A photovoltaic power station output power prediction method, a terminal and a medium

Abstract

The application provides a photovoltaic power station output power prediction method, a terminal and a medium, the method comprising: collecting real-time meteorological data and preprocessing; constructing a self-organizing mapping network to identify and classify different environmental modes under unsupervised conditions, and simultaneously predicting photovoltaic power generation power; designing a regional model of a cross-region knowledge transfer algorithm based on a meta-learning model to transfer knowledge between different regions; predicting the category of the current meteorological environment according to newly collected real-time meteorological data, then calculating the prediction error of each regional model under the new environmental data, and selecting the regional model with the minimum error to predict the power; and constructing a personalized recommendation strategy model according to the predicted power to realize real-time feedback and optimization of the power prediction. The application has strong adaptability and generalization ability, and provides more accurate decision support for intelligent scheduling of photovoltaic power generation and power grid management.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of machine learning, and particularly relates to a photovoltaic power station output power prediction method, a terminal and a medium. BACKGROUND

[0002] As one of the main forces of renewable energy, photovoltaic power generation technology plays a crucial role in the transformation of global energy structure. With the continuous expansion of photovoltaic power generation scale, the power prediction of photovoltaic power stations has gradually become a key link in energy dispatching and power grid stability guarantee. Due to the characteristics of photovoltaic power generation, i.e. the significant influence of light intensity, weather conditions and seasonal changes on power generation, accurate power prediction not only helps energy dispatching and avoids power grid fluctuations, but also improves the economic benefits and operational efficiency of photovoltaic power generation. However, existing photovoltaic power prediction methods still face multiple technical bottlenecks, affecting the prediction accuracy and system adaptability, especially in the face of dynamic weather changes, cross-regional differences and environmental sudden factors.

[0003] Traditional photovoltaic power prediction methods rely on statistical models, physical models and machine learning models. Statistical models usually fit the relationship between photovoltaic power generation and meteorological factors through historical data, but these methods have great limitations in dealing with complex weather changes. Physical models are based on the physical principles of photovoltaic power generation process, simulating the power generation process under different weather conditions, but due to the need to consider too many parameters and the high complexity of the model, real-time response is often difficult to achieve. In addition, with the rise of machine learning technology, regression analysis or deep learning methods based on historical data and weather conditions have been widely applied. However, although these methods can improve prediction accuracy in some cases, they still have many problems. For example, in the face of sudden weather changes (such as heavy rain, strong light, sudden wind, etc.), traditional machine learning models cannot dynamically adjust model parameters, resulting in decreased prediction accuracy. In addition, due to differences in geography and climate conditions, cross-regional photovoltaic stations often have weak generalization ability when using the same model for prediction, resulting in unsatisfactory prediction results. Especially in newly deployed photovoltaic power stations, there is usually a lack of sufficient local data, making it difficult to effectively train and optimize the model, which makes the accuracy and stability of cross-regional photovoltaic power prediction particularly prominent.

[0004] Currently, although some advanced deep learning methods (such as convolutional neural networks, long short-term memory networks, etc.) have been used for photovoltaic power prediction, these methods usually rely on a large amount of training data and are not sensitive enough to real-time changes in weather conditions. In particular, in different regions and under different weather conditions, it is difficult for traditional deep learning models to achieve rapid model migration and real-time adaptation. In addition, existing methods usually lack the ability to dynamically switch between different environmental conditions, so their performance is still not ideal when dealing with complex weather changes and diverse meteorological environments. At the same time, although some methods attempt to enhance stability by combining multiple prediction models, these methods still fail to effectively address the rapid adaptability and generalization ability of cross-regional prediction. SUMMARY

[0005] The purpose of the present application is to provide a photovoltaic power station output power prediction method, terminal and medium, which realizes the real-time, dynamic and cross-regional adaptability of photovoltaic power station power prediction, can effectively cope with the challenges brought by different meteorological conditions and regional differences, and significantly improves the accuracy and reliability of photovoltaic power prediction. Especially in the face of sudden weather changes, regional climate differences and the application scenarios of newly built photovoltaic stations, the present application provides strong adaptability and generalization ability, providing more accurate decision support for intelligent scheduling of photovoltaic power generation and power grid management.

[0006] To achieve the above purpose, in the first aspect of the present application, a photovoltaic power station output power prediction method is provided, the method comprising:

[0007] S1, collecting real-time meteorological data and preprocessing;

[0008] S2, constructing a self-organizing mapping network, taking the preprocessed data as input, identifying and classifying different environmental patterns under unsupervised conditions, and simultaneously predicting photovoltaic power generation power;

[0009] S3, designing a regional model of a cross-regional knowledge transfer algorithm based on a meta-learning model, and transferring knowledge between different regions;

[0010] S4, predicting the category of the current meteorological environment according to newly collected real-time meteorological data, then calculating the prediction error of each regional model under the new environmental data, and selecting the regional model with the smallest error for power prediction;

[0011] S5, constructing a personalized recommendation strategy model according to the predicted power to provide real-time feedback and optimization for power prediction.

[0012] Further, the self-organizing mapping network first combines the pre-processed data with the environment model for identification training, then matches the environment data with the time series data, and finally predicts the photovoltaic power generation to capture the photovoltaic power generation law under different environment modes.

[0013] Further, the self-organizing mapping network first combines the pre-processed data with the environment model for identification training, and specifically includes:

[0014] Suppose the grid size of the self-organizing mapping network is M x M, each node w j , j ∈ [1, M 2 ], each node w j represents a neuron, w j ∈ R d , the weight vector of each neuron w j is the same as the dimension of the meteorological data x′ i , which is used to represent the meteorological mode of the region;

[0015] For each input data x′ i , calculate the Euclidean distance between it and all neuron weight vectors:

[0016]

[0017] Then select the neuron w BMU with the smallest distance, i.e. the best matching unit:

[0018]

[0019] Through the mechanism of competitive learning, the neuron weights in the neighborhood of the best matching unit are updated to move closer to the input data x′ i , and the specific update formula is:

[0020]

[0021] where w j (t) represents the weight of neuron w j at time t, η(t) is the learning rate, h BMU,j (t) is the influence function of the BMU neighborhood;

[0022] At the same time in the training process, the self-organizing mapping network maps the meteorological mode of the input data set x onto a two-dimensional grid, where each neuron w j corresponding region is regarded as an environment mode region;

[0023] After the training is completed, the mode recognition is performed on the entire grid, the regions with similar neuron weights are clustered into a class, and different meteorological mode regions {C1, C2,..., CM} are formed.K}, each region represents a weather pattern category;

[0024] For the time sequence of weather data, a time regularization term λ t is introduced to distinguish weather data at different times by weighting:

[0025]

[0026] where λ t represents the time weight, used to adjust the influence of weather data at different times on the clustering results, thus solving the problem of changes in weather patterns over time.

[0027] After training, the self-organizing mapping network maps the weather data onto a two-dimensional grid to generate the clustering results of environmental pattern recognition {C1, C2, …, C K} represents different pattern regions in the weather data, and the center point w k of each category C k is the typical weather state of the current pattern.

[0028] Further, the environmental data is matched with the time sequence data, and finally the photovoltaic power is predicted to capture the photovoltaic power law under different environmental patterns, which specifically includes:

[0029] For new weather data x′ i , calculate its Euclidean distance with the center point w k of each environmental pattern category:

[0030] D(x′ i , C k ) = ||x′ i -w k ||

[0031] Then select the category with the smallest distance as the matching pattern of the current weather data:

[0032]

[0033] After environmental pattern recognition and matching, the photovoltaic power is predicted in combination with the environmental category C :

[0034] A power prediction model based on the environmental category is designed The power prediction model performs regression analysis on the environmental category:

[0035] Assuming that for each environmental category C k , a power prediction model is designed, which is based on historical data P historical ​​Training, assuming P historical For the photovoltaic power sequence under the environmental mode C k The goal of training the model is:

[0036]

[0037] Finally, based on meteorological pattern recognition and matching, the output power of the photovoltaic power generation system under the current meteorological condition is predicted

[0038] Further, the S3 specifically comprises:

[0039] The meta-learning model training target based on irrelevant meta-learning is designed, photovoltaic power generation prediction is carried out under different meteorological modes, and the meta-learning model can be optimized on historical data by minimizing the loss function, and a regional weighting mechanism is introduced to enable the meta-learning model to transfer knowledge and adapt to changes in different meteorological regions.

[0040] After the training phase is completed, the meta-learning model is quickly adapted through irrelevant meta-learning, and the meta-learning model after quick adaptation will be used for photovoltaic power generation prediction under new meteorological modes, and finally the predicted power generation power of the region is output.

[0041] Further, the meta-learning model training target is represented as follows:

[0042]

[0043] Wherein, is the meta-learning model loss function, x′ i is the input feature of the region C k , P i is the actual power generation power in the region, is the predicted power generation power, D k is the training data set of the region C k , and |D k | represents the number of training samples;

[0044] The regional weighting mechanism specifically comprises:

[0045] For each region C k , a weighted loss function

[0046]

[0047] Wherein, is the weighting factor of the region C k , reflecting the influence degree of the region on the model training; the weighting factor is calculated as follows:

[0048]

[0049] where |D k is the data volume of region C k , and |D is the sum of data volumes of all regions.

[0050] After the training phase is completed, the meta-learning model is quickly adapted by irrelevant meta-learning, and the meta-learning model after quick adaptation will be used for photovoltaic power prediction under a new meteorological mode, and finally the predicted power of the region is output, which specifically comprises:

[0051] In the test phase of a new region C new , first, the parameters θ' optimized by irrelevant meta-learning are used for prediction, and according to the irrelevant meta-learning algorithm, the quick adaptation process is as follows:

[0052]

[0053] where β is the adaptation learning rate, is the test loss on region C new .

[0054] is the loss function A regularization term is introduced to suppress the overfitting of the model under small sample data:

[0055]

[0056] where λ reg is the regularization coefficient, is the L2 norm of the gradient of the meta-learning model, which can effectively control the complexity of the model and prevent overfitting in new small sample regions.

[0057] Further, the S4 specifically comprises:

[0058] The new meteorological data is input into the environmental pattern classification network to identify the category of the current meteorological environment, and the output is the predicted environmental pattern label. The environmental pattern classification network is optimized using cross-entropy loss;

[0059] For each region model, calculate its prediction error under new environmental data, and select the model with the smallest error for power prediction, which is represented as follows:

[0060]

[0061] Select the model with the smallest error:

[0062]

[0063] Introducing an L2 regularization term avoids overfitting and ensures the model's generalization ability:

[0064]

[0065] Where λ is the regularization coefficient. It is the L2 norm of the model parameters;

[0066] If the prediction error of the selected regional model exceeds a preset threshold, the model parameters are adjusted using an optimization algorithm to better fit the new environmental data; the optimization algorithm is as follows:

[0067]

[0068] Where γ is the learning rate. It is the loss function with respect to the parameters gradient, These are the updated parameters;

[0069] Power prediction is performed on new environmental data by selecting and adjusting the meta-learning model.

[0070] Furthermore, S5 specifically includes:

[0071] A personalized recommendation strategy model is designed based on the user's personalized demand vector and the global state output by the meta-learning model. Then, a deep neural network is used to optimize the personalized recommendation strategy model. The user's personalized demand vector includes the user's electricity consumption behavior characteristics and the user's demand for dispatch response. The user's personalized demand vector is represented as follows:

[0072]

[0073] in, It is the power load of user i at time t; This refers to the time delay in the user's response to the scheduling request; It represents the user's maximum response capability at time t; This is the user's preferred scheduling period, representing the time window they prioritize.

[0074] The output of the personalized recommendation strategy model is:

[0075]

[0076] Among them, f θ It is a deep neural network model obtained through training, where θ represents the weights of the network. It is a personalized scheduling strategy, representing the power dispatch instructions that user i needs to execute at time t;

[0077] The comprehensive utility function is designed to quantify the scheduling decision at a specific time, and a dynamic adjustment mechanism is designed, so that when the global state or the user's personalized demand vector changes, the personalized recommendation strategy will be updated according to real-time feedback; the comprehensive utility function is represented as follows:

[0078]

[0079] wherein, is the load satisfaction degree of the user i at the time t, which is usually measured by the difference between the actual power supply and the user demand; is the response efficiency, which considers the speed of the user's response to the scheduling within a specified time; is the power cost saved by the user, which measures the influence of the scheduling strategy on the economic benefit of the user; γ1, γ2 and γ3 represent the corresponding weights;

[0080] When the global state or the user's personalized demand vector changes, the personalized recommendation strategy will be updated according to real-time feedback. After each strategy execution, the effect of the current strategy is evaluated by calculating the real-time reward , which is represented as:

[0081]

[0082] wherein, and are the user load satisfaction degree and the saved cost mentioned before; α1, α2 are adjustment factors, which control the priority of different effects;

[0083] A weighted fusion mechanism is designed to synthesize the global state a t and the personalized strategy to generate the final scheduling strategy , which is represented as:

[0084]

[0085] wherein, λ1 and λ2 are weighting coefficients;

[0086] The long-term benefit of the personalized strategy is evaluated, and the strategy is updated according to the user's continuous feedback. At the same time, whenever a new load mode, user demand change or unpredictable emergency occurs, the final scheduling strategy needs to be adjusted in time. The long-term benefit of the personalized strategy is evaluated by periodically calculating the average value of the user utility function to evaluate the long-term benefit of the strategy:

[0087]

[0088] wherein, T is the evaluation period length, is the user utility value at each time t.

[0089] The final scheduling strategy needs to be adjusted in time when encountering new load patterns, user demand changes or unforeseen emergencies, and an adaptive regularization term is introduced Ensure efficient updating of personalized recommendation models under real-time feedback:

[0090]

[0091] wherein, is a regularization term, controlling the difference between the personalized strategy and the global strategy; β1, β2 are regularization factors, controlling the weights of the two regularization terms; is the update amount of the final scheduling strategy.

[0092] According to a second aspect of the present application, a terminal is provided, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor being configured to execute the program to perform the method of any one of the above.

[0093] According to a third aspect of the present application, a computer readable storage medium is provided, having a computer program stored thereon, the program being executable by a processor to perform the method of any one of the above.

[0094] The present application has at least the following beneficial technical effects:

[0095] (1) The present application uses a self-organizing mapping network (SOM) to cluster and identify weather data, classifies photovoltaic power generation behavior under different weather conditions into multiple modes (such as sunny, cloudy, overcast, storm, etc.), and automatically identifies the current environmental mode according to real-time environmental data. Under each environmental mode, the power variation law of photovoltaic power generation has significant differences, so the present application dynamically selects the power prediction model most suitable for the current environmental conditions through the environmental mode switching mechanism, thereby achieving adaptive dynamic adjustment when facing rapidly changing weather. This method overcomes the problem of slow response to environmental changes in traditional methods, improving the real-time and accuracy of prediction.

[0096] (2) The present application combines meta-learning technology and can solve the data scarcity and knowledge transfer problem in cross-regional photovoltaic power prediction. By jointly training photovoltaic data from different regions, the meta-learning framework is used to learn the common features between regions, thereby realizing the rapid transfer and adjustment of the cross-regional photovoltaic power prediction model. In newly deployed photovoltaic power stations, the system can quickly adjust the existing model parameters with a small amount of local data to adapt to new climate and environmental conditions, significantly improving the cross-regional adaptability of the model. This method solves the problem of insufficient prediction accuracy of traditional models in cross-regional applications, and is particularly suitable for the data scarcity problem of newly built sites.

[0097] (3) The application integrates multiple photovoltaic power prediction models for different environmental patterns and combines real-time environmental pattern recognition to propose a dynamic model switching mechanism. When changes in meteorological conditions are detected, the system can quickly switch to the model most suitable for the current environmental pattern for power prediction, thereby achieving flexible adaptation to sudden weather events (such as heavy rain, strong light, etc.) and long-term seasonal changes. This overcomes the shortcomings of fixed models and lack of flexibility in existing methods, improving the robustness and accuracy of the system in complex environments.

[0098] (4) The application realizes real-time, dynamic and cross-regional adaptability of photovoltaic power station power prediction, which can effectively cope with the challenges brought by different meteorological conditions and regional differences, significantly improving the accuracy and reliability of photovoltaic power prediction. Especially in the face of sudden weather changes, regional climate differences and the application scenarios of newly built photovoltaic stations, the application provides strong adaptability and generalization ability, providing more accurate decision support for intelligent scheduling of photovoltaic power generation and power grid management. BRIEF DESCRIPTION OF DRAWINGS

[0099] The application is further illustrated by the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the application. For ordinary skilled in the art, other drawings can be obtained without creative labor on the basis of the following drawings.

[0100] Figure 1 A photovoltaic power station output power prediction method flowchart is provided for the embodiments of the application. DETAILED DESCRIPTION

[0101] The embodiments of the application will be described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference signs represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary and are only used to explain the application, and cannot be understood as a limitation on the application.

[0102] As shown in Figure 1 A photovoltaic power station output power prediction method provided by the embodiments of the application includes the following steps S1-S5:

[0103] S1, collect real-time meteorological data and perform preprocessing.

[0104] Specifically, real-time meteorological data is collected from multiple sensors and meteorological service platforms (such as light, temperature, humidity, wind speed, etc.). Each meteorological data source may have a time delay or different data quality, and the reliability information of the data source needs to be identified and recorded in the collection stage.

[0105] Furthermore, noise removal and missing value handling:

[0106] The Local Weighted Regression (LWR) algorithm is applied to remove random noise from the data. Specifically, assume the original meteorological dataset is D = {d1, d2, ..., d...} n}, where each d i This indicates at time point t i Meteorological data collected above. The LWR-smoothed data is D′={d1',d2',...,d n The formula is as follows:

[0107]

[0108] Among them, W(t) i ,t j () is based on time point t i and t j The weighted function of the distance between them, d j For time point t j Meteorological data.

[0109] A K-Nearest Neighbors (KNN)-based interpolation method is used to fill in missing values. For a missing value at position i, its distance d to other known values ​​is calculated. i Select the K nearest neighbor data points {d j Then, the missing values ​​are filled using the weighted average method, as shown in the following formula:

[0110]

[0111] Where, ω j The weight of data point j is usually inversely proportional to the distance between data points.

[0112] Furthermore, to ensure consistent scaling across different meteorological data, Z-score standardization was applied to all meteorological data. Assume each meteorological data point is x. i If the mean and standard deviation are μ and σ respectively, then the standardized data x' i for:

[0113]

[0114] Where x' is the standardized meteorological data, μ is the mean of the data, and σ is the standard deviation of the data.

[0115] Furthermore, the preprocessed data is converted into a unified format, and the output is a multidimensional matrix X = {x1, x2, ..., x...}. n}, where each x irepresenting the normalized weather features at each time point.

[0116] Meanwhile, in order to support the subsequent environmental pattern recognition step (such as SOM), the weather data and time stamp of each time point are combined into a time series dataset for pattern recognition of different time periods.

[0117] Further, the normalized and denoised weather dataset X is in the form of a matrix, each sample containing a time stamp, weather parameters and related identification information.

[0118] S2, construct a self-organizing mapping network, take the preprocessed data as input, identify and classify different environmental patterns under unsupervised conditions, and predict photovoltaic power generation.

[0119] Specifically, in step 1, the input weather dataset X = {x'1, x'2,..., x'N} has been processed, which includes data normalization and noise removal. Next, the weather data is subjected to environmental pattern recognition and clustering through a self-organizing mapping network (SOM). As an unsupervised learning model, SOM can effectively extract potential patterns from complex weather data and map them to a two-dimensional grid, thereby completing the spatial division of weather patterns. n} has been processed, which includes data normalization and noise removal. Next, the weather data is subjected to environmental pattern recognition and clustering through a self-organizing mapping network (SOM). As an unsupervised learning model, SOM can effectively extract potential patterns from complex weather data and map them to a two-dimensional grid, thereby completing the spatial division of weather patterns.

[0120] Further, the preprocessed weather dataset X, each x' i represents a set of weather features at a time point, including measured values of factors such as light intensity, temperature, humidity, wind speed, etc. The data dimension is d.

[0121] Further, SOM training and environmental pattern recognition: assuming the grid size of the SOM model is M x M, each node w j (j∈[1,M 2 ]) represents a neuron, w j ∈R d . The weight vector of each neuron w j is the same dimension as the weather data x' i , used to represent the weather pattern in that area.

[0122] For each input data x' i , the invention calculates the Euclidean distance between it and all neuron weight vectors:

[0123]

[0124] Then select the neuron w BMU with the smallest distance, i.e. the Best Matching Unit (BMU):

[0125]

[0126] Through the mechanism of competitive learning, the neuron weights of the BMU neighborhood are updated to make them close to the direction of the input data x i . The specific update formula is:

[0127] w j (t+1)=w j (t)+η(t)·h BMU,j (t)·(x i -w j (t))

[0128] where w j (t) represents the weight of neuron w j at time t, η(t) is the learning rate, h BMU,j (t) is the influence function of the BMU neighborhood, which is usually defined as a Gaussian function:

[0129]

[0130] where r j is the two-dimensional grid position of neuron w j , and σ(t) is the neighborhood function width that decays with the training process.

[0131] Further, during the training process, the SOM maps the meteorological patterns of the input data set X onto a two-dimensional grid. Each neuron w j corresponds to a region that can be regarded as an environmental pattern region. After the training is completed, the present application performs pattern recognition on the entire grid, and clusters regions with similar neuron weights of adjacent neurons into a class to form different meteorological pattern regions {C1, C2,..., C K}, each of which represents a meteorological pattern category.

[0132] Due to the strong spatiotemporal correlation of meteorological patterns, some special regularization terms are needed to avoid overfitting and over-segmentation during the clustering process. In order to enhance the sensitivity to time series data, a time regularization term λ t is introduced for the time series nature of meteorological data to distinguish meteorological data at different times by weighting:

[0133]

[0134] where λ t represents the time weight, which is used to adjust the influence of meteorological data at different times on the clustering result, thereby solving the problem of changes in meteorological patterns over time.

[0135] Further, SOM is trained to map the weather data onto a two-dimensional grid, generating clustering results of environmental pattern recognition. These clustering results {C1, C2,..., C K} represent different pattern categories in the weather data, and the center point w k of each category C k is the typical weather state of the pattern.

[0136] Further, in step 1, the environmental patterns of the weather data are extracted and clustered by the SOM network. Next, support is provided for photovoltaic power prediction by matching the current weather data with these environmental patterns. Assuming that the new weather input is x' i , the present invention first calculates its similarity with each environmental pattern category C k , and selects the most similar pattern for prediction.

[0137] Further, the environmental matching formula: for the new weather data x', the Euclidean distance between it and the center point w k of each environmental pattern category is calculated:

[0138] D(x', C k ) = ||x' i -w k ||.

[0139] Then the category with the smallest distance is selected as the matching pattern of the current weather data:

[0140]

[0141] Further, the photovoltaic power prediction model: after environmental pattern recognition and matching, the environmental category obtained in the previous steps is combined to further predict the photovoltaic power. In order to capture the photovoltaic power law under different environmental patterns, a power prediction model based on environmental categories is designed This model combines environmental categories with weather factors such as sunlight and temperature for regression analysis.

[0142] Further, assuming that a power prediction model is designed for each environmental category C k , the model is trained based on historical data P historical . Assuming that P historical is the photovoltaic power sequence under the environmental pattern C k , the goal of training the model is:

[0143]

[0144] Furthermore, based on meteorological model identification and matching, the output power of the photovoltaic power generation system under the current meteorological conditions is predicted. This output power value will be used in the power dispatch and energy management system of the photovoltaic power plant to support subsequent load forecasting and grid dispatch.

[0145] S3. Design a regional model based on a meta-learning model for cross-regional knowledge transfer algorithms to perform knowledge transfer between different regions.

[0146] Specifically, the meta-learning model adopts the Model-Agnostic Meta-Learning (MAML) framework, which can be applied to multiple meteorological regions C k Training on the data, after training, it can be used in a new region C new This allows for rapid adaptation, achieving ideal predictive results with only a small amount of sample data. Assume that through step 2 above, different regions C have been obtained. k Clustering results and corresponding historical photovoltaic power generation datasets This invention will train the model on these data and optimize the model parameters θ.

[0147] The training objective of this invention is to train a photovoltaic power generation prediction model. It can predict photovoltaic power generation under different weather patterns and optimize it on historical data by minimizing the loss function. Loss function for:

[0148]

[0149] Where, x' i For region C k Input features (e.g., meteorological data), P i This represents the actual power generation capacity in this region. For the predicted power generation, D k For region C k The training dataset, |D k | indicates the number of training samples.

[0150] Furthermore, in order to enable the model to effectively transfer knowledge and adapt to changes in different meteorological regions, this invention introduces a regional weighting mechanism. This weighting factor The weighting mechanism determines the impact of different regions on the global model training by evaluating the quality of historical data for each region, the similarity between regions, and the region's importance within the overall photovoltaic power generation system. This allows for the reasonable adjustment of the impact of different regions on the global model training. This weighting mechanism helps transfer knowledge from important regions to new regions, improving cross-regional adaptability.

[0151] Weighted loss function: for each region C k In the model training process, the present application introduces a weighted loss function

[0152]

[0153] where, is the weighted factor of region C k , reflecting the influence degree of the region on model training. The factor can be dynamically adjusted based on the historical data volume, meteorological condition stability, prediction accuracy, and other factors of the region.

[0154] Further, the present application defines as the weight of region C k , which can usually be calculated as follows:

[0155]

[0156] where, |D k | is the data volume of region C k , is the sum of data volumes of all regions. This calculation method makes the region with more historical data have greater influence in the training process, helping the model better transfer learned knowledge.

[0157] Further, after the training phase is completed, the model is quickly adapted through the MAML algorithm. Assuming that the new meteorological region C new has new meteorological data x' new , the present application will quickly adjust the model parameters through the meta-learning framework, so that the model can make accurate predictions under C new region. In the test phase of the new region C new , first, the present application uses the optimized parameters θ' of MAML for prediction, but θ' needs to be further adjusted through a small number of test samples. According to the MAML algorithm, the quick adaptation process is as follows:

[0158]

[0159] where, β is the adaptation learning rate, is the test loss on region C new .

[0160] Further, in order to better adapt to the new environment mode and avoid overfitting, the present application introduces a regularization term for the loss function to suppress overfitting of the model under small sample data:

[0161]

[0162] where, λreg is the regularization coefficient, is the L2 norm of the model gradient, which can effectively control the complexity of the model and prevent overfitting in new small sample areas.

[0163] Further, the meta-learning model after rapid adaptation will be used for new meteorological mode C new photovoltaic power prediction, and finally output the predicted power of the region

[0164] Through the above design, the model can perform transfer learning in different meteorological regions, and when facing new meteorological modes, it can quickly adjust the model parameters, so that the photovoltaic power prediction model can effectively adapt to new environmental conditions and provide more accurate predictions.

[0165] This step mainly realizes knowledge transfer and rapid adaptation between different meteorological regions by combining the meta-learning (MAML) framework, regional weighting mechanism and specific regularization term design. This scheme can help the photovoltaic power prediction model to learn from historical data and adapt to new meteorological modes, improve the accuracy of cross-region prediction, and optimize the scheduling and management of photovoltaic power generation systems.

[0166] S4, according to the real-time re-collected new meteorological data, predict the category of the current meteorological environment, then for each regional model, calculate its prediction error under the new environment data, and select the regional model with the smallest error for power prediction.

[0167] Specifically, collect new environmental data: x new = {I new , T new , H new , …}, including radiation I new , temperature T new , humidity H new , etc., which reflect the current meteorological and environmental mode.

[0168] Further, input the environmental data x new into the environmental mode classification network The network is used to identify the category of the current meteorological environment, and the output is the predicted environmental mode label

[0169] Among them, the classification network output:

[0170]

[0171] Among them is the environmental mode prediction of the environmental data x new , and φ is the parameter of the network.

[0172] Further, the environment pattern classification network is optimized with cross-entropy loss

[0173]

[0174] where II is an indicator function, y i is the real environment pattern label, is the predicted label, P(y i | x new ) is the predicted probability of the label.

[0175] Further, the selection criterion is: for each regional model , calculate its prediction error under the new environment data x new , and select the model with the smallest error for power prediction.

[0176] where the error of each regional model is:

[0177]

[0178] Select the model with the smallest error:

[0179]

[0180] Further, introduce the L2 regularization term to avoid overfitting and ensure the generalization ability of the model:

[0181]

[0182] where λ is the regularization coefficient, is the L2 norm of the model parameters.

[0183] Further, parameter fine-tuning:

[0184] If the prediction error of the selected regional model exceeds the preset threshold ∈ threshold , adjust the parameters of the model through the optimization algorithm to better fit the new environment data.

[0185] where the parameter update formula is:

[0186]

[0187] where γ is the learning rate, is the gradient of the loss function with respect to the parameter , and is the updated parameter.

[0188] Further, power prediction:

[0189] through the selected and adjusted model For environmental data x new Perform power prediction:

[0190]

[0191] in, This indicates the updated region model. For new environmental data x new Power prediction was performed.

[0192] This step, through dynamically selecting the most suitable regional model and adaptive adjustment mechanism, can effectively improve the accuracy of photovoltaic power prediction in complex meteorological changes, and ensure the stability and adaptability of the system under different environmental modes.

[0193] S5. Based on the predicted power, a personalized recommendation strategy model is constructed to provide real-time feedback and optimization for power prediction.

[0194] Specifically, in this stage, the input is the scheduling strategy optimization result a from the previous stage. t System status s t and real-time feedback data y t Using this data, personalized needs are identified and modeled for each user or user group, further refining user requirements. Historical power load, response patterns, and preferences data are combined with time series analysis to construct personalized demand vectors. This demand vector includes users' electricity consumption behavior characteristics and their needs for dispatch response:

[0195]

[0196] in, It is the power load of user i at time t. This is the time delay in the user's response to the scheduling request. It represents the user's maximum response capability at time t. It represents the user's preferred scheduling period, indicating the time window they prioritize.

[0197] Furthermore, user groups can be segmented using clustering algorithms (such as K-means or the density-based clustering algorithm DBSCAN) to form different demand groups. These groups can help to perform more accurate personalized scheduling.

[0198] Furthermore, in personalized recommendation strategy models, the core objective is to determine the user's personalized needs based on their individual needs. and system global state s t Output the user's personalized scheduling strategy This enables dynamic optimization of the power system. Here, the invention employs a recommendation system model f based on a deep neural network. θ Optimize decision-making.

[0199] Furthermore, the recommendation strategy model generates personalized scheduling decisions based on user demand vectors and system states using a deep learning model:

[0200]

[0201] Among them, f θ It is a deep neural network model obtained through training, where θ represents the weights of the network. It is a personalized scheduling strategy, representing the power scheduling instructions that user i needs to execute at time t.

[0202] Furthermore, user utility function This utility function is used to quantify user satisfaction under specific scheduling decisions. It considers not only the satisfaction of users' electricity demand but also response efficiency and related cost savings. Therefore, this invention designs a comprehensive utility function that includes multiple aspects:

[0203]

[0204] in, It represents the load satisfaction of user i at time t, which is usually measured by the difference between the actual power supply and the user's demand. It refers to response efficiency, taking into account how quickly users can respond to scheduling within a specified time. This represents the electricity cost savings for users, measuring the impact of dispatch strategies on user economic benefits. The weighting coefficients γ1, γ2, and γ3 in the user utility function reflect the relative importance of different factors to the user and depend on the user's individual needs.

[0205] Furthermore, to respond to real-time changes in the system, this invention designs a dynamic adjustment mechanism. Whenever the system state s... t Or user feedback y t Personalized recommendation strategy when changes occur It will be updated based on real-time feedback. After each strategy execution, the effectiveness of the current strategy is evaluated by calculating the real-time reward.

[0206]

[0207] in, and Similarly, the previously mentioned user load satisfaction and cost savings; α1 and α2 are adjustment factors that control the priority of different utilities. Based on the current reward The personalized recommendation model f is trained using reinforcement learning or incremental training methods.θ Adjustment is made to update the model parameters θ to better adapt to real-time feedback.

[0208] Further, while achieving personalized scheduling, it is necessary to combine personalized strategies with global scheduling strategies to ensure the coordination and optimization of the entire power system. Therefore, the present application designs a weighted fusion mechanism to synthesize the global strategy a t and the personalized strategy to generate the final scheduling strategy

[0209] By introducing the weighting coefficients λ1 and λ2, the global strategy and the personalized strategy are weighted and fused according to different priorities:

[0210]

[0211] where a t is the global scheduling strategy, ensuring the optimal scheduling of the power system at the global level; is the personalized scheduling strategy, which optimizes for the needs of specific users or user groups. The weighting coefficients λ1 and λ2 can be dynamically adjusted according to system load, user demand changes, and real-time feedback, ensuring the flexibility and adaptability of the system.

[0212] Further, long-term personalized scheduling strategies need to be adjusted and optimized through periodic effectiveness evaluation. At the end of each period, the present application evaluates the long-term benefits of the personalized strategy and updates the strategy according to the user's continuous feedback.

[0213] The long-term benefits of the strategy are evaluated by periodically calculating the average value of the user utility function :

[0214]

[0215] where T is the evaluation period length, is the user utility value at each time t.

[0216] Further, according to the long-term feedback results, the parameters θ of the personalized recommendation model are further adjusted through optimization algorithms (such as genetic algorithm or simulated annealing) to improve the overall benefits.

[0217] Further, as the power scheduling system continues to evolve, the diversity and real-time nature of user demand require the system to be continuously optimized. Whenever the system encounters new load patterns, changes in user demand, or unforeseen emergencies, the personalized recommendation system needs to be adjusted in a timely manner.

[0218] Further, an adaptive regularization term is introduced to ensure efficient updating of the personalized recommendation model under real-time feedback:

[0219]

[0220] wherein, is a regularization term, controlling the difference between the personalized policy and the global policy; β1, β2 are regularization factors, controlling the weights of the two regularization terms.

[0221] Another embodiment of the present application provides a terminal, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, the processor being configured to execute the method of any one of the above embodiments of the present application when executing the program.

[0222] Optionally, the memory is configured to store the program; the memory can comprise volatile memory (e.g., random-access memory (RAM), such as static random-access memory (SRAM), Double Data Rate synchronous dynamic random-access memory (DDR SDRAM), etc.), and / or non-volatile memory (e.g., flash memory). The memory is configured to store computer programs (e.g., application programs, functional modules, etc. for implementing the above method), computer instructions, etc., which can be stored in one or more memories in partitions. The computer programs, computer instructions, etc. can be invoked by the processor.

[0223] The computer programs, computer instructions, etc. can be stored in one or more memories in partitions. The computer programs, computer instructions, data, etc. can be invoked by the processor.

[0224] The processor is configured to execute the computer program stored in the memory to implement each step in the method according to the above embodiments. For details, refer to the related description in the method embodiments above.

[0225] The processor and the memory can be independent structures or integrated structures. When the processor and the memory are independent structures, the memory and the processor can be coupled and connected through a bus.

[0226] Another embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, the program being executable by a processor to implement the method of any one of the above embodiments of the present application.

[0227] The above embodiments of the present application provide a photovoltaic power station output power prediction method, the method comprising: collecting real-time meteorological data and preprocessing; constructing a self-organizing mapping network, taking the preprocessed data as input, identifying and classifying different environmental patterns under unsupervised conditions, while predicting photovoltaic power generation power; design a regional model of cross-regional knowledge transfer algorithm based on meta-learning model, transfer knowledge between different regions; according to the newly collected real-time meteorological data, predict the category of the current meteorological environment, then for each regional model, calculate its prediction error under new environmental data, and select the regional model with the smallest error for power prediction; according to the predicted power, construct a personalized recommendation strategy model to realize real-time feedback and optimization of power prediction. The present application realizes the real-time, dynamic and cross-regional adaptability of photovoltaic power station power prediction, can effectively cope with the challenges brought by different meteorological conditions and regional differences, and significantly improves the accuracy and reliability of photovoltaic power prediction. Especially in the face of sudden weather changes, regional climate differences and the application scenarios of newly built photovoltaic stations, the present application provides strong adaptability and generalization ability, and provides more accurate decision support for intelligent scheduling of photovoltaic power generation and power grid management.

[0228] The specific embodiments of the present application are described above. It should be understood that the present application is not limited to the above specific embodiments, and those skilled in the art can make various modifications or changes within the scope of the claims, which does not affect the essential content of the present application.

Claims

1. A method of photovoltaic power plant output power prediction, characterized in that, The method comprises: S1, collecting real-time meteorological data and preprocessing; S2, constructing a self-organizing mapping network, taking the preprocessed data as input, identifying and classifying different environmental patterns under unsupervised conditions, and predicting photovoltaic power generation power; S3, designing a regional model of a cross-regional knowledge transfer algorithm based on a meta-learning model, and transferring knowledge between different regions; S4, predicting the category of the current environmental pattern according to the newly collected real-time meteorological data, then calculating the prediction error of each regional model under the new meteorological data, and selecting the regional model with the smallest error for power prediction; S5, constructing a personalized recommendation strategy model according to the predicted power for real-time feedback and optimization of power prediction, specifically including: Design a personalized recommendation strategy model according to the user personalized demand vector and the global state output by the meta-learning model, and then optimize the personalized recommendation strategy model using a deep neural network; wherein the user personalized demand vector includes the user's power consumption behavior characteristics and the user's demand for dispatching response; wherein the user personalized demand vector is represented as: wherein, is the power load of user i at time instant t; is the time delay of user i in responding to the dispatch request at time instant t; is the maximum response capability of user i at time instant t; is the preferred dispatch period of user i, representing its prioritized time window; The output of the personalized recommendation strategy model is: wherein f θ is a deep neural network model obtained by training, and θ represents the weights of the network; is a personalized strategy, representing the power dispatch instruction that user i needs to perform at time t; s t is the global state of the system; Design a comprehensive utility function to quantify a specific dispatching decision, and design a dynamic adjustment mechanism, which updates the personalized strategy according to real-time feedback when the global state or user personalized demand vector changes; the comprehensive utility function is represented as follows: wherein, is the response efficiency of user i, considering the speed of user's response to the dispatch within a specified time; γ1, γ2 and γ3 represent the corresponding weights; is the comprehensive utility function value of each time t; is the load satisfaction degree of user i at time t, which is usually measured by the difference between the actual power supply and the user demand; is the power saving cost of user i at time t, which measures the influence of the dispatch strategy on the economic benefit of the user; The personalized strategy is updated according to real-time feedback when the global state or user personalized demand vector changes, and the effect of the current strategy is evaluated after each strategy execution by calculating the real-time reward , which is represented as: Wherein, α1, α2 are adjustment factors, controlling the priority of different utilities; A weighted fusion mechanism is designed to combine the global policy a t with the personalized policy to generate the final scheduling policy is represented as: Wherein, λ1 and λ2 are weighting coefficients; The long-term benefit of the final scheduling strategy is evaluated, and the strategy is updated according to the continuous feedback of the user, and the final scheduling strategy needs to be adjusted in time whenever a new load mode, user demand change or unpredictable emergency is encountered; wherein the long-term benefit is evaluated by periodically calculating the average value of the comprehensive utility function value to evaluate the long-term benefit of the strategy: where T is the length of the evaluation period, is the value of the overall utility function at each time instant t.

2. A method of predicting the output power of a photovoltaic power plant according to claim 1, characterized in that, The self-organizing mapping network first identifies and trains the preprocessed data and the environmental model, then matches the meteorological data and the time series data, and finally predicts the photovoltaic power generation power to capture the photovoltaic power generation law under different environmental patterns.

3. A method of predicting the output power of a photovoltaic power plant according to claim 2, characterized in that, The self-organizing mapping network first identifies and trains the preprocessed data and the environmental model, specifically including: Assume the self-organizing map network has a grid size of M×M, and each node w j j∈[1,M 2 ], each node w j Representing a neuron, w j ∈R d The weight vector of each neuron is related to the meteorological data x′. i The dimensions are the same, used to characterize the environmental patterns of the region; For each input weather data x' i , the Euclidean distance between it and all neuron weight vectors is calculated: The neuron w with the smallest distance is then selected BMU i.e. the best matching unit: Through the mechanism of competitive learning, update the neuron weights of the neighborhood of the best matching unit to make them approach the direction of the input meteorological data x'i, and the specific update formula is: where w j (t) represents the neuron w j The weight at time t, η(t) is the learning rate, h BMU,j (t) is the best matching unit w BMU The influence function of the neighborhood; At the same time in the training process, self-organizing mapping network maps the environmental pattern of input data set x onto a two-dimensional grid, wherein each neuron w j The corresponding area is regarded as an environmental pattern area; After the training is completed, pattern recognition is performed on the entire grid, and regions with similar weights of adjacent neurons are clustered into a class to form different environmental pattern regions {C1, C2,..., C K}, each of which represents an environmental pattern category; For the time series of meteorological data, distinguish the meteorological data at different times by weighting: where λ t represents the time weight, used to adjust the influence degree of meteorological data in different time periods on the clustering results, so as to solve the problem of time-varying environment mode; After training, the self-organizing map network maps the weather data onto a two-dimensional grid, generating clustering results {C1, C2,..., C K} representing different pattern regions in the weather data, each region C K having a center point w k representing a typical weather state of the current pattern.

4. A method of predicting the output power of a photovoltaic power plant according to claim 3, characterized in that, The meteorological data and the time series data are matched, and finally the photovoltaic power generation power is predicted to capture the photovoltaic power generation law under different environmental patterns, specifically including: For new weather data x'j, compute its Euclidean distance to each environmental pattern region center point w k : D(x', y' ) = ||x' - w i ,C K ) = ||x' - w i -w k || The region with the smallest distance is then selected As a matching pattern for current weather data: After the environmental pattern recognition and matching, the region is combined Prediction of photovoltaic power generation Design a regional power prediction model f, which performs regression analysis on the region: Assume for each region Design a power prediction model f that is based on historical data P historical Train, assuming P historical For the region The goal of the model is to: Finally, based on the recognition and matching of the environmental patterns, the output power of the photovoltaic power generation system under the current meteorological conditions is predicted 5. The method of claim 1, wherein, S3 specifically includes: Design a meta-learning model training target based on irrelevant meta-learning to predict photovoltaic power generation under different environmental patterns, and can be optimized on historical data by minimizing the loss function, while introducing a regional weighting mechanism to enable the meta-learning model to transfer knowledge and adapt to changes in different meteorological regions; After the training phase is completed, the meta-learning model is quickly adapted through irrelevant meta-learning, and the meta-learning model after quick adaptation is used for photovoltaic power generation power prediction in a new environmental pattern, and finally the predicted power generation power of the region is output.

6. A method of predicting the output power of a photovoltaic power plant according to claim 5, characterized in that, The meta-learning model training target is represented as follows: wherein, is a meta-learning model loss function, x′ i is weather data for region C K , P i is actual power generation under the current region, is predicted power generation for region C K , D K is a training data set for region C K ; The regional weighting mechanism specifically includes: For each region C K In the meta-learning model training process, a weighted loss function is introduced wherein, is a weighting factor for region C K reflecting the influence degree of the region on the model training; the weighting factor is calculated as follows: where |D K is the data volume of region C K is the data volume of region C is the sum of all region data volumes; After the training phase is completed, the meta-learning model is quickly adapted through irrelevant meta-learning, the meta-learning model after quick adaptation is used for photovoltaic power generation power prediction in a new environment mode, and finally the predicted power generation power of the region is output, and specifically includes: In the new region C new of the test phase, first use the irrelevant meta-learning optimized parameters θ' to make predictions, according to the irrelevant meta-learning algorithm, the process of rapid adaptation is as follows: where β is an adaptive learning rate, is the test loss on region C new . For the loss function A regularizer term is introduced To suppress overfitting of the model on small sample data: wherein, λ reg is a regularization coefficient, is the L2 norm of the meta-learning model gradient, which can effectively control the complexity of the model and prevent overfitting in new small sample areas.

7. A method of predicting the output power of a photovoltaic power plant according to claim 6, characterized in that, The S4 specifically includes: The new meteorological data is input into the environment mode classification network to identify the category of the current environment mode, and the output is the predicted environment mode label, and the environment mode classification network is optimized using cross-entropy loss; For each regional model, calculate the prediction error under the new meteorological data, and select the model with the smallest error for power prediction, which is represented as follows: Select the model with the smallest error: An L2 regularization term is introduced to avoid overfitting and ensure the generalization ability of the model: where λ is a regularization coefficient, is the L2 norm of the model parameters; If the prediction error of the selected regional model exceeds the preset threshold, the parameters of the model are adjusted through an optimization algorithm to better fit the new meteorological data; the optimization algorithm is represented as follows: Where γ is the learning rate. It is the loss function with respect to the parameters gradient, These are the updated parameters; Through the selected and adjusted meta-learning model, the new meteorological data is used for power prediction.

8. The photovoltaic power station output power prediction method according to claim 1, characterized in that, The final scheduling strategy needs to be adjusted in time when encountering new load patterns, user demand changes or unpredictable emergencies, and an adaptive regularization term is introduced Ensure efficient updating of personalized recommendation strategy model under real-time feedback: wherein, is a regularization term, controlling the difference between the final scheduling policy and the global policy; β1, β2 are regularization factors, controlling the weights of the two regularization terms; is the update amount of the final scheduling policy.

9. A terminal comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program and can be used to execute the method of any one of claims 1-8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor and can be used to execute the method of any one of claims 1-8.

Images