used to pre-train the power prediction model using historical photovoltaic power station data with abundant data, and to retrain the task adaptation module in the power prediction model using historical data from newly built plateau power stations, so as to obtain the final power prediction model after training.

[0030] The photovoltaic power prediction module is used to predict the photovoltaic power of newly built plateau power stations using the trained final power prediction model.

[0031] Furthermore, the present invention adopts the following technical solution:

[0032] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the photovoltaic power prediction method for newly built plateau power stations based on transfer learning as described above.

[0033] Furthermore, the present invention adopts the following technical solution:

[0034] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the photovoltaic power prediction method for newly built plateau power stations based on transfer learning as described above.

[0035] Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0036] 1. This invention proposes a deep learning framework that integrates convolutional neural network (CNN) feature extraction, attention mechanism, and transfer learning to solve the problem of photovoltaic power prediction under the condition of scarce data for new power generation stations in plateau areas. By explicitly embedding physical mechanisms into the feature space, the prediction model can more accurately capture the laws of photovoltaic power generation in plateau areas, providing reliable decision support for grid dispatch and energy management. This feature engineering method based on physical cognition has stronger interpretability and generalization ability compared with pure data-driven black box models.

[0037] 2. This invention proposes a deep learning framework that integrates convolutional neural network (CNN) feature extraction, attention mechanism and transfer learning, which effectively extracts the temporal dependencies contained in features such as photovoltaic power, irradiance, clear sky irradiance, temperature and altitude, and can better realize photovoltaic power forecasting under the condition of scarce field data in plateau areas. Attached Figure Description

[0038] Figure 1 A schematic diagram of the network structure of the power prediction model provided in an embodiment of the present invention;

[0039] Figure 2 A schematic diagram illustrating the principle of photovoltaic power forecasting for newly built power plants based on transfer learning, provided in an embodiment of the present invention.

[0040] Figure 3 This is a schematic diagram of the transfer learning training structure of the power prediction model provided in an embodiment of the present invention;

[0041] Figure 4 A comparison chart of photovoltaic power forecasts provided for embodiments of the present invention. Detailed Implementation

[0042] The photovoltaic power prediction method and system for newly built plateau power stations based on transfer learning provided by the present invention will be further described clearly and completely below with reference to the accompanying drawings:

[0043] Example 1

[0044] Given the unique characteristics of high-altitude regions, it is necessary to select characteristic variables that fully reflect their physical mechanisms when constructing photovoltaic power forecasting models. In addition to historical photovoltaic power data, this application also selected four features: irradiance, clear-sky irradiance, temperature, and altitude. This constitutes a complete physics-driven forecasting framework. Irradiance, as the core input variable, directly quantifies the energy source of the photovoltaic system. Clear-sky irradiance provides a reference benchmark for irradiance under ideal cloudless conditions, characterizing the maximum possible value of solar irradiance in cloudless situations. Clear-sky irradiance is mainly determined by the relative motion of the Earth and the Sun, and is a function of longitude, latitude, and time, serving as important future prior information in photovoltaic power forecasting. Temperature captures the thermal effect loss of photovoltaic cells, which is particularly crucial in high-altitude regions—the efficiency improvement brought by the low-temperature environment can partially offset the additional costs of high-altitude construction and operation and maintenance. Altitude, as a static geographical feature, implicitly includes environmental factors such as air pressure and oxygen content that affect equipment performance. These features are complementary and indispensable. To better predict photovoltaic power at power plants in high-altitude areas, it is necessary to effectively extract the temporal dependencies contained in features such as photovoltaic power, irradiance, clear-sky irradiance, temperature, and altitude. This embodiment proposes a deep learning framework that integrates convolutional neural network (CNN) feature extraction, attention mechanism, and transfer learning to solve the problem of photovoltaic power prediction under the condition of scarce data at new power plants in high-altitude areas.

[0045] This invention provides a method for predicting photovoltaic power at newly built plateau power stations based on transfer learning, comprising the following steps:

[0046] S1. Obtain the historical photovoltaic power, irradiance, clear sky irradiance, temperature, altitude, and clear sky irradiance for the future time to be predicted for the newly built plateau power station to be predicted. Based on the altitude, solar radiation characteristics, climate conditions, and hardware configuration factors of the newly built plateau power station, determine the photovoltaic power stations with rich historical data that match it through multi-factor quantitative similarity ranking, and obtain the historical data of the photovoltaic power stations with rich historical data, including the historical photovoltaic power, irradiance, clear sky irradiance, temperature, altitude, and clear sky irradiance for the future time to be predicted.

[0047] S2. Construct a power prediction model, which includes a feature extraction module and a task adaptation module. The feature extraction module contains multiple convolutional layers (Conv1-Conv5) connected in sequence and a feature attention module. Each convolutional layer is followed by a batch normalization and pooling layer to progressively reduce the feature dimension and enhance the feature expressive power, used to automatically extract multi-level feature representations from the input data. The task adaptation module contains at least two convolutional layers and a fully connected layer connected in sequence. The output of the last convolutional layer in the feature extraction module is connected to the feature attention module to obtain attention weights. The obtained attention weights are multiplied element-wise with the output of the last convolutional layer in the feature extraction module to obtain a new matrix, which is used as the input of the convolutional layer in the task adaptation module.

[0048] Specifically, such as Figure 1 As shown, Figure 1 A schematic diagram of the network structure of the power prediction model provided in this embodiment is given. The input of the network is the photovoltaic power at a historical time, the irradiance at a historical time, the clear sky irradiance at a historical time, the temperature at a historical time, the altitude, and the clear sky irradiance at a future time. The output of the network is the photovoltaic power at a future time.

[0049] It should be noted that the convolutional layers in the one-dimensional convolutional feature extraction module for temporal data used in this application are recommended to be 2 to 5 layers. This is because this range can extract the instantaneous changes, fluctuation trends, periodic patterns, and long-range dependencies of temporal data layer by layer, constructing a complete temporal feature pyramid from the bottom to the top. Under the constraint of a lightweight structure, it can ensure that the receptive field gradually expands without losing temporal details, avoid feature redundancy and gradient vanishing, and make the extracted temporal features more comprehensive and discriminative, thereby improving the model performance and generalization ability. In this embodiment, the feature extraction module includes 5 convolutional layers (Conv1-Conv5) connected in sequence. Each convolutional layer is followed by a batch normalization and pooling layer to gradually reduce the feature dimension and enhance the feature expressive power, which is used to automatically extract multi-level feature representations from the input data.

[0050] The calculation formula for each convolutional layer is as follows:

[0051] ;

[0052] in, These are the input features of the i-th layer. It is the output feature of the j-th layer. It is a weight matrix. It is a bias matrix; symbol It's a convolution operation. ReLU (Rectified Linear Unit) represents a non-linear activation function.

[0053] Traditional neural networks assign equal weights to all feature variables, while attention mechanisms can adaptively assign weights to different feature variables to characterize their impact on the prediction task. Therefore, by introducing an attention mechanism, the weights of each feature in future photovoltaic power prediction can be dynamically and adaptively adjusted.

[0054] When applying attention mechanisms, Figure 1 The output of the Conv5 layer in the convolutional neural network obtains attention weights through the feature attention module. Then attention weights and the output matrix of the Conv5 layer of the convolutional neural network Element-wise multiplication yields a new matrix C, which serves as the input to the Conv6 layer.

[0055] ;

[0056] Attention weight Adaptive adjustments can be made through backpropagation during network training. Through an attention mechanism, the power prediction model can automatically identify the most important historical time steps for the current prediction task, thereby achieving adaptive weighted fusion of features such as photovoltaic power, irradiance, clear-sky irradiance, temperature, and altitude, and thus effectively improving the prediction accuracy of photovoltaic power.

[0057] Attention mechanisms can more accurately focus on key temporal features and important channels, suppress redundant information interference, and significantly improve the effectiveness and discriminativeness of features under lightweight structural constraints, thereby enhancing the model's ability to represent temporal data and adapt to tasks. The task adaptation module achieves secondary refinement and dimensional alignment of temporal features through two convolutional layers. Compared with single-layer convolution, multi-layer convolution can introduce nonlinear transformations, expand the receptive field to capture long temporal dependencies, and achieve multi-level abstraction and noise reduction refinement of features. Finally, the fully connected layer completes the global nonlinear fusion of features and task space mapping.

[0058] S3. Use historical photovoltaic power station data to pre-train the power prediction model, and use historical data from newly built plateau power stations to retrain the task adaptation module in the power prediction model to obtain the final power prediction model after training.

[0059] In practical applications, due to the long accumulation period of photovoltaic power data, newly built photovoltaic power plants typically require months or even years to obtain sufficient training data. In scenarios where data is scarce for newly built photovoltaic power plants, traditional machine learning methods often suffer from severely insufficient photovoltaic power forecasting accuracy due to a lack of adequate training data. To address this issue, this embodiment proposes a photovoltaic power forecasting method for newly built power plants based on deep transfer learning. This method aims to significantly improve the photovoltaic power forecasting accuracy of newly built power plants by introducing information from power plants with abundant historical data and achieving knowledge sharing between them.

[0060] like Figure 2 As shown, Figure 2 The schematic diagram of the photovoltaic power forecasting for newly built photovoltaic power plants based on transfer learning provided in this embodiment is given; this application introduces the idea of ​​transfer learning into the power forecasting of newly built photovoltaic power plants in plateau areas, and designs a method such as... Figure 3 The deep CNN transfer learning framework shown is Figure 3 The network shown is divided into two main parts:

[0061] (1) Feature extraction module: It includes 5 convolutional layers (Conv1-Conv5) and a feature attention module, which are used to automatically extract multi-level feature representations from the input data. The weights of this part (convolutional kernel parameters in the convolutional layers and weight parameters in the feature attention module) are fixed after pre-training on photovoltaic power station data with rich historical data.

[0062] (2) Task adaptation module: It includes two convolutional layers (Conv6-Conv7) and a fully connected layer (FC). The weights of this part (including convolutional kernel parameters and fully connected layer weights) are fine-tuned and optimized using a small amount of data from newly built photovoltaic power stations.

[0063] By using deep CNN transfer learning, this application realizes information sharing between photovoltaic power stations with abundant historical data and newly built power stations in plateau areas with scarce data, thereby significantly improving the photovoltaic power forecasting accuracy of newly built power stations in plateau areas with scarce data;

[0064] When training the power prediction model, the model is first pre-trained using photovoltaic power plants with abundant historical data to obtain a pre-trained model. Then, the parameters of the pre-trained model are fine-tuned using a small amount of historical data from newly built power plants. During the fine-tuning process, the feature extraction module of the pre-trained model is frozen, the task adaptation module of the model is fine-tuned, and the convolution kernel and bias term in the task adaptation module are updated through error backpropagation. After retraining the pre-trained model with a small amount of historical data from newly built power plants, the fine-tuned power prediction model is obtained.

[0065] It should be noted that the photovoltaic power stations with abundant historical data used in the pre-training of the power prediction model were selected based on the following criteria: photovoltaic power stations with abundant historical data were chosen that were similar to the newly built plateau power stations in terms of altitude, solar radiation characteristics, climate conditions, and hardware configuration. This is explained in detail below:

[0066] In transfer learning, existing photovoltaic (PV) power plants with similar altitudes and sufficient historical data to newly built PV power plants on plateaus are selected as the source domain. The core basis is the consistency of the altitude-dominated PV power generation physical mechanism, and screening and verification are conducted based on four key factors: altitude, solar radiation characteristics, climate conditions, and hardware configuration. Altitude is the core control factor, and a quantitative screening standard of ≤500m altitude difference between the source and target power plants is recommended to ensure that the atmospheric physical characteristics such as atmospheric pressure and molecular density are basically the same in both locations, ensuring strong similarity in radiation transmission paths and temperature environments from a mechanistic perspective. Simultaneously, the deviation of total horizontal radiation should be ≤15%, and the deviation of direct radiation proportion should be ≤10% to ensure the effective transfer of the "radiation-output" mapping law. Climate conditions are matched with an annual average relative humidity deviation of ≤12% and an annual average number of dusty days deviation of ≤8% to avoid domain distribution mismatch caused by differences in attenuation mechanisms. Hardware configuration prioritizes a component efficiency deviation of ≤3%, an installed capacity ratio in the range of 0.5 to 2, and a consistent system topology to reduce the interference of equipment differences on the effectiveness of transfer. The optimal source domain is determined by ranking similarity using multi-factor quantification. First, a rigid screening is performed using altitude as the primary controlling factor, retaining only existing stations with an altitude difference ≤ 500m from the target station in the candidate set. For candidate stations that pass the altitude screening, a relative difference mapping method is used to convert the original differences between them and the target station in three categories—solar radiation, climate conditions, and hardware configuration—into a unified similarity interval. Specifically, the absolute difference between the source and target station indicators is calculated, then divided by the larger of the two station indicator values ​​to obtain the relative difference. Finally, subtracting this relative difference from 1 normalizes the similarity of individual indicators to the [0,1] interval; the closer the value is to 1, the more similar the indicator features. This method eliminates differences in dimensions and orders of magnitude between different indicators, making various features directly comparable. Then, an equal-weighted arithmetic average is used to obtain a comprehensive similarity score, thus completing the candidate source domain ranking. This ensures both the physical consistency dominated by altitude and the matching of multi-dimensional features, avoiding negative transfer at the source. By utilizing long-term time-series data from the source domain to complete model pre-training, the problem of insufficient samples for newly built power plants can be solved. While ensuring the model's lightweight nature, the prediction generalization accuracy can be improved, giving full play to the application value of transfer learning in high-altitude photovoltaic scenarios.

[0067] As an example, in this embodiment, to verify the effectiveness of the proposed method in photovoltaic power forecasting, an experiment was conducted using one year of historical data from two photovoltaic power plants. First, photovoltaic power, irradiance, clear-sky irradiance, and temperature data from two power plants (Zele and Gar) were selected, with a time resolution of one hour.

[0068] First, the power prediction model was pre-trained using data from the Zele power plant. The first 70% of the data from the Zele power plant was selected as the training set, and the last 30% as the test set. After pre-training, the model was fine-tuned using data from the first three days of the Gar power plant. The last 30% of the data from the Gar power plant was used as the test set, which did not overlap with the training set, to evaluate the accuracy of the fine-tuned model in photovoltaic power prediction at the Gar power plant.

[0069] In contrast, we directly used the data from the first 3 days of the Gar site as the training set to train a CNN model (the CNN model here is a comparative experimental model with a network structure of 5 convolutional layers plus a fully connected architecture, which is a mature existing network model), and used the last 30% of the data from the Gar site as the test set to evaluate the prediction accuracy of the model trained directly on 3 days of historical data at the Gar site.

[0070] This example uses a 4-hour advance timescale as the forecast target. The network inputs are the historical irradiance, clear sky irradiance, temperature, and photovoltaic power of the past 24 hours, altitude (1 data copy 24 times), and clear sky irradiance of the 4th hour in the future (1 data copy 24 times). The network output is the photovoltaic power of the 4th hour in the future.

[0071] For the evaluation of forecast accuracy, the standardized mean absolute error (nMAE) and the standardized root mean square error (nRMSE) are used as evaluation indicators. The smaller the nMAE and nRMSE, the higher the forecast accuracy.

[0072]

[0073]

[0074] in, The rated power of the photovoltaic panel. This represents the actual value of photovoltaic power. This is the predicted value for photovoltaic power. The number of samples.

[0075] The results are shown in Table 1 and Figure 4As shown, this demonstrates that the method of this patent can significantly improve the accuracy of photovoltaic power forecasting on newly built power plants with scarce data.

[0076] Table 1. Accuracy of Photovoltaic Power Forecast 4 Hours Ahead for Gar Sites

[0077] The CNN was directly trained using three days of data from the new research station (traditional method). 0.1360 0.2741 Transfer learning in deep CNNs (method used in this application) 0.0515 0.1076

[0078] From Table 1 and Figure 4 As can be seen, the method proposed in this application achieves significantly higher forecast accuracy than traditional methods, with the forecast values ​​closely matching the actual values, resulting in excellent forecast performance. For newly constructed photovoltaic power plants with scarce data, traditional methods suffer from insufficient forecast accuracy due to data limitations. The method proposed in this application incorporates information from power plants with abundant historical data, enabling information sharing between data-rich photovoltaic power plants and newly constructed power plants in data-scarce plateau regions. This significantly improves the photovoltaic power forecast accuracy for newly constructed power plants in data-scarce plateau regions.

[0079] Example 2

[0080] This invention also provides a photovoltaic power prediction system for newly built plateau power stations based on transfer learning, used to implement the method in Example 1, including:

[0081] The historical data and photovoltaic power station acquisition module is used to acquire the historical photovoltaic power, historical irradiance, historical clear sky irradiance, historical temperature, altitude, and clear sky irradiance for future moments of the newly built plateau power station to be predicted. Based on the altitude, solar radiation characteristics, climate conditions, and hardware configuration factors of the newly built plateau power station, the module uses multi-factor quantitative similarity ranking to identify photovoltaic power stations with rich historical data that match it, and acquires the historical data of the photovoltaic power stations with rich historical data.

[0082] A power prediction model construction module is used to construct a power prediction model. The prediction model includes a feature extraction module and a task adaptation module. The feature extraction module contains multiple convolutional layers and a feature attention module connected in sequence. The task adaptation module contains at least two convolutional layers and a fully connected layer connected in sequence. The output of the last convolutional layer in the feature extraction module is connected to the feature attention module to obtain attention weights. The obtained attention weights are multiplied element-wise with the output of the last convolutional layer in the feature extraction module to obtain a new matrix, which is used as the input to the convolutional layer in the task adaptation module.

[0083] The power prediction model training module is used to pre-train the power prediction model using historical photovoltaic power station data with abundant data, and to retrain the task adaptation module in the power prediction model using historical data from newly built plateau power stations, so as to obtain the final power prediction model after training.

[0084] The photovoltaic power prediction module is used to predict the photovoltaic power of newly built plateau power stations using the trained final power prediction model.

[0085] Furthermore, the present invention adopts the following technical solution:

[0086] A non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements the photovoltaic power prediction method for newly built plateau power stations based on transfer learning in Example 1.

[0087] Furthermore, the present invention adopts the following technical solution:

[0088] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the photovoltaic power prediction method for newly built plateau power stations based on transfer learning in Embodiment 1.

[0089] From the above description of the embodiments, those skilled in the art will clearly understand that the facilities of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Embodiments of the present invention can be implemented using existing processors, or by dedicated processors used for this or other purposes for suitable systems, or by hardwired systems. Embodiments of the present invention also include non-transitory computer-readable storage media, comprising machine-readable media for carrying or having machine-executable instructions or data structures stored thereon; such machine-readable media can be any available medium accessible by a general-purpose or special-purpose computer or other machine with a processor. For example, such machine-readable media can include RAM, ROM, EPROM, EEPROM, CD-ROM or other optical disc storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store the required program code in the form of machine-executable instructions or data structures and is accessible by a general-purpose or special-purpose computer or other machine with a processor. When information is transmitted or provided to a machine via a network or other communication connection (hardwired, wireless, or a combination of hardwired and wireless), that connection is also considered a machine-readable medium.

[0090] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for predicting photovoltaic power at newly built plateau power stations based on transfer learning, characterized in that, Includes the following steps: S1. Obtain the historical photovoltaic power, irradiance, clear sky irradiance, temperature, and altitude of the newly built plateau power station to be predicted, as well as the clear sky irradiance of the future time to be predicted. Based on the altitude, solar radiation characteristics, climate conditions, and hardware configuration factors of the newly built plateau power station, determine the photovoltaic power stations with rich historical data that match it through multi-factor quantitative similarity ranking, and obtain the historical data of the photovoltaic power stations with rich historical data. S2. Construct a power prediction model, which includes a feature extraction module and a task adaptation module. The feature extraction module contains multiple convolutional layers and a feature attention module connected in sequence. The task adaptation module contains at least two convolutional layers and a fully connected layer connected in sequence. The output of the last convolutional layer in the feature extraction module is connected to the feature attention module to obtain attention weights. The obtained attention weights are multiplied element-wise with the output of the last convolutional layer in the feature extraction module to obtain a new matrix, which is used as the input of the convolutional layer in the task adaptation module. S3. Use historical photovoltaic power station data to pre-train the power prediction model, and use historical data from newly built plateau power stations to retrain the task adaptation module in the power prediction model to obtain the final power prediction model after training. S4. Use the trained final power prediction model to predict the photovoltaic power of the newly built plateau power station. Among them, when determining the photovoltaic power stations with abundant historical data that match the newly built plateau power stations based on their altitude, solar radiation characteristics, climate conditions, and hardware configuration factors, the following methods are used: Altitude is the core dominant factor, with an altitude difference of ≤500m between the source and target domains as the quantitative standard. This ensures that the atmospheric physical characteristics of the source and target domains are consistent, fundamentally guaranteeing the similarity of radiation transmission and temperature environment, while ensuring that the total horizontal radiation deviation is ≤15% and the deviation of the direct radiation ratio is ≤10%. Climate conditions are matched with an annual average relative humidity deviation of ≤12% and an annual average number of dusty days deviation of ≤8% to avoid regional distribution mismatch caused by differences in attenuation mechanisms. The efficiency deviation of hardware configuration control components is ≤3%, the installed capacity ratio is in the range of 0.5~2, and the system topology is consistent, so as to reduce the interference of equipment differences on the effectiveness of migration. The method for pre-training the power prediction model using historical photovoltaic power plant data and then re-training the task adaptation module of the power prediction model using historical data from newly built plateau power plants to obtain the final trained power prediction model is as follows: First, the model is pre-trained using photovoltaic power plants with abundant historical data to obtain a pre-trained model. Then, the parameters of the pre-trained model are fine-tuned using a small amount of historical data from newly built power plants. During the fine-tuning process, the feature extraction module of the pre-trained model is frozen, the task adaptation module of the model is fine-tuned, and the convolution kernel and bias term in the task adaptation module are updated through error backpropagation. After retraining the pre-trained model with a small amount of historical data from newly built power plants, a fine-tuned power prediction model is obtained.

2. The photovoltaic power prediction method for newly built plateau power stations based on transfer learning according to claim 1, characterized in that, When determining which photovoltaic power stations with abundant historical data match the newly built plateau power stations based on their altitude, solar radiation characteristics, climate conditions, and hardware configuration, multi-factor quantitative similarity ranking is also used. First, a rigid screening is performed using altitude as the main control factor, retaining only existing stations with an altitude difference of ≤500m from the target station into the candidate set. For candidate stations that pass the altitude screening, the relative difference mapping method is used to convert the original differences between them and the target station in three categories of factors—solar radiation, climate conditions, and hardware configuration—into similarity within a unified interval. Then, an equal-weighted arithmetic mean is used to obtain a comprehensive similarity score, thereby completing the ranking of candidate source domains.

3. The photovoltaic power prediction method for newly built plateau power stations based on transfer learning according to claim 1, characterized in that, The calculation formula for the convolutional layer in the feature extraction module is as follows: ； in, These are the input features of the i-th layer. It is the output feature of the j-th layer. It is a weight matrix. It is a bias matrix; symbol It's a convolution operation. This represents the ReLU function, a non-linear activation function.

4. The photovoltaic power prediction method for newly built plateau power stations based on transfer learning according to claim 1, characterized in that, The method for predicting photovoltaic power at newly built plateau power stations using the trained final power prediction model is as follows: The historical irradiance, historical clear-sky irradiance, historical temperature, historical photovoltaic power, altitude at the prediction time, and clear-sky irradiance at the prediction time of the newly built plateau power station are input into the trained final power prediction model. The network output of the final power prediction model is the photovoltaic power at the prediction time.

5. A photovoltaic power prediction system for newly built plateau power stations based on transfer learning, used to implement the method described in any one of claims 1-4, characterized in that, include: The historical data and photovoltaic power station acquisition module is used to acquire the historical photovoltaic power, historical irradiance, historical clear sky irradiance, historical temperature, altitude, and clear sky irradiance for future moments of the newly built plateau power station to be predicted. Based on the altitude, solar radiation characteristics, climate conditions, and hardware configuration factors of the newly built plateau power station, the module uses multi-factor quantitative similarity ranking to identify photovoltaic power stations with rich historical data that match it, and acquires the historical data of the photovoltaic power stations with rich historical data. A power prediction model construction module is used to construct a power prediction model. The prediction model includes a feature extraction module and a task adaptation module. The feature extraction module contains multiple convolutional layers and a feature attention module connected in sequence. The task adaptation module contains at least two convolutional layers and a fully connected layer connected in sequence. The output of the last convolutional layer in the feature extraction module is connected to the feature attention module to obtain attention weights. The obtained attention weights are multiplied element-wise with the output of the last convolutional layer in the feature extraction module to obtain a new matrix, which is used as the input to the convolutional layer in the task adaptation module. The power prediction model training module is used to pre-train the power prediction model using historical photovoltaic power station data with abundant data, and to retrain the task adaptation module in the power prediction model using historical data from newly built plateau power stations, so as to obtain the final power prediction model after training. The photovoltaic power prediction module is used to predict the photovoltaic power of newly built plateau power stations using the trained final power prediction model.

6. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the photovoltaic power prediction method for newly built plateau power stations based on transfer learning as described in any one of claims 1 to 4.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the photovoltaic power prediction method for newly built plateau power stations based on transfer learning as described in any one of claims 1 to 4.

Images