A method of intelligent photovoltaic power generation prediction and self-learning based on deep learning

Abstract

The invention discloses a method for intelligent photovoltaic power generation prediction and self-learning based on deep learning, which belongs to the field of photovoltaic power generation technology, including: collecting multi-source heterogeneous data of photovoltaic power stations and preprocessing them; performing online incremental learning of standardized feature sets through neural networks, and quantitatively evaluating and optimizing the initial power generation prediction results through reward functions; performing knowledge migration on historical photovoltaic power generation data, generating a knowledge migration matrix, and embedding the knowledge migration matrix into a photovoltaic power generation prediction model; dynamically adjusting the optimal confidence threshold of an advanced power generation prediction model through a Bayesian optimization strategy to generate an advanced power generation prediction result. The present invention achieves a comprehensive improvement in data quality, model adaptability, generalization ability and result reliability of photovoltaic power generation prediction through multi-level collaborative innovation of data, models and optimization strategies, effectively meeting the needs of accurate prediction and intelligent management of photovoltaic power generation in complex dynamic environments.

Description

Technical Field

[0001] The present invention relates to the field of photovoltaic power generation technology, and in particular to a method for intelligent photovoltaic power generation prediction and self-learning based on deep learning. Background Art

[0002] With the widespread adoption of photovoltaic power generation technology, accurate prediction of photovoltaic power generation has become a key technology for improving grid dispatch efficiency and the economic efficiency of photovoltaic systems. However, existing photovoltaic power generation prediction methods still have significant shortcomings in complex meteorological conditions, dynamic environmental changes, and long-term prediction accuracy, making it difficult to meet the needs of integrating a high proportion of renewable energy into the grid.

[0003] After searching, the Chinese invention patent with the announcement number CN117220287B discloses a power generation prediction method, with a publication date of February 6, 2024. The patent proposes a power generation prediction method based on PCA dimensionality reduction technology and CNN-NLSTM-Attention model. The original data is reduced in dimensionality through principal component analysis, and time series features are extracted using convolutional neural networks, and historical data are analyzed using long and short-term memory networks. The prediction results are optimized by combining the attention mechanism. However, this technical solution mainly relies on static model training, lacks online incremental learning capabilities, and cannot dynamically adjust model parameters according to real-time operating data, resulting in poor robustness in dealing with dynamic environmental changes such as rapid cloud movement and seasonal radiation fluctuations. In addition, this method does not fully consider the problem of cross-regional and cross-seasonal prediction knowledge transfer, which limits its generalization ability in multiple scenarios.

[0004] After searching, the Chinese invention patent with the announcement number CN117239737B discloses a photovoltaic power generation prediction method and terminal, with a publication date of January 26, 2024. The patent proposes a photovoltaic power generation prediction method based on three independent backbone networks, which respectively models long-term prediction, short-term prediction and sudden weather characteristics, and improves the prediction accuracy by fusing the prediction results. However, although this technical solution has improved the prediction accuracy to a certain extent, its model update still relies on manual intervention, lacks autonomous learning and dynamic optimization mechanisms, and is difficult to adapt to the long-term operation requirements of photovoltaic power generation systems. At the same time, this method has limited ability to identify the risk of prediction failure under extreme weather conditions, and has not introduced an uncertainty quantification module, which may affect the reliability of power grid dispatching decisions.

[0005] The above problems show that the existing photovoltaic power generation prediction methods still have obvious deficiencies in terms of adaptive learning ability, dynamic environmental adaptability, extreme working condition risk identification, and cross-scenario knowledge transfer. Therefore, the present invention provides a method for intelligent photovoltaic power generation prediction and self-learning based on deep learning, which aims to build an online incremental learning framework by integrating multi-source heterogeneous data and deep reinforcement learning technology, and realize cross-regional and cross-seasonal knowledge transfer by combining meta-learning modules, and dynamically adjust the prediction confidence threshold through Bayesian optimization strategy to form a "prediction-correction-feedback" closed-loop control link, thereby significantly improving the timeliness, accuracy and robustness of photovoltaic power generation prediction, and promoting the development of photovoltaic power generation systems towards high-proportion absorption and intelligent operation and maintenance. Summary of the Invention

[0006] In order to solve the technical problems existing in the prior art, the present invention provides a method for intelligent photovoltaic power generation prediction and self-learning based on deep learning. Through multi-level collaborative innovation of data, models and optimization strategies, it achieves a comprehensive improvement in photovoltaic power generation prediction in terms of data quality, model adaptability, generalization ability and result reliability, which can effectively meet the needs of accurate prediction and intelligent management of photovoltaic power generation in complex dynamic environments.

[0007] In a first aspect, the present invention provides a method for intelligent photovoltaic power generation prediction and self-learning based on deep learning, comprising the steps of:

[0008] S1. Collecting multi-source heterogeneous data during the operation of a photovoltaic power station, constructing an initial feature set based on the multi-source heterogeneous data, and preprocessing the initial feature set to generate a standardized feature set; the multi-source heterogeneous data includes meteorological data, real-time irradiation intensity data, and equipment operation status data;

[0009] S2. Performing online incremental learning on the standardized feature set through a neural network to generate an initial power generation prediction result; performing quantitative evaluation and optimization on the initial power generation prediction result through a reward function to generate initial parameters of a power generation prediction model;

[0010] S3. Performing knowledge migration on historical photovoltaic power generation data for multiple regions and seasons to generate a cross-regional and cross-seasonal knowledge migration matrix; embedding the knowledge migration matrix into the photovoltaic power generation prediction model to form an advanced power generation prediction model with generalization capability;

[0011] S4. Dynamically adjust the optimal confidence threshold of the advanced power generation prediction model through a Bayesian optimization strategy to generate an advanced power generation prediction result.

[0012] Preferably, the collecting of multi-source heterogeneous data during the operation of the photovoltaic power station comprises the following steps:

[0013] Deploy multiple sensor nodes to collect meteorological data, real-time radiation intensity data, and equipment operation status data; set the data collection frequency to once per minute;

[0014] The collected multi-source heterogeneous data is stored in a distributed database and sorted by timestamp to generate an initial feature set;

[0015] Preprocessing the initial feature set to generate a standardized feature set;

[0016] The standardized feature set is divided into a training set and a test set, where the training set accounts for 80% and the test set accounts for 20%.

[0017] Preferably, the pretreatment includes:

[0018] Interpolation algorithms are used to fill in missing data;

[0019] The abnormal data is smoothed using a sliding window filtering algorithm.

[0020] Preferably, the method of performing online incremental learning on the standardized feature set through a neural network to generate an initial power generation prediction result; and performing quantitative evaluation and optimization on the initial power generation prediction result through a reward function to generate initial parameters of a power generation prediction model comprises the following steps:

[0021] Initialize the neural network structure in the deep reinforcement learning framework, which includes an input layer, a hidden layer, and an output layer; set the number of hidden layers to 3, and the number of neurons in each layer to 128, 64, and 32 respectively;

[0022] The input layer is used to import the standardized data set into the neural network;

[0023] The hidden layer is used to reduce the dimension of the standardized feature set layer by layer and extract an abstract representation of the standardized feature set;

[0024] The output layer is used to output a preliminary prediction result of power generation according to the abstract representation;

[0025] A reward function is defined, which includes prediction error, model convergence speed, and dynamic environment adaptability; the reward function is used to quantitatively evaluate and optimize the output results of the neural network and update the weight parameters of the neural network, wherein,

[0026] The prediction error is the difference between the initial power generation prediction result and the actual power generation. It also serves as the reward value of the reward function, which is used to drive the prediction model to learn the "environment-power generation" association law in historical data;

[0027] The convergence speed is used to quickly reduce the prediction error during the neural network training process to prevent the training process from falling into a local optimum or taking too long;

[0028] The dynamic environment adaptability requires the prediction model to maintain the stability of prediction error when facing new data, while driving the dynamic pattern of neural network learning generalization;

[0029] At each time step, the standardized feature set of the current time step is input into the neural network, which then outputs an initial prediction of power generation. This prediction is then compared with the actual power generation, and the weight parameters of the neural network are adjusted using a reward function.

[0030] When the reward value reaches the preset threshold, the training process is stopped and the weight parameters of the current neural network are saved as the initial parameters.

[0031] Preferably, the step of performing knowledge migration on historical photovoltaic power generation data of multiple regions and seasons to generate a cross-regional and cross-seasonal knowledge migration matrix comprises the following steps:

[0032] Historical photovoltaic power generation data extracted from a standardized feature set, the historical photovoltaic power generation data including meteorological data before a current time step, real-time irradiation intensity data, and equipment operating status data;

[0033] Classify historical PV power generation data by region and season to generate a historical standardized dataset;

[0034] Constructing a meta-learning module, the meta-learning module includes a feature extraction unit and a knowledge transfer unit; the feature extraction unit is used to extract common features in the historical standardized data set, and the knowledge transfer unit is used to generate a cross-regional and cross-seasonal knowledge transfer matrix;

[0035] The knowledge transfer matrix is ​​embedded into a power generation prediction model, and the weight parameters of the prediction model are adjusted through matrix multiplication operations to form an advanced power generation prediction model with generalization capability.

[0036] Preferably, dynamically adjusting the optimal confidence threshold of the advanced power generation prediction model by using a Bayesian optimization strategy to generate an advanced power generation prediction result comprises the following steps:

[0037] Defining an objective function, the objective function including an advanced prediction error and a confidence threshold, wherein the advanced prediction error is the difference between the power generation prediction result output by the advanced power generation prediction model and the actual power generation, and is used to measure the accuracy of the prediction result; and the confidence threshold is used to measure the reliability of the prediction result;

[0038] Through Gaussian process regression modeling, the objective function is transformed into a global distribution map;

[0039] Combining the high-level prediction error of the current time step and the global distribution map, the optimal confidence threshold is searched through the Bayesian optimization algorithm; including:

[0040] The confidence threshold interval is limited to [0.1, 0.9] to avoid the confidence threshold being too low or too high, which leads to unreliability;

[0041] Taking multiple candidate confidence thresholds in Gaussian process regression models;

[0042] Calculate the objective function corresponding to each confidence threshold, and select the confidence threshold with the minimum objective function as the optimal confidence threshold.

[0043] Draw a large number of samples from the global distribution map, calculate the mean and standard deviation of the samples, and then generate the confidence interval and confidence interval width of the advanced prediction results;

[0044] The reliability level of the advanced prediction results is divided according to the matching degree between the confidence interval width and the optimal confidence threshold.

[0045] Preferably, the calculation formula of the objective function is:

[0046] ;

[0047] Where, is the confidence threshold, Is the objective function, used to measure the current confidence threshold The comprehensive performance of the advanced power generation forecasting model is evaluated. By minimizing the objective function, the optimal confidence threshold can be found to make the forecast results both accurate and reliable. The power generation forecast result output by the advanced power generation forecast model; is the actual power generation; is the high-level prediction error; is the weight coefficient of the advanced prediction error; is the weight coefficient of the confidence threshold; It is the proportional coefficient between the advanced prediction error and the confidence threshold, and is used to adjust the impact of the advanced prediction error on the confidence threshold.

[0048] Preferably, the objective function is converted into a global distribution map by Gaussian process regression modeling, comprising the steps of:

[0049] Constructing a mean function, where the mean function is the prediction result of the expected value of the objective function under a given confidence threshold, wherein the expected value of the objective function is the average value of the objective function given the same confidence threshold, which is used to reflect the average level of model prediction error under the confidence threshold;

[0050] Constructing a covariance function, wherein the covariance function is used to describe the error correlation between different confidence thresholds;

[0051] By training the mean function and covariance function of the Gaussian process regression model, the Gaussian process regression model can output the target function distribution corresponding to any confidence threshold, providing a global distribution map for the subsequent search for the optimal confidence threshold.

[0052] Preferably, the reliability level of the advanced prediction results is divided according to the matching degree between the confidence interval width and the optimal confidence threshold, including:

[0053] Set the confidence interval width to , the optimal confidence threshold is , the base width is ;

[0054] when When , the reliability level is high reliability;

[0055] when When , the reliability level is medium reliability;

[0056] when When , the reliability level is low reliability;

[0057] When the reliability is high or medium, the optimal confidence threshold can be directly used for scheduling. When the reliability is low, it is prompted to adjust the weight of the objective function and re-adjust the optimal confidence threshold dynamically.

[0058] Generate advanced energy production forecasts based on optimal confidence thresholds.

[0059] In a second aspect, the present invention further provides a system for intelligent photovoltaic power generation prediction and self-learning based on deep learning, wherein the system applies the method for intelligent photovoltaic power generation prediction and self-learning based on deep learning as described above, and the system includes a multi-source heterogeneous data acquisition module, an online incremental learning module, a knowledge transfer module, and a Bayesian optimization module;

[0060] The multi-source heterogeneous data acquisition module is used to collect meteorological data, real-time irradiation intensity data, and equipment operating status data during the operation of the photovoltaic power station, and pre-process the collected data to generate a standardized feature set;

[0061] An online incremental learning module is configured to perform online incremental learning on the standardized feature set using a neural network to generate an initial power generation prediction result; quantitatively evaluate and optimize the initial power generation prediction result using a reward function to generate initial parameters of a power generation prediction model;

[0062] A knowledge transfer module is used to transfer knowledge of historical photovoltaic power generation data of multiple regions and seasons, generate a cross-regional and cross-seasonal knowledge transfer matrix, and embed the knowledge transfer matrix into the photovoltaic power generation prediction model to form an advanced power generation prediction model with generalization capability;

[0063] The Bayesian optimization module dynamically adjusts the optimal confidence threshold of the advanced power generation prediction model through a Bayesian optimization strategy to generate an advanced power generation prediction result.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] 1. The multi-source heterogeneous data collection and standardization processing mechanism effectively integrates multi-dimensional operational data such as meteorological data, real-time radiation intensity, and equipment operating status. It overcomes the limitation of traditional methods that rely on a single data source and lead to insufficient information dimensions. It provides a more comprehensive and representative input basis for model training, and ensures the completeness of prediction information from the data level.

[0066] 2. The online incremental learning module, combined with the reward function mechanism, enables the model to continuously learn the relationship between the environment and power generation from new data during operation, dynamically optimize parameters, and avoid the generalization ability degradation problem of traditional static models caused by environmental changes. It significantly improves the model's adaptability to dynamic operation scenarios.

[0067] 3. The construction of a cross-regional and cross-seasonal knowledge transfer matrix effectively breaks through regional and seasonal restrictions by extracting common features of historical data and embedding them into the prediction model, enabling the model to quickly adapt to the differences in photovoltaic power generation characteristics in different regions and seasons, greatly expanding the application scenarios and generalization performance of the prediction model.

[0068] 4. The Bayesian optimization strategy dynamically adjusts the optimal confidence threshold through objective function and Gaussian process regression analysis. While ensuring prediction accuracy, it quantifies the reliability level of the evaluation results and provides a graded reference basis for actual scheduling. When the reliability of the prediction results is insufficient, the system can trigger a parameter adjustment mechanism to further improve the scientific nature and robustness of the decision. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0070] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0071] Figure 1 It is a flow chart of a method for intelligent photovoltaic power generation prediction and self-learning based on deep learning.

[0072] Figure 2 The present invention is a flowchart of S4 in a method for intelligent photovoltaic power generation prediction and self-learning based on deep learning.

[0073] Figure 3 It is a structural diagram of a system for intelligent photovoltaic power generation prediction and self-learning based on deep learning. DETAILED DESCRIPTION

[0074] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0075] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative position relationship, movement status, etc. between the various components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0076] In addition, the descriptions of "first", "second", etc. in the present invention are for descriptive purposes only and should not be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" or "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between the various embodiments can be combined with each other, but this must be based on the fact that they can be implemented by ordinary technicians in this field. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0077] Example 1

[0078] See Figure 1 As shown, the present invention provides a method for intelligent photovoltaic power generation prediction and self-learning based on deep learning, comprising the following steps:

[0079] S1. Collecting multi-source heterogeneous data during the operation of a photovoltaic power station, constructing an initial feature set based on the multi-source heterogeneous data, and preprocessing the initial feature set to generate a standardized feature set; the multi-source heterogeneous data includes meteorological data, real-time irradiation intensity data, and equipment operation status data;

[0080] S2. Performing online incremental learning on the standardized feature set through a neural network to generate an initial power generation prediction result; performing quantitative evaluation and optimization on the initial power generation prediction result through a reward function to generate initial parameters of a power generation prediction model;

[0081] S3. Performing knowledge migration on historical photovoltaic power generation data for multiple regions and seasons to generate a cross-regional and cross-seasonal knowledge migration matrix; embedding the knowledge migration matrix into the photovoltaic power generation prediction model to form an advanced power generation prediction model with generalization capability;

[0082] S4. Dynamically adjust the optimal confidence threshold of the advanced power generation prediction model through a Bayesian optimization strategy to generate an advanced power generation prediction result.

[0083] In this embodiment, first, multi-source heterogeneous data are collected from the operating environment of the photovoltaic power station through several sensor nodes of the multi-source heterogeneous data acquisition module. The multi-source heterogeneous data include meteorological data, real-time irradiation intensity data and equipment operation status data, and the collection frequency is once per minute. Among them, meteorological data can indirectly affect the power generation efficiency by changing the irradiation intensity, component temperature and air transparency; irradiation intensity is the direct energy source of photovoltaic power generation, and real-time irradiation intensity data includes total irradiation intensity per unit time, real-time irradiation angle, real-time direct irradiation amount and real-time scattered irradiation amount; equipment operation status data is used to reflect the The health level and working efficiency of all equipment, including the voltage, current, power and other parameters of photovoltaic modules; the collected multi-source heterogeneous data are stored in a distributed database and sorted according to timestamps to generate an initial feature set; the incomplete data in the initial feature set are completed by an interpolation algorithm, and the abnormal data are smoothed by a sliding window filtering algorithm to generate a standardized feature set; the standardized feature set is used as the input data of subsequent modules; and the standardized feature set is then divided into a training set and a test set, of which the training set accounts for 80% and the test set accounts for 20%. This division method ensures the reliability of subsequent prediction model training and verification.

[0084] Next, the online incremental learning module uses a neural network and a reward function to perform online incremental learning on the standardized feature set. The core components of the online incremental learning module are a neural network and a reward function. The neural network includes an input layer, a hidden layer, and an output layer. The input layer is used to import the standardized data set into the neural network. The hidden layer is used to reduce the dimension of the standardized feature set layer by layer and extract the abstract representation of the standardized feature set. In the present invention, the hidden layer is set to three layers, and the number of neurons in each layer is 128, 64, and 32 respectively. The first layer captures the basic associations of the standardized feature set through 128 neurons, the second layer filters redundant data through 64 neurons, and focuses on key associations. The third layer abstracts the key associations into an abstract representation of the standardized feature set through 32 neurons; the output layer outputs the preliminary prediction results of power generation based on the abstract representation; The output results of the neural network are then quantitatively evaluated and optimized through a reward function, and the weight parameters of the neural network are then updated. The reward function includes prediction error, model convergence speed, and dynamic environmental adaptability. The prediction error is the difference between the initial prediction result of power generation and the actual power generation, and is also used as the reward value of the reward function to drive the prediction model to learn the "environment-power generation" association law in historical data. The convergence speed is used to quickly reduce the prediction error during the neural network training process to avoid the training process falling into local optimality or taking too long. The dynamic environmental adaptability requires the prediction model to maintain the stability of the prediction error when facing new data, and at the same time drive the neural network to learn generalized dynamic patterns, such as cloud mutation laws.

[0085] At each time step, the standardized feature set for the current time step is input into the neural network, which then outputs an initial prediction of power generation. This prediction is then compared with the actual power generation, and the neural network's weight parameters are adjusted using a reward function. When the reward reaches a preset threshold, training stops and the current neural network weight parameters are saved as the initial parameters. The output of the online incremental learning module becomes the initial parameters of the power generation prediction model, which are then passed to the knowledge transfer module.

[0086] Then, knowledge transfer is performed on historical photovoltaic power generation data of multiple regions and seasons to generate a cross-regional and cross-season knowledge transfer matrix. The knowledge transfer matrix is ​​embedded into the power generation forecast model to form an advanced power generation forecast model with generalization capability, including:

[0087] First, historical photovoltaic power generation data is extracted from the standardized data set. The historical photovoltaic power generation data includes meteorological data before the current time step, real-time irradiation intensity data, and equipment operating status data.

[0088] Then, the historical photovoltaic power generation data is classified according to region and season to generate a historical standardized data set; the historical standardized data set includes a regional historical standardized data set and a seasonal historical standardized data set; the regional historical standardized data set represents a historical standardized data set within a fixed area; the seasonal historical standardized data set represents a historical standardized data set of a certain season;

[0089] The meta-learning module includes a feature extraction unit and a knowledge transfer unit; the feature extraction unit extracts common features from the historical standardized data set, and the knowledge transfer unit generates a cross-regional and cross-seasonal knowledge transfer matrix based on the historical standardized data set;

[0090] The knowledge transfer matrix is ​​embedded into the power generation prediction model, and the weight parameters of the power generation prediction model are adjusted through matrix multiplication operations to form an advanced power generation prediction model with generalization ability. The model can adapt to the photovoltaic power generation prediction needs of different regions and seasons.

[0091] Finally, the optimal confidence threshold of the advanced power generation prediction model is dynamically adjusted through the Bayesian optimization strategy to generate the photovoltaic power generation prediction result; Figure 2 Shown, including:

[0092] Define an objective function, which includes an advanced prediction error and a confidence threshold. The advanced prediction error is the difference between the power generation prediction result output by the advanced power generation prediction model and the actual power generation, and is used to measure the accuracy of the prediction result. The confidence threshold is used to measure the reliability of the prediction result. It is essentially a dynamically adjustable value with a numerical range of [0.1, 0.9].

[0093] The calculation formula of the objective function is:

[0094] ;

[0095] Where, is the confidence threshold, Is the objective function, used to measure the current confidence threshold The comprehensive performance of the advanced power generation forecasting model is evaluated. By minimizing the objective function, the optimal confidence threshold can be found to make the forecast results both accurate and reliable. The power generation forecast result output by the advanced power generation forecast model; is the actual power generation; is the high-level prediction error; is the weight coefficient of the advanced prediction error; is the weight coefficient of the confidence threshold; is the proportional coefficient between the advanced prediction error and the confidence threshold, which is used to adjust the impact of the advanced prediction error on the confidence threshold;

[0096] Through Gaussian process regression modeling, the objective function is transformed into a global distribution map, including:

[0097] Constructing a mean function, where the mean function is the prediction result of the expected value of the objective function under a given confidence threshold, wherein the expected value of the objective function is the average value of the objective function given the same confidence threshold, which is used to reflect the average level of model prediction error under the confidence threshold;

[0098] Construct a covariance function, which is used to describe the error correlation between different confidence thresholds; for example, let the covariance function be ,like is large, it means that the confidence threshold is Adjust to The objective function will change significantly when is very small, it means that the confidence threshold is Adjust to The objective function has almost no change;

[0099] By training the mean function and covariance function of the Gaussian process regression model, the Gaussian process regression model can output the target function distribution corresponding to any confidence threshold, providing a global distribution map for the subsequent search for the optimal confidence threshold.

[0100] It can be understood that the Bayesian optimization strategy in this implementation can balance the accuracy and reliability of the power generation prediction results by adjusting the confidence threshold. In order to achieve this goal, it is necessary to clarify how the objective function will change under different confidence thresholds. Since a large number of historical confidence thresholds and corresponding historical objective function samples are recorded in the historical data, these discrete historical data can be converted into a continuous global probability distribution graph through Gaussian process regression modeling, among which the global relationship between the historical confidence threshold and the historical objective function can be described by the mean function and covariance function, thereby providing a global distribution graph for searching the optimal confidence threshold.

[0101] Combining the high-level prediction error of the current time step and the global distribution map, the optimal confidence threshold is searched through the Bayesian optimization algorithm; including:

[0102] The confidence threshold interval is limited to [0.1, 0.9] to avoid the confidence threshold being too low or too high, which leads to unreliability;

[0103] Taking multiple candidate confidence thresholds in Gaussian process regression models;

[0104] Calculate the objective function corresponding to each confidence threshold, and select the confidence threshold with the minimum objective function as the optimal confidence threshold.

[0105] Draw a large number of samples from the global distribution map, calculate the mean and standard deviation of the samples, and then generate confidence intervals for advanced prediction results. For example, perform 1000 Monte Carlo samplings on the global distribution map and calculate the mean of the 1000 sampled samples. and standard deviation , and then intercept the interval at the 95% confidence level As the confidence interval, the confidence interval width is .

[0106] The reliability level of the advanced prediction results is divided according to the matching degree between the confidence interval width and the optimal confidence threshold, as shown below:

[0107] Set the confidence interval width to , the optimal confidence threshold is , the base width is ;

[0108] when When , the reliability level is high reliability;

[0109] when When , the reliability level is medium reliability;

[0110] when When , the reliability level is low reliability;

[0111] When the reliability is high or medium, the optimal confidence threshold can be directly used for scheduling; when the reliability is low, it is prompted that the weight of the objective function needs to be adjusted and the optimal confidence threshold needs to be dynamically adjusted again.

[0112] Generate advanced energy production forecasts based on optimal confidence thresholds.

[0113] Example 2

[0114] See Figure 3 As shown, the present invention also provides a system for intelligent photovoltaic power generation prediction and self-learning based on deep learning provided by the present invention, applying the above-mentioned method for intelligent photovoltaic power generation prediction and self-learning based on deep learning, the system includes a multi-source heterogeneous data acquisition module, an online incremental learning module, a knowledge transfer module and a Bayesian optimization module;

[0115] The multi-source heterogeneous data acquisition module is used to collect meteorological data, real-time irradiation intensity data, and equipment operating status data during the operation of the photovoltaic power station, and pre-process the collected data to generate a standardized feature set;

[0116] An online incremental learning module is configured to perform online incremental learning on the standardized feature set using a neural network to generate an initial power generation prediction result; quantitatively evaluate and optimize the initial power generation prediction result using a reward function to generate initial parameters of a power generation prediction model;

[0117] A knowledge transfer module is used to transfer knowledge of historical photovoltaic power generation data of multiple regions and seasons, generate a cross-regional and cross-seasonal knowledge transfer matrix, and embed the knowledge transfer matrix into the photovoltaic power generation prediction model to form an advanced power generation prediction model with generalization capability;

[0118] The Bayesian optimization module dynamically adjusts the optimal confidence threshold of the advanced power generation prediction model through a Bayesian optimization strategy to generate an advanced power generation prediction result.

[0119] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is intended to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for intelligent photovoltaic power generation prediction and self-learning based on deep learning, characterized in that: The following steps are involved: S1. Collecting multi-source heterogeneous data during the operation of a photovoltaic power station, constructing an initial feature set based on the multi-source heterogeneous data, and preprocessing the initial feature set to generate a standardized feature set; the multi-source heterogeneous data includes meteorological data, real-time irradiation intensity data, and equipment operation status data; S2. Performing online incremental learning on the standardized feature set through a neural network to generate an initial power generation prediction result; performing quantitative evaluation and optimization on the initial power generation prediction result through a reward function to generate initial parameters of a power generation prediction model; S3. Performing knowledge migration on historical photovoltaic power generation data for multiple regions and seasons to generate a cross-regional and cross-seasonal knowledge migration matrix; embedding the knowledge migration matrix into the power generation prediction model to form an advanced power generation prediction model with generalization capability; S4. Dynamically adjusting the optimal confidence threshold of the advanced power generation prediction model through a Bayesian optimization strategy to generate an advanced power generation prediction result, including the following steps: Defining an objective function, the objective function including an advanced prediction error and a confidence threshold, wherein the advanced prediction error is the difference between the power generation prediction result output by the advanced power generation prediction model and the actual power generation, and is used to measure the accuracy of the prediction result; and the confidence threshold is used to measure the reliability of the prediction result; Through Gaussian process regression modeling, the objective function is transformed into a global distribution map; Combining the high-level prediction error of the current time step and the global distribution map, the optimal confidence threshold is searched through the Bayesian optimization algorithm; include: Limit the confidence threshold interval to [0.1, 0.9]; Taking multiple candidate confidence thresholds in Gaussian process regression models; Calculate the objective function corresponding to each confidence threshold, and select the confidence threshold with the minimum objective function as the optimal confidence threshold; Draw a large number of samples from the global distribution map, calculate the mean and standard deviation of the samples, and then generate the confidence interval and confidence interval width of the advanced prediction results; The reliability level of the advanced prediction results is divided according to the matching degree between the confidence interval width and the optimal confidence threshold.

2. The method for intelligent photovoltaic power generation prediction and self-learning based on deep learning according to claim 1, characterized in that: The method of collecting multi-source heterogeneous data during the operation of a photovoltaic power station includes the following steps: Deploy multiple sensor nodes to collect meteorological data, real-time radiation intensity data, and equipment operation status data; set the data collection frequency to once per minute; The collected multi-source heterogeneous data is stored in a distributed database and sorted by timestamp to generate an initial feature set; Preprocessing the initial feature set to generate a standardized feature set; The standardized feature set is divided into a training set and a test set, where the training set accounts for 80% and the test set accounts for 20%.

3. The method for intelligent photovoltaic power generation prediction and self-learning based on deep learning according to claim 2, characterized in that: The preprocessing is specifically as follows: Interpolation algorithms are used to fill in missing data; The abnormal data is smoothed using a sliding window filtering algorithm.

4. The method for intelligent photovoltaic power generation prediction and self-learning based on deep learning according to claim 1, characterized in that: The method includes the following steps: performing online incremental learning on the standardized feature set through a neural network to generate an initial power generation prediction result; and performing quantitative evaluation and optimization on the initial power generation prediction result through a reward function to generate initial parameters of a power generation prediction model. Initialize the neural network structure in the deep reinforcement learning framework, which includes an input layer, a hidden layer, and an output layer; set the number of hidden layers to 3, and the number of neurons in each layer to 128, 64, and 32 respectively; The input layer is used to import the standardized data set into the neural network; The hidden layer is used to reduce the dimension of the standardized feature set layer by layer and extract an abstract representation of the standardized feature set; The output layer is used to output a preliminary prediction result of power generation according to the abstract representation; A reward function is defined, which includes prediction error, model convergence speed, and dynamic environment adaptability; the reward function is used to quantitatively evaluate and optimize the output results of the neural network and update the weight parameters of the neural network, wherein, The prediction error is the difference between the initial power generation prediction result and the actual power generation. It also serves as the reward value of the reward function, which is used to drive the prediction model to learn the "environment-power generation" association law in historical data. The convergence rate is used to reduce prediction errors during neural network training and prevent the training process from falling into a local optimum or taking too long; The dynamic environment adaptability requires the prediction model to maintain the stability of prediction error when facing new data, while driving the dynamic pattern of neural network learning generalization; In each time step, the standardized feature set of the current time step is input into the neural network, and the neural network outputs the initial prediction result of power generation. The initial prediction result of power generation is compared with the actual power generation, and the weight parameters of the neural network are adjusted through the reward function; When the reward value reaches the preset threshold, the training process is stopped and the weight parameters of the current neural network are saved as the initial parameters.

5. The method for intelligent photovoltaic power generation prediction and self-learning based on deep learning according to claim 1, characterized in that: The method of performing knowledge migration on historical photovoltaic power generation data of multiple regions and seasons to generate a cross-regional and seasonal knowledge migration matrix includes the following steps: Historical photovoltaic power generation data extracted from a standardized feature set, the historical photovoltaic power generation data including meteorological data before a current time step, real-time irradiation intensity data, and equipment operating status data; Classify historical PV power generation data by region and season to generate a historical standardized dataset; Constructing a meta-learning module, wherein the meta-learning module includes a feature extraction unit and a knowledge transfer unit; The feature extraction unit is used to extract common features from the historical standardized data set, and the knowledge transfer unit is used to generate a cross-regional and cross-season knowledge transfer matrix; The knowledge transfer matrix is ​​embedded into a power generation prediction model, and the weight parameters of the prediction model are adjusted through matrix multiplication operations to form an advanced power generation prediction model with generalization capability.

6. The method for intelligent photovoltaic power generation prediction and self-learning based on deep learning according to claim 1, characterized in that: The calculation formula of the objective function is: ； Where, is the confidence threshold, Is the objective function, used to measure the current confidence threshold The comprehensive advantages and disadvantages of the performance of the advanced power generation forecasting model; The power generation forecast result output by the advanced power generation forecast model; is the actual power generation; is the high-level prediction error; is the weight coefficient of the advanced prediction error; is the weight coefficient of the confidence threshold; It is the proportional coefficient between the advanced prediction error and the confidence threshold, and is used to adjust the impact of the advanced prediction error on the confidence threshold.

7. The method of intelligent photovoltaic power generation prediction and self-learning based on deep learning according to claim 1, characterized in that: The objective function is converted into a global distribution map through Gaussian process regression modeling, including the steps of: Constructing a mean function, where the mean function is the prediction result of the expected value of the objective function under a given confidence threshold, wherein the expected value of the objective function is the average value of the objective function given the same confidence threshold, which is used to reflect the average level of model prediction error under the confidence threshold; Constructing a covariance function, wherein the covariance function is used to describe the error correlation between different confidence thresholds; By training the mean function and covariance function of the Gaussian process regression model, the Gaussian process regression model is used to output the target function distribution corresponding to any confidence threshold, providing a global distribution map for the subsequent search for the optimal confidence threshold.

8. The method of intelligent photovoltaic power generation prediction and self-learning based on deep learning according to claim 1, characterized in that: The reliability level of the advanced prediction results is divided according to the matching degree between the confidence interval width and the optimal confidence threshold, including: Set the confidence interval width to , the optimal confidence threshold is , the base width is ; when When , the reliability level is high reliability; when When , the reliability level is medium reliability; when When , the reliability level is low reliability; When the reliability is high or medium, the optimal confidence threshold is directly used for scheduling. When the reliability is low, it is prompted to adjust the weight of the objective function and re-adjust the optimal confidence threshold dynamically. Generate advanced energy production forecasts based on optimal confidence thresholds.

9. A system for intelligent photovoltaic power generation prediction and self-learning based on deep learning, applying the method for intelligent photovoltaic power generation prediction and self-learning based on deep learning according to any one of claims 1 to 8, characterized in that: The system includes a multi-source heterogeneous data acquisition module, an online incremental learning module, a knowledge transfer module and a Bayesian optimization module; The multi-source heterogeneous data acquisition module is used to collect meteorological data, real-time irradiation intensity data, and equipment operating status data during the operation of the photovoltaic power station, and pre-process the collected data to generate a standardized feature set; An online incremental learning module is configured to perform online incremental learning on the standardized feature set using a neural network to generate an initial power generation prediction result; quantitatively evaluate and optimize the initial power generation prediction result using a reward function to generate initial parameters of a power generation prediction model; A knowledge transfer module is used to transfer knowledge of historical photovoltaic power generation data of multiple regions and seasons, generate a cross-regional and cross-seasonal knowledge transfer matrix, and embed the knowledge transfer matrix into the power generation forecast model to form an advanced power generation forecast model with generalization capability; The Bayesian optimization module dynamically adjusts the optimal confidence threshold of the advanced power generation prediction model through a Bayesian optimization strategy to generate an advanced power generation prediction result.

Images