Conversion power prediction method and system considering air humidity

Abstract

The present invention provides a conversion power prediction method and system considering air humidity, which relates to the field of electric power technology, including obtaining historical meteorological data of a target photovoltaic power station and corresponding actual conversion power data; extracting design parameters of the photovoltaic power station; constructing a nonlinear mapping model of air relative humidity and photovoltaic component conversion efficiency; based on pre-processed historical meteorological data and conversion power data of the photovoltaic power station, combined with the design parameters of the photovoltaic power station, forming a photovoltaic power station conversion power prediction model; using a variational mode decomposition algorithm to adaptively decompose the conversion power data of the photovoltaic power station conversion power prediction model, and constructing a photovoltaic power station conversion power prediction model driven by multi-source heterogeneous data.

Description

Technical Field

[0001] The present invention relates to electric power technology, and in particular to a conversion power prediction method and system considering air humidity. Background Art

[0002] There are two main types of photovoltaic power generation prediction methods: data-driven and physical modeling. These methods have some shortcomings:

[0003] Most models only consider the impact of solar radiation intensity and ambient temperature on power generation, ignoring the effect of air humidity. In fact, air humidity affects the transmittance of solar radiation and the heat conduction on the surface of photovoltaic modules, thus affecting the photoelectric conversion efficiency and power generation.

[0004] Most existing models use fixed model structures and parameters, lack optimization mechanisms for prediction models, and are difficult to adapt to the time-varying characteristics of photovoltaic systems and dynamic changes in the external environment.

[0005] Traditional parameter identification algorithms such as the least squares method and gradient descent method are prone to fall into local optimality, have low solution efficiency, and are difficult to meet the needs of real-time prediction. Summary of the invention

[0006] The embodiments of the present invention provide a conversion power prediction method and system taking air humidity into consideration, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention,

[0008] Provides a conversion power prediction method considering air humidity, including:

[0009] Obtain the historical meteorological data of the target photovoltaic power station and the corresponding actual conversion power data, the historical meteorological data includes the radiation intensity of the whole day, the scattered radiation intensity, the ambient temperature and the relative humidity of the air, sort the historical meteorological data according to the time series, and identify and correct the missing values ​​and abnormal values; extract the design parameters of the photovoltaic power station, the design parameters include the type of photovoltaic modules, the installation angle, the installation spacing, and the total installed capacity;

[0010] Construct a nonlinear mapping model between relative air humidity and photovoltaic module conversion efficiency, obtain photovoltaic module conversion efficiency data under different relative air humidity levels through controlled variable experiments, and use support vector regression machine to fit the nonlinear relationship curve between photovoltaic module conversion efficiency and relative air humidity; Based on the pre-processed historical meteorological data and conversion power data of the photovoltaic power station, combined with the design parameters of the photovoltaic power station, use the least squares support vector regression machine to construct a nonlinear mapping relationship model between historical meteorological data and photovoltaic power station conversion power, and couple the nonlinear mapping model between relative air humidity and photovoltaic module conversion efficiency with the nonlinear mapping relationship model between historical meteorological data and photovoltaic power station conversion power to form a photovoltaic power station conversion power prediction model;

[0011] The conversion power data of the photovoltaic power station conversion power prediction model is adaptively decomposed by the variational mode decomposition algorithm, and the multi-scale periodic characteristics and local fluctuation characteristics of the photovoltaic power station conversion power are extracted. The extracted characteristic variables and meteorological parameters are used as the input of the least squares support vector regression machine to construct a photovoltaic power station conversion power prediction model driven by multi-source heterogeneous data; the training set data is randomly shuffled and the established photovoltaic power station conversion power prediction model is trained and verified by cross-validation. Regularization and early stopping strategies are used to avoid overfitting of the photovoltaic power station conversion power prediction model. The optimal hyperparameters of the photovoltaic power station conversion power prediction model are obtained through grid search, and the prediction sub-models trained with multiple different random seeds are weighted combined using an ensemble learning strategy to improve the robustness and generalization performance of the photovoltaic power station conversion power prediction model.

[0012] In an optional embodiment,

[0013] The nonlinear mapping model between relative air humidity and photovoltaic module conversion efficiency is constructed, and the photovoltaic module conversion efficiency data under different relative air humidity levels are obtained through the control variable experiment. The steps of fitting the nonlinear relationship curve between photovoltaic module conversion efficiency and relative air humidity by support vector regression machine include:

[0014] The conversion efficiency data of photovoltaic modules under different relative air humidity conditions were obtained through controlled variable experiments. During the experiment, the relative air humidity in the experimental chamber was adjusted by a humidifier and a dehumidifier, and multiple humidity gradients were set. The electrical parameters and environmental parameters of the photovoltaic modules were continuously monitored under each humidity gradient, and other environmental factors were kept constant to eliminate interference. The real-time conversion efficiency of the photovoltaic modules under different relative air humidity was calculated using open circuit voltage, short circuit current, maximum power point voltage and current parameters.

[0015] For each type of photovoltaic module, the average conversion efficiency of multiple measurements under different relative air humidity levels is extracted to construct an experimental data set. Considering the significant nonlinear relationship between relative air humidity and photovoltaic module conversion efficiency, support vector regression machine is used for curve fitting. The support vector regression machine realizes nonlinear mapping from the original space to the high-dimensional feature space by introducing the kernel function, constructs a linear regression model in the feature space, introduces slack variables and ε-insensitive loss function, and establishes the optimization objective equation of the support vector regression machine.

[0016] The Gaussian radial basis kernel function is selected as the kernel function of the support vector regression machine. The Lagrange multiplier method and the Carlo-Kuhn-Tucker condition are used to transform the optimization objective equation into a dual problem for solution. The support vector, Lagrange multiplier and bias term parameters of the support vector regression machine model are obtained through training, and the nonlinear regression function of the photovoltaic module conversion efficiency with respect to the relative humidity of the air is constructed.

[0017] The regularization coefficient and kernel function bandwidth parameters of the support vector regression model are optimized by combining grid search with cross validation. The nonlinear function relationship curve between the conversion efficiency of photovoltaic modules and the relative humidity of air is fitted using the optimized support vector regression model.

[0018] In an optional embodiment,

[0019] Based on the pre-processed historical meteorological data and conversion power data of the photovoltaic power station, combined with the design parameters of the photovoltaic power station, the steps of using the least squares support vector regression mechanism to construct a nonlinear mapping relationship model between the historical meteorological data and the conversion power of the photovoltaic power station include:

[0020] The historical operation data of photovoltaic power stations are cleaned and normalized, and the input feature vector is constructed according to the time-lag characteristics of the parameters of all-day radiation intensity, scattered radiation intensity, ambient temperature and relative humidity of air; a nonlinear mapping function is introduced to map the input feature vector to a high-dimensional feature space, and a linear regression model is constructed in the high-dimensional feature space to establish a nonlinear mapping relationship between meteorological parameters and the conversion power of photovoltaic power stations;

[0021] The time-series correlation constraint term is introduced into the objective function of the least squares support vector regression machine. The similarity between samples in the feature space after mapping is measured by the kernel function, so that samples that are close in time and have similar features have similar weight coefficients, thereby describing the time-dependent characteristics of meteorological parameters.

[0022] Construct the optimization objective function of the least squares support vector regression machine with time-series correlation constraint term, which consists of three parts: empirical risk term, structural risk term and time-series correlation constraint term;

[0023] The whale optimization algorithm is used to search for the optimal parameter combination of the least squares support vector regression machine model, and the regularization parameter, time series correlation constraint weight coefficient and kernel function scale parameter of the model are optimized by updating the position of the whale population; the least squares support vector regression machine model is trained on all training data using the optimal parameter combination obtained by the whale optimization algorithm, and the meteorological parameters at the time to be predicted are input into the trained least squares support vector regression machine model to obtain the corresponding photovoltaic power station conversion power prediction value, forming a nonlinear mapping relationship model between meteorological data and photovoltaic power station conversion power;

[0024] By introducing time-series correlation constraints into the objective function and using the whale optimization algorithm for parameter optimization, the ability of the least squares support vector regression model to characterize the nonlinear relationship between meteorological parameters and conversion power is improved, thereby improving the accuracy and stability of photovoltaic power station conversion power prediction.

[0025] In an optional embodiment,

[0026] The steps of using the whale optimization algorithm to search for the optimal parameter combination of the least squares support vector regression model and optimizing the regularization parameters, time series correlation constraint weight coefficients and kernel function scale parameters of the model by updating the position of the whale population include:

[0027] The regularization parameters, time-series correlation constraint weights and kernel function scale parameters in the least squares support vector regression model are encoded into the position coordinates of the whales in the whale optimization algorithm, and the initial positions of a group of whales are randomly generated to form an initial population.

[0028] For each whale in the population, decode its position coordinates into least squares support vector regression model parameters, train the least squares support vector regression model and calculate the prediction error on the validation set as the fitness value of the whale;

[0029] The whale with the highest fitness value in the current population is taken as the leader whale, and the position of each whale in the population is updated by adopting the siege predation mode or the spiral attack mode according to the probability parameter; in the siege predation mode, the whale calculates the coefficient vector according to the linear decreasing variable and the random vector, and updates its own position based on the position and coefficient vector of the leader whale; in the spiral attack mode, the whale updates its own position according to the distance between it and the leader whale and the spiral function to enhance the global exploration ability of the algorithm;

[0030] For whales whose updated positions are out of the range, their positions are adjusted to the boundary of the parameters, and it is determined whether the termination conditions of the maximum number of iterations or the performance improvement threshold are met; if the termination conditions are not met, the fitness value calculation step is returned to continue the iterative optimization; if the termination conditions are met, the position coordinates of the leading whale are decoded into the optimal parameter combination of the least squares support vector regression machine.

[0031] In an optional embodiment, the method further includes:

[0032] An adaptive hybrid search strategy is introduced to dynamically adjust the selection probability of the siege predation mode and the spiral attack mode. The adaptive weight coefficient w is designed to represent the probability of selecting the siege predation mode. The probability of selecting the spiral attack mode is 1-w. The value of w is adaptively adjusted according to the current number of iterations t and the population diversity index d. The calculation formula is as follows:

[0033] ;

[0034] Among them, T max is the maximum number of iterations, d max is the maximum value of population diversity, and the population diversity index d is measured by the average Euclidean distance between individuals in the population:

[0035] ;

[0036] Where N represents the total number of samples in the data set, X represents the samples in the data set, i and j represent the index numbers of the samples, which are used to traverse and compare the samples pairwise. i -X j || represents the distance between the i-th sample and the j-th sample;

[0037] As the iteration proceeds, the value of w gradually decreases, indicating that the probability of selecting the spiral attack mode increases, and the local search capability is enhanced; w is related to the population diversity index d. When the population diversity is low, the value of w is relatively large, indicating that the probability of selecting the siege predation mode increases, and the global exploration capability is enhanced to avoid the algorithm from converging prematurely;

[0038] In the siege predation mode and spiral attack mode, an adaptive parameter control mechanism is introduced to dynamically adjust the algorithm parameters and improve the optimization performance; for the shrinkage factor a in the siege predation mode, an adaptive update strategy is designed:

[0039] ;

[0040] Among them, a min and a max are the minimum and maximum values ​​of a respectively, and k is a control parameter used to adjust the decay speed of a. As the iteration proceeds, the value of a changes from a to max Gradually decrease to a min , which enables the algorithm to have stronger global exploration capabilities in the early stage and stronger local development capabilities in the later stage;

[0041] For the spiral shape parameter b in the spiral attack mode, an adaptive update strategy is designed:

[0042] ;

[0043] Among them, b min and b max are the minimum and maximum values ​​of b respectively. As the iteration proceeds, the value of b changes from b max Gradually decrease to b min , so that the algorithm performs global search with a larger spiral step size in the early stage and local search with a smaller spiral step size in the later stage, improving the convergence speed and accuracy;

[0044] A global exploration strategy based on Lévy flight is introduced. Lévy flight exploration is performed with a certain probability in each iteration. The random step size that obeys the Lévy distribution is generated by the Lévy flight function to update the current solution, so as to enhance the algorithm's ability to escape from the local optimum. During the iteration process, the siege and predation mode and the spiral attack mode are dynamically selected through an adaptive hybrid search strategy, and the algorithm parameters are adjusted using an adaptive parameter control mechanism. At the same time, the Lévy flight strategy is combined for global exploration to balance the development and utilization of the algorithm.

[0045] In an optional embodiment,

[0046] The conversion power data of the photovoltaic power station conversion power prediction model is adaptively decomposed using the variational mode decomposition algorithm to extract the multi-scale periodic characteristics and local fluctuation characteristics of the photovoltaic power station conversion power. The extracted characteristic variables and meteorological parameters are used as the input of the least squares support vector regression machine. The steps of constructing a photovoltaic power station conversion power prediction model driven by multi-source heterogeneous data include:

[0047] The original conversion power sequence of the photovoltaic power station is preprocessed to remove outliers and missing values, and the power data is smoothed and normalized; according to the spectral characteristics of the preprocessed conversion power sequence, the number and penalty parameters of the intrinsic mode function components in the variational mode decomposition are preliminarily determined;

[0048] The preprocessed power sequence is adaptively decomposed using the variational mode decomposition algorithm. The original power sequence is decomposed into several intrinsic mode function components by solving the constraint optimization problem of variational mode decomposition.

[0049] Hilbert transform is performed on each intrinsic mode function component obtained by variational mode decomposition to extract the envelope and instantaneous frequency characteristics of each component, and the multi-scale periodic characteristics and local fluctuation characteristics of the conversion power sequence are reconstructed according to the frequency characteristics and energy distribution of each component; the reconstructed multi-scale periodic characteristics and local fluctuation characteristics are integrated with the meteorological parameters of the photovoltaic power station to form a multi-source heterogeneous feature set as the input of the photovoltaic power station conversion power prediction model.

[0050] In an optional embodiment,

[0051] After randomly shuffling the training set data, the established photovoltaic power station conversion power prediction model is trained and verified by cross-validation. Regularization and early stopping strategies are used to avoid overfitting of the photovoltaic power station conversion power prediction model. The steps of obtaining the optimal hyperparameters of the photovoltaic power station conversion power prediction model through grid search include:

[0052] The training set data of the photovoltaic power station conversion power prediction model is randomly shuffled, and the photovoltaic power station conversion power prediction model is trained and verified by K-fold cross-validation. The training set is divided into K mutually exclusive subsets, and a different subset is selected as the verification set each time. The remaining subsets constitute the training set. The photovoltaic power station conversion power prediction model is trained with the training set, and the performance of the photovoltaic power station conversion power prediction model is evaluated on the verification set.

[0053] The L2 norm penalty term of the parameter vector is introduced into the objective function of the photovoltaic power station conversion power prediction model. The size of the photovoltaic power station conversion power prediction model parameters is constrained by L2 regularization, the complexity of the photovoltaic power station conversion power prediction model is controlled, and the generalization ability of the photovoltaic power station conversion power prediction model is improved when the training data is insufficient.

[0054] After each training round, the performance of the photovoltaic power plant conversion power prediction model on the validation set is evaluated. When the performance of the validation set has not improved for many consecutive times, the early stopping strategy is used to terminate the training process in advance to prevent the photovoltaic power plant conversion power prediction model from overfitting the training data.

[0055] The hyperparameters of the photovoltaic power station conversion power prediction model are optimized through grid search, and the preset hyperparameter combinations are exhaustively searched. Cross-validation is used to evaluate the performance of the photovoltaic power station conversion power prediction model corresponding to each hyperparameter combination, and the hyperparameter combination with the best performance in the validation set is selected as the final parameter configuration of the photovoltaic power station conversion power prediction model.

[0056] According to a second aspect of the embodiments of the present invention,

[0057] Provides a conversion power prediction system that takes air humidity into account, including:

[0058] The first unit is used to obtain the historical meteorological data of the target photovoltaic power station and the corresponding actual conversion power data, the historical meteorological data includes the radiation intensity of the whole day, the scattered radiation intensity, the ambient temperature and the relative humidity of the air, sort the historical meteorological data according to the time series, and identify and correct the missing values ​​and abnormal values; extract the design parameters of the photovoltaic power station, the design parameters include the type of photovoltaic modules, the installation angle, the installation spacing, and the total installed capacity;

[0059] The second unit is used to construct a nonlinear mapping model between relative air humidity and photovoltaic module conversion efficiency. Through the control variable experiment, the photovoltaic module conversion efficiency data under different relative air humidity levels are obtained, and the nonlinear relationship curve between the photovoltaic module conversion efficiency and the relative air humidity is fitted by the support vector regression machine. Based on the pre-processed historical meteorological data and conversion power data of the photovoltaic power station, combined with the design parameters of the photovoltaic power station, the least squares support vector regression machine is used to construct a nonlinear mapping relationship model between the historical meteorological data and the conversion power of the photovoltaic power station. The nonlinear mapping model between relative air humidity and photovoltaic module conversion efficiency is coupled with the nonlinear mapping relationship model between historical meteorological data and the conversion power of the photovoltaic power station to form a photovoltaic power station conversion power prediction model.

[0060] The third unit is used to use the variational mode decomposition algorithm to adaptively decompose the conversion power data of the photovoltaic power station conversion power prediction model, extract the multi-scale periodic characteristics and local fluctuation characteristics of the photovoltaic power station conversion power, and use the extracted characteristic variables and meteorological parameters as the input of the least squares support vector regression machine to build a photovoltaic power station conversion power prediction model driven by multi-source heterogeneous data; after randomly shuffling the training set data, the established photovoltaic power station conversion power prediction model is trained and verified by cross-validation, and the regularization and early stopping strategies are used to avoid overfitting of the photovoltaic power station conversion power prediction model. The optimal hyperparameters of the photovoltaic power station conversion power prediction model are obtained through grid search, and the prediction sub-models obtained by training with multiple different random seeds are weighted combined using an integrated learning strategy to improve the robustness and generalization performance of the photovoltaic power station conversion power prediction model.

[0061] According to a third aspect of the embodiments of the present invention,

[0062] An electronic device is provided, comprising:

[0063] processor;

[0064] a memory for storing processor-executable instructions;

[0065] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0066] A fourth aspect of the embodiments of the present invention is:

[0067] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0068] This paper introduces an adaptive hybrid search strategy, dynamically selects the siege predation mode and the spiral attack mode, and uses the adaptive parameter control mechanism to adjust the algorithm parameters, which can balance the global exploration and local development capabilities of the algorithm and improve the optimization efficiency. At the same time, combined with the Levy flight strategy for global exploration, the algorithm's ability to jump out of the local optimum is further enhanced to avoid premature convergence. The improved whale optimization algorithm proposed in this paper has stronger optimization performance and convergence speed in the parameter optimization of the photovoltaic power station conversion power prediction model.

[0069] This paper uses the variational mode decomposition algorithm to adaptively decompose the conversion power data of photovoltaic power plants, extract the multi-scale periodic characteristics and local fluctuation characteristics of the conversion power series, and integrate them with meteorological parameters to construct a multi-source heterogeneous feature set. By inputting this feature set into the least squares support vector regression machine, the multi-source heterogeneous features contained in the conversion power data and meteorological data can be fully utilized to improve the prediction accuracy and robustness of the photovoltaic power plant conversion power prediction model.

[0070] This paper uses cross-validation and early stopping strategies to train and verify the photovoltaic power station conversion power prediction model, and combines grid search to obtain the optimal hyperparameters of the prediction model, which can effectively improve the generalization performance and robustness of the prediction model and avoid overfitting the model to the training data. In practical applications, the number of cross-validation folds, the number of tolerance rounds of the early stopping method, and the search range of hyperparameters can be flexibly selected according to the characteristics of the data set and the complexity of the model to balance model performance and computational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 A schematic flow chart of a conversion power prediction method taking air humidity into consideration according to an embodiment of the present invention;

[0072] Figure 2 It is a structural schematic diagram of a conversion power prediction system taking air humidity into consideration according to an embodiment of the present invention. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0075] Figure 1 FIG. 1 is a flow chart of a conversion power prediction method considering air humidity according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0076] S1. Obtain the historical meteorological data of the target photovoltaic power station and the corresponding actual conversion power data, the historical meteorological data includes the radiation intensity, scattered radiation intensity, ambient temperature and relative humidity of the air throughout the day, sort the historical meteorological data according to the time series, and identify and correct the missing values ​​and abnormal values; extract the design parameters of the photovoltaic power station, the design parameters include the type of photovoltaic modules, installation angle, installation spacing, and total installed capacity;

[0077] S2. Construct a nonlinear mapping model between relative air humidity and photovoltaic module conversion efficiency, obtain photovoltaic module conversion efficiency data under different relative air humidity levels through controlled variable experiments, and use support vector regression machine to fit the nonlinear relationship curve between photovoltaic module conversion efficiency and relative air humidity; based on the pre-processed historical meteorological data and conversion power data of the photovoltaic power station, combined with the design parameters of the photovoltaic power station, use the least squares support vector regression machine to construct a nonlinear mapping relationship model between historical meteorological data and photovoltaic power station conversion power, and couple the nonlinear mapping model between relative air humidity and photovoltaic module conversion efficiency with the nonlinear mapping relationship model between historical meteorological data and photovoltaic power station conversion power to form a photovoltaic power station conversion power prediction model;

[0078] S3. The conversion power data of the photovoltaic power station conversion power prediction model is adaptively decomposed by using the variational mode decomposition algorithm to extract the multi-scale periodic characteristics and local fluctuation characteristics of the photovoltaic power station conversion power. The extracted characteristic variables and meteorological parameters are used together as the input of the least squares support vector regression machine to construct a photovoltaic power station conversion power prediction model driven by multi-source heterogeneous data; the training set data is randomly shuffled and the established photovoltaic power station conversion power prediction model is trained and verified by cross-validation. Regularization and early stopping strategies are used to avoid overfitting of the photovoltaic power station conversion power prediction model. The optimal hyperparameters of the photovoltaic power station conversion power prediction model are obtained through grid search, and the prediction sub-models trained with multiple different random seeds are weighted combined using an ensemble learning strategy to improve the robustness and generalization performance of the photovoltaic power station conversion power prediction model.

[0079] In an optional embodiment,

[0080] The nonlinear mapping model between relative air humidity and photovoltaic module conversion efficiency is constructed, and the photovoltaic module conversion efficiency data under different relative air humidity levels are obtained through the control variable experiment. The steps of fitting the nonlinear relationship curve between photovoltaic module conversion efficiency and relative air humidity by support vector regression machine include:

[0081] The conversion efficiency data of photovoltaic modules under different relative air humidity conditions were obtained through controlled variable experiments. During the experiment, the relative air humidity in the experimental chamber was adjusted by a humidifier and a dehumidifier, and multiple humidity gradients were set. The electrical parameters and environmental parameters of the photovoltaic modules were continuously monitored under each humidity gradient, and other environmental factors were kept constant to eliminate interference. The real-time conversion efficiency of the photovoltaic modules under different relative air humidity was calculated using open circuit voltage, short circuit current, maximum power point voltage and current parameters.

[0082] For each type of photovoltaic module, the average conversion efficiency of multiple measurements under different relative air humidity levels is extracted to construct an experimental data set. Considering the significant nonlinear relationship between relative air humidity and photovoltaic module conversion efficiency, support vector regression machine is used for curve fitting. The support vector regression machine realizes nonlinear mapping from the original space to the high-dimensional feature space by introducing the kernel function, constructs a linear regression model in the feature space, introduces slack variables and ε-insensitive loss function, and establishes the optimization objective equation of the support vector regression machine.

[0083] The Gaussian radial basis kernel function is selected as the kernel function of the support vector regression machine. The Lagrange multiplier method and the Carlo-Kuhn-Tucker condition are used to transform the optimization objective equation into a dual problem for solution. The support vector, Lagrange multiplier and bias term parameters of the support vector regression machine model are obtained through training, and the nonlinear regression function of the photovoltaic module conversion efficiency with respect to the relative humidity of the air is constructed.

[0084] The regularization coefficient and kernel function bandwidth parameters of the support vector regression model are optimized by combining grid search with cross validation. The nonlinear function relationship curve between the conversion efficiency of photovoltaic modules and the relative humidity of air is fitted using the optimized support vector regression model.

[0085] For example, in order to study the effect of relative air humidity on the conversion efficiency of photovoltaic modules, it is necessary to obtain efficiency data under different humidity conditions through controlled variable experiments. The specific experimental steps are as follows:

[0086] Several monocrystalline silicon photovoltaic modules with consistent performance parameters were selected as experimental objects, and their key parameters such as nominal power, open circuit voltage, and short circuit current were recorded. An indoor experimental test platform was built, including artificial climate chamber, programmable DC power supply, electronic load, data acquisition instrument and other equipment. The artificial climate chamber was used to simulate different temperature and humidity environments, the programmable DC power supply provided stable light intensity, the electronic load realized the maximum power point tracking (MPPT) of the photovoltaic module, and the data acquisition instrument recorded the voltage and current data of the experimental process.

[0087] The photovoltaic module was installed in an artificial climate chamber, and the parameters of the climate chamber were adjusted to keep the temperature at 25°C to simulate the standard test conditions (STC). A programmable DC power supply was used to provide a stable light of 1000W / m², and an electronic load was connected to the output end of the photovoltaic module to achieve MPPT operation. The above conditions were kept unchanged, and only the relative humidity of the air in the climate chamber was adjusted. Five humidity gradients were set, namely 30%, 45%, 60%, 75%, and 90%, as the control variables.

[0088] At each relative humidity level, after the output power of the photovoltaic module is stable, the voltage and current data are continuously collected for 10 minutes at a frequency of 1 Hz by the data acquisition instrument. The start and end time of data collection, as well as the real-time temperature and humidity values ​​displayed by the climate box are recorded. Repeat the above process to complete the data collection under 5 humidity gradients, and repeat the experiment at least 3 times for each gradient to reduce accidental errors.

[0089] For example, under the condition of 30% relative humidity, the output voltage of the photovoltaic module collected in a certain experiment was 32.85V, and the output current was 8.12A, from which the output power was calculated to be 266.74W. The nominal power of the photovoltaic module is 270W, so the conversion efficiency is 266.74 / 270=98.8%. The results of three repeated experiments were averaged, and the average conversion efficiency under the condition of 30% relative humidity was 98.5%.

[0090] The collected raw data were sorted and preprocessed, outliers were removed, the average output power and average conversion efficiency of photovoltaic modules under each humidity condition were calculated, and the experimental error was estimated. The corresponding relationship between conversion efficiency and relative humidity was plotted into a curve to intuitively reflect the influence trend of humidity on photovoltaic performance.

[0091] The influence mechanism of relative air humidity on the conversion efficiency of photovoltaic modules is analyzed, and the reasons why humidity causes the efficiency to decrease are discussed. The efficiency differences under different humidity conditions are compared to determine the critical impact value of humidity. Combined with the geographical location and climate characteristics of photovoltaic power stations, engineering measures to reduce air humidity and alleviate the impact of humidity are suggested, such as installing ventilation and drying devices, optimizing cleaning cycles, etc.

[0092] Through the above controlled variable experiments, the data on the change of photovoltaic module conversion efficiency under different relative humidity of air was obtained, revealing the influence of humidity on photovoltaic power generation performance, and providing data support and theoretical basis for humidity management and efficiency improvement of photovoltaic power stations. In practical applications, photovoltaic module selection and power station design can be optimized for different humidity areas to minimize humidity loss and improve photovoltaic power generation efficiency.

[0093] In an optional embodiment,

[0094] Based on the pre-processed historical meteorological data and conversion power data of the photovoltaic power station, combined with the design parameters of the photovoltaic power station, the steps of using the least squares support vector regression mechanism to construct a nonlinear mapping relationship model between the historical meteorological data and the conversion power of the photovoltaic power station include:

[0095] The historical operation data of photovoltaic power stations are cleaned and normalized, and the input feature vector is constructed according to the time-lag characteristics of the parameters of all-day radiation intensity, scattered radiation intensity, ambient temperature and relative humidity of air; a nonlinear mapping function is introduced to map the input feature vector to a high-dimensional feature space, and a linear regression model is constructed in the high-dimensional feature space to establish a nonlinear mapping relationship between meteorological parameters and the conversion power of photovoltaic power stations;

[0096] The time-series correlation constraint term is introduced into the objective function of the least squares support vector regression machine. The similarity between samples in the feature space after mapping is measured by the kernel function, so that samples that are close in time and have similar features have similar weight coefficients, thereby describing the time-dependent characteristics of meteorological parameters.

[0097] Construct the optimization objective function of the least squares support vector regression machine with time-series correlation constraint term, which consists of three parts: empirical risk term, structural risk term and time-series correlation constraint term;

[0098] The whale optimization algorithm is used to search for the optimal parameter combination of the least squares support vector regression machine model, and the regularization parameter, time series correlation constraint weight coefficient and kernel function scale parameter of the model are optimized by updating the position of the whale population; the least squares support vector regression machine model is trained on all training data using the optimal parameter combination obtained by the whale optimization algorithm, and the meteorological parameters at the time to be predicted are input into the trained least squares support vector regression machine model to obtain the corresponding photovoltaic power station conversion power prediction value, forming a nonlinear mapping relationship model between meteorological data and photovoltaic power station conversion power;

[0099] By introducing time-series correlation constraints into the objective function and using the whale optimization algorithm for parameter optimization, the ability of the least squares support vector regression model to characterize the nonlinear relationship between meteorological parameters and conversion power is improved, thereby improving the accuracy and stability of photovoltaic power station conversion power prediction.

[0100] For example, in order to construct a nonlinear mapping relationship model between meteorological data and photovoltaic power station conversion power, it is necessary to preprocess the historical operation data of the photovoltaic power station. This paper selects the operation data of a photovoltaic power station for one year, with a data sampling interval of 15 minutes, including parameters such as all-day radiation intensity, scattered radiation intensity, ambient temperature, relative humidity of air and conversion power.

[0101] The original data was cleaned to remove missing values, outliers, and unavailable values. Missing data were filled using linear interpolation, and outliers were identified and corrected using local mean and standard deviation. Then, the cleaned data was normalized and each parameter was mapped to the interval [0,1] to eliminate the dimension effect and numerical range differences.

[0102] Considering that the impact of meteorological parameters on the conversion power of photovoltaic power stations has a certain lag, the historical data is divided according to the time window to construct the input feature vector. Taking the prediction of the conversion power in the next hour as an example, the meteorological parameters 3 hours before the current moment are selected as the input features, and each 30 minutes is a time step. The input feature vector consists of 4 meteorological parameters of 6 time steps, a total of 24 features. The output label is the conversion power 1 hour after the current moment.

[0103] By sliding the time window, the input feature vector and the corresponding output label are generated to form the training data set.

[0104] The traditional support vector regression machine constructs a regression model through the ε-insensitive loss function and the principle of structural risk minimization. For nonlinear problems, it is necessary to introduce a kernel function to map samples to a high-dimensional feature space and perform linear regression in the high-dimensional space. However, the standard SVR does not consider the temporal correlation between samples. The contributions of samples adjacent in time to the output should be more similar.

[0105] In order to characterize the time-dependent characteristics of meteorological parameters, this paper introduces a time-series correlation constraint term into the objective function of the least squares support vector regression machine (LS-SVR). LS-SVR uses the square loss function instead of the ε-insensitive loss function to transform the inequality constraint into an equality constraint, which simplifies the computational complexity. The time-series correlation constraint term measures the similarity between samples in the feature space after mapping through the kernel function, so that samples that are close in time and have similar features have similar weight coefficients.

[0106] Considering the complex nonlinear relationship between meteorological parameters and conversion power, Gaussian radial basis kernel function (RBF) is selected as the kernel function of LS-SVR. The feature space dimension corresponding to RBF kernel function is infinite and has good nonlinear mapping ability.

[0107] The timing correlation constraint term is embedded into the objective function of LS-SVR, and the optimization objective function of the improved LS-SVR model proposed in this paper is obtained.

[0108] The objective function consists of three parts: the first is structural risk, which controls the complexity of the model; the second is empirical risk, which measures the degree to which the model fits the training data; the third is the time series correlation constraint, which makes the weight coefficients of adjacent samples close.

[0109] The improved LS-SVR model contains multiple parameters that need to be optimized. These parameters have a significant impact on the generalization performance of the model and need to be reasonably selected. Traditional parameter tuning methods such as grid search and cross-validation are computationally intensive and prone to fall into local optimality.

[0110] This paper uses the Whale Optimization Algorithm (WOA) to intelligently search and optimize the parameters of the LS-SVR model. WOA simulates the predation behavior of whales and uses spiral position update and shrinking surround strategy to achieve global search and local development.

[0111] First, the parameters to be optimized are encoded as the position vectors of individual whales, and the whale population is initialized. Each individual whale represents a set of parameter combinations. The LS-SVR model is trained according to the parameter combinations, and the prediction accuracy of the model is calculated on the validation set as the individual fitness value. The goal of the WOA algorithm is to find the whale individual with the highest fitness value, that is, the corresponding optimal parameter combination.

[0112] During the search process, WOA updates the position of the whale population according to the current optimal solution. By randomly selecting the search strategy, global exploration and local development are balanced. When the conditions are met, the whale will spirally update the position in the direction of the current optimal solution to strengthen the local search; otherwise, the whale will randomly search near the optimal solution to expand the exploration range. At the same time, an adaptive weight coefficient is introduced to adjust the search step size, maintaining a larger search step size for global exploration in the early stage of iteration, and reducing the search step size for fine local search in the later stage.

[0113] By setting the maximum number of iterations, WOA continuously updates the position of the whale population until the termination condition is reached. The parameter combination corresponding to the optimal whale individual obtained in the end is the optimal parameter setting of the LS-SVR model.

[0114] The optimal parameter combination obtained by WOA optimization is used to retrain the LS-SVR model on all training data to obtain the final nonlinear mapping model between meteorological data and photovoltaic power station conversion power. In the prediction stage, the meteorological parameters at the time to be predicted are input into the trained LS-SVR model to obtain the corresponding photovoltaic power station conversion power prediction value.

[0115] By introducing the time-series correlation constraint term, the LS-SVR model can better characterize the nonlinear relationship and time-dependent characteristics between meteorological parameters and conversion power. Using WOA for parameter optimization can adaptively search for the optimal parameter combination and improve the prediction accuracy and generalization performance of the model. The improved LS-SVR model proposed in this paper combines data-driven and physical characteristics, providing a new idea and method for photovoltaic power station conversion power prediction.

[0116] In an optional embodiment,

[0117] The steps of using the whale optimization algorithm to search for the optimal parameter combination of the least squares support vector regression model and optimizing the regularization parameters, time series correlation constraint weight coefficients and kernel function scale parameters of the model by updating the position of the whale population include:

[0118] The regularization parameters, time-series correlation constraint weights and kernel function scale parameters in the least squares support vector regression model are encoded into the position coordinates of the whales in the whale optimization algorithm, and the initial positions of a group of whales are randomly generated to form an initial population.

[0119] For each whale in the population, decode its position coordinates into least squares support vector regression model parameters, train the least squares support vector regression model and calculate the prediction error on the validation set as the fitness value of the whale;

[0120] The whale with the highest fitness value in the current population is taken as the leader whale, and the position of each whale in the population is updated by adopting the siege predation mode or the spiral attack mode according to the probability parameter; in the siege predation mode, the whale calculates the coefficient vector according to the linear decreasing variable and the random vector, and updates its own position based on the position and coefficient vector of the leader whale; in the spiral attack mode, the whale updates its own position according to the distance between it and the leader whale and the spiral function to enhance the global exploration ability of the algorithm;

[0121] For whales whose updated positions are out of the range, their positions are adjusted to the boundary of the parameters, and it is determined whether the termination conditions of the maximum number of iterations or the performance improvement threshold are met; if the termination conditions are not met, the fitness value calculation step is returned to continue the iterative optimization; if the termination conditions are met, the position coordinates of the leading whale are decoded into the optimal parameter combination of the least squares support vector regression machine.

[0122] For example, in order to further improve the performance of the photovoltaic power station conversion power prediction model, this paper introduces the whale optimization algorithm to optimize the key parameters of the model. The whale optimization algorithm is a novel and efficient swarm intelligence optimization algorithm. By simulating the predation behavior of whales, it performs global search and local development in the solution space to quickly find the optimal solution. The specific steps of applying the whale optimization algorithm to the optimization of the prediction model parameters are as follows:

[0123] The model parameters to be optimized are mapped to the search variables of the whale optimization algorithm to form the solution space of the optimization problem. Taking the support vector regression machine as an example, the parameters that need to be optimized include the penalty coefficient, the kernel function type and its corresponding kernel parameter. For these three parameters, the real number encoding method is used to represent each whale individual as a three-dimensional real number vector, corresponding to the penalty coefficient, the kernel function type index, and the kernel parameter respectively. The value range of the penalty coefficient is set to 0.1 to 100, and the value range of the kernel function type index is set to 0 to 1, where less than 0.5 represents a linear kernel function and greater than or equal to 0.5 represents a radial basis kernel function. The kernel parameter of the radial basis kernel function ranges from 0.01 to 10. For example, a whale individual is encoded as (12.8, 0.3, 1.5), which means that the penalty coefficient is 12.8, the linear kernel function is selected, and the kernel parameter is 1.5.

[0124] The performance evaluation index of the prediction model is used as the fitness function of the whale optimization algorithm to guide the algorithm to search for the optimal parameter combination. This paper selects the root mean square error (RMSE) of the prediction model on the validation set as the fitness function value. The smaller the RMSE, the better the model performance and the better the corresponding parameter combination. The RMSE of the prediction model under the corresponding parameter combination of each whale individual is calculated through cross-validation, and it is used as the fitness value of the individual. For example, for the whale individual coded as (12.8, 0.3, 1.5), the support vector regression machine is trained using a penalty coefficient of 12.8, a linear kernel function, and a kernel parameter of 1.5. The RMSE calculated on the validation set is 0.18, and the fitness value of the individual is 0.18.

[0125] Initialize the whale population and randomly generate a certain number of whale individuals, each of which corresponds to a set of parameters to be optimized. Calculate the fitness value of each whale in the population, and record the current optimal whale individual and the global optimal whale individual. During the iterative optimization process, each whale updates its position at the next moment based on its own position, the current optimal whale position, and the global optimal whale position. The position update strategy includes three behaviors: siege predation, bubble network attack, and exploration search. When siege predation, the whale individual moves closer to the current optimal whale position; when bubble network attack, the whale individual searches for a better solution on a spiral path with the current optimal whale as the center; when exploring and searching, the whale individual randomly walks in the global range, expands the search range, and jumps out of the local optimum. Through iterative optimization, the position of the whale individual is continuously updated, so that the population gradually converges to the global optimal solution.

[0126] Assume that the number of whale population is 50, the maximum number of iterations is 100, and the selection probabilities of siege predation, bubble network attack, and exploration search are 0.6, 0.3, and 0.1 respectively at each position update. In the initial population, the global optimal whale individual code is (10.2, 0.8, 2.1), and the fitness value is 0.15. After 100 iterations of optimization, the global optimal whale individual code is updated to (15.6, 0.7, 1.8), and the fitness value is reduced to 0.12. The corresponding optimal parameter combination is a penalty coefficient of 15.6, a radial basis kernel function, and a kernel parameter of 1.8.

[0127] The prediction model was retrained using the optimal parameter combination obtained by the whale optimization algorithm, and the performance was evaluated on the test set. The performance of the optimized model was compared with the model performance under the default parameters to quantitatively analyze the improvement effect of the whale optimization algorithm on the model performance. Taking the support vector regression machine as an example, the RMSE of the model trained with the default parameters (penalty coefficient 1.0, radial basis kernel function, kernel parameter 1.0) on the test set was 0.21, and the mean absolute percentage error (MAPE) was 5.8%. The RMSE of the model trained with the optimal parameter combination obtained by the whale optimization algorithm (penalty coefficient 15.6, radial basis kernel function, kernel parameter 1.8) on the test set was reduced to 0.18, and the MAPE was reduced to 4.5%, and the prediction performance was significantly improved.

[0128] The parameters of the prediction model are optimized by the whale optimization algorithm, which fully explores the performance potential of the model. Without changing the model structure, the prediction accuracy and generalization ability of the model are improved by parameter tuning. This optimization strategy can be widely used in parameter optimization problems of various machine learning models, providing a useful reference for photovoltaic power generation prediction and even modeling practices in other fields.

[0129] In an optional embodiment, the method further includes:

[0130] An adaptive hybrid search strategy is introduced to dynamically adjust the selection probability of the siege predation mode and the spiral attack mode. The adaptive weight coefficient w is designed to represent the probability of selecting the siege predation mode. The probability of selecting the spiral attack mode is 1-w. The value of w is adaptively adjusted according to the current number of iterations t and the population diversity index d. The calculation formula is as follows:

[0131] ;

[0132] Among them, T max is the maximum number of iterations, d max is the maximum value of population diversity, and the population diversity index d is measured by the average Euclidean distance between individuals in the population:

[0133] ;

[0134] Where N represents the total number of samples in the data set, X represents the samples in the data set, i and j represent the index numbers of the samples, which are used to traverse and compare the samples pairwise. i -X j || represents the distance between the i-th sample and the j-th sample;

[0135] As the iteration proceeds, the value of w gradually decreases, indicating that the probability of selecting the spiral attack mode increases, and the local search capability is enhanced; w is related to the population diversity index d. When the population diversity is low, the value of w is relatively large, indicating that the probability of selecting the siege predation mode increases, and the global exploration capability is enhanced to avoid the algorithm from converging prematurely;

[0136] In the siege predation mode and spiral attack mode, an adaptive parameter control mechanism is introduced to dynamically adjust the algorithm parameters and improve the optimization performance; for the shrinkage factor a in the siege predation mode, an adaptive update strategy is designed:

[0137] ;

[0138] Among them, a min and a max are the minimum and maximum values ​​of a respectively, and k is a control parameter used to adjust the decay speed of a. As the iteration proceeds, the value of a changes from a to max Gradually decrease to a min , which enables the algorithm to have stronger global exploration capabilities in the early stage and stronger local development capabilities in the later stage;

[0139] For the spiral shape parameter b in the spiral attack mode, an adaptive update strategy is designed:

[0140] ;

[0141] Among them, b min and b max are the minimum and maximum values ​​of b respectively. As the iteration proceeds, the value of b changes from b max Gradually decrease to b min , so that the algorithm performs global search with a larger spiral step size in the early stage and local search with a smaller spiral step size in the later stage, improving the convergence speed and accuracy;

[0142] A global exploration strategy based on Lévy flight is introduced. Lévy flight exploration is performed with a certain probability in each iteration. The random step size that obeys the Lévy distribution is generated by the Lévy flight function to update the current solution, so as to enhance the algorithm's ability to escape from the local optimum. During the iteration process, the siege and predation mode and the spiral attack mode are dynamically selected through an adaptive hybrid search strategy, and the algorithm parameters are adjusted using an adaptive parameter control mechanism. At the same time, the Lévy flight strategy is combined for global exploration to balance the development and utilization of the algorithm.

[0143] For example, in order to further improve the performance of the whale optimization algorithm, this paper introduces an adaptive hybrid search strategy to dynamically adjust the selection probability of the siege predation mode and the spiral attack mode, and balance the global exploration and local development capabilities of the algorithm. At the same time, an adaptive parameter control mechanism is introduced in the siege predation mode and the spiral attack mode to dynamically adjust the algorithm parameters and improve the optimization efficiency. In addition, this paper also introduces a global exploration strategy based on Levy flight to enhance the algorithm's ability to jump out of the local optimum.

[0144] Design the adaptive weight coefficient w, which represents the probability of selecting the siege predation mode, and the probability of selecting the spiral attack mode is 1-w. The value of w is adaptively adjusted according to the current iteration number t and the population diversity index d.

[0145] In the calculation formula of w, T max is the maximum number of iterations, d max is the maximum value of population diversity. Let T max =100,d max = 10. The population diversity index d is measured by the average Euclidean distance between individuals in the population. Assume that the current population is pop = {X1, X2, ..., X N}, the population size N = 50, then in the calculation formula of d, X i and X j are the position vectors of the i-th and j-th individuals in the population, and pop is the population.

[0146] As the iteration proceeds, the value of w gradually decreases, indicating that the probability of selecting the spiral attack mode increases, and the local search capability is enhanced. At the same time, w is related to the population diversity index d. When the population diversity is low, the value of w is relatively large, indicating that the probability of selecting the siege predation mode increases, and the global exploration capability is enhanced to avoid premature convergence of the algorithm.

[0147] In the siege predation mode, an adaptive parameter control mechanism is introduced to dynamically adjust the value of the shrinkage factor a. The value range of a is [a min , a max ], let a min =0.1,a max = 2. The control parameter k is used to adjust the decay rate of a, and its value is usually 5.

[0148] As the iteration progresses, the value of a changes from a to max Gradually decrease to a min , so that the algorithm has a strong global exploration capability in the early stage and a strong local development capability in the later stage. For example, the current iteration number t=30, the maximum iteration number T max =100, then the calculation process of a is as follows:

[0149] t / T max =0.3 is substituted into the adaptive update formula of a, and we get:

[0150] a = 2 - 1.9 · (1 - exp(-5 · 0.3)) = 0.7525;

[0151] Therefore, in the current iteration, the shrinkage factor a of the siege-predation mode is 0.7525.

[0152] In the spiral attack mode, an adaptive parameter control mechanism is introduced to dynamically adjust the value of the spiral shape parameter b. The value range of b is [b min , b max ], let b min =0.1,b max =1.

[0153] As the iterations proceed, the value of b changes from b max Gradually decrease to b min , so that the algorithm performs global search with a larger spiral step size in the early stage and local search with a smaller spiral step size in the later stage, improving the convergence speed and accuracy. For example, the current iteration number t=30, the maximum iteration number T max =100, then the calculation process of b is as follows:

[0154] t / T max =0.3 is substituted into the adaptive update formula of b, and we get:

[0155] b = 1 - 0.9 · (1 - exp(-5 · 0.3)) = 0.3975;

[0156] Therefore, in the current iteration, the spiral shape parameter b of the spiral attack mode takes the value of 0.3975.

[0157] In each iteration, the Levy flight exploration is performed with a certain probability. The Levy flight function is used to generate random steps that obey the Levy distribution and update the current solution to enhance the algorithm's ability to jump out of the local optimum.

[0158] By introducing an adaptive hybrid search strategy, dynamically selecting the siege predation mode and the spiral attack mode, and using the adaptive parameter control mechanism to adjust the algorithm parameters, the global exploration and local development capabilities of the algorithm can be balanced to improve the optimization efficiency. At the same time, combined with the Levy flight strategy for global exploration, the algorithm's ability to jump out of the local optimum is further enhanced to avoid premature convergence. The improved whale optimization algorithm proposed in this paper has stronger optimization performance and convergence speed in the parameter optimization of the photovoltaic power station conversion power prediction model.

[0159] In an optional embodiment,

[0160] The conversion power data of the photovoltaic power station conversion power prediction model is adaptively decomposed using the variational mode decomposition algorithm to extract the multi-scale periodic characteristics and local fluctuation characteristics of the photovoltaic power station conversion power. The extracted characteristic variables and meteorological parameters are used as the input of the least squares support vector regression machine. The steps of constructing a photovoltaic power station conversion power prediction model driven by multi-source heterogeneous data include:

[0161] The original conversion power sequence of the photovoltaic power station is preprocessed to remove outliers and missing values, and the power data is smoothed and normalized; according to the spectral characteristics of the preprocessed conversion power sequence, the number and penalty parameters of the intrinsic mode function components in the variational mode decomposition are preliminarily determined;

[0162] The preprocessed power sequence is adaptively decomposed using the variational mode decomposition algorithm. The original power sequence is decomposed into several intrinsic mode function components by solving the constraint optimization problem of variational mode decomposition.

[0163] Hilbert transform is performed on each intrinsic mode function component obtained by variational mode decomposition to extract the envelope and instantaneous frequency characteristics of each component, and the multi-scale periodic characteristics and local fluctuation characteristics of the conversion power sequence are reconstructed according to the frequency characteristics and energy distribution of each component; the reconstructed multi-scale periodic characteristics and local fluctuation characteristics are integrated with the meteorological parameters of the photovoltaic power station to form a multi-source heterogeneous feature set as the input of the photovoltaic power station conversion power prediction model.

[0164] For example, in order to deeply explore the inherent law of photovoltaic power station conversion power, this paper proposes a feature extraction method based on variational mode decomposition. Variational mode decomposition is an adaptive signal processing technology that can decompose complex non-stationary signals into multiple stable intrinsic mode functions and a residual term to achieve signal time-frequency domain feature extraction. The specific steps of applying variational mode decomposition to photovoltaic power station conversion power feature extraction are as follows:

[0165] The historical conversion power data of photovoltaic power stations are preprocessed to remove outliers and missing values, and the data are normalized to map them between 0 and 1. Taking the hourly conversion power data of a photovoltaic power station for 30 consecutive days as an example, there are 720 data points in total. After preprocessing, a normalized conversion power time series with a length of 720 is obtained, which is recorded as the original signal.

[0166] The preprocessed conversion power time series is subjected to variational mode decomposition, which is adaptively decomposed into several intrinsic mode functions and a residual term. The core idea of ​​variational mode decomposition is to solve the constrained variational model through iterative optimization, so that the decomposed intrinsic mode function has good bandwidth properties in the envelope spectrum domain, while maintaining the orthogonality between the eigenmode functions.

[0167] Initialize the first intrinsic mode function as the original signal, and continuously update the center frequency and bandwidth of each intrinsic mode function through iterative optimization, so that its envelope spectrum approaches the theoretical optimal value. In each iteration, calculate the envelope spectrum of the current intrinsic mode function through Hilbert transform, construct the penalty term function, and use the gradient descent method to solve the constrained variational model and update the intrinsic mode function. Iterate the optimization until the center frequency and bandwidth of all intrinsic mode functions converge, or reach the preset maximum number of iterations.

[0168] Subtract the decomposed intrinsic mode function from the original signal to get the residual term. Repeat the above process, use the residual term as the new original signal, and continue to decompose the next intrinsic mode function until the residual term becomes a monotonic function or its energy is lower than the preset threshold.

[0169] Assuming that the conversion power time series is subjected to variational mode decomposition, a total of 6 intrinsic mode functions and 1 residual term are obtained. The first intrinsic mode function extracts the high-frequency fluctuation information in the original signal, corresponding to the random fluctuation component of the photovoltaic power station conversion power; the second to fifth intrinsic mode functions extract the medium-frequency oscillation information in the original signal, corresponding to the periodic fluctuation component of the photovoltaic power station conversion power; the sixth intrinsic mode function extracts the low-frequency trend information in the original signal, corresponding to the long-term trend component of the photovoltaic power station conversion power; the residual term corresponds to the constant bias component of the photovoltaic power station conversion power.

[0170] The eigenmode functions and residual terms obtained by variational mode decomposition are analyzed to extract the characteristic quantities that have an important impact on the conversion power prediction. For each eigenmode function, its statistical characteristics, such as mean, variance, skewness, kurtosis, etc., are calculated to characterize its fluctuation law and distribution characteristics; for the envelope spectrum of each eigenmode function, its energy characteristics, such as total energy, frequency band energy ratio, etc., are calculated to characterize its energy distribution on different frequency scales. For the residual term, its mean and slope are calculated to characterize its constant bias and slow-changing trend.

[0171] Through feature importance evaluation and cross-validation, the feature subsets that contribute more to the prediction of conversion power are selected as the input of the subsequent prediction model. Taking the first eigenmode function as an example, the mean and variance of its statistical features contribute the most to the prediction of conversion power, so these two features are selected as input features. Similarly, 2 to 3 important features are selected from the second to the sixth eigenmode functions, and the mean feature is selected from the residual term, and finally a feature subset containing 12 feature quantities is obtained.

[0172] The feature subset extracted by variational mode decomposition is used to build a photovoltaic power station conversion power prediction model. The feature subset is divided into training set, validation set and test set, accounting for 60%, 20% and 20% of the total data respectively. The prediction model is trained on the training set, and the validation set is used for hyperparameter tuning and model selection. Finally, the prediction performance of the model is evaluated on the test set.

[0173] Taking the support vector regression machine as an example, the radial basis kernel function is used to optimize the penalty coefficient and kernel parameters through grid search, and the optimal parameter combination is a penalty coefficient of 10.0 and a kernel parameter of 0.1. The root mean square error of the support vector regression machine trained with this parameter combination on the test set is 0.05, and the mean absolute percentage error is 2.1%. The prediction performance is better than the original model without feature extraction.

[0174] The photovoltaic power station conversion power feature extraction method based on variational mode decomposition makes full use of the multi-scale time-frequency domain characteristics of the conversion power signal, adaptively extracts the intrinsic mode function that reflects its inherent fluctuation law, and obtains information-rich feature subsets through feature analysis and selection, effectively improving the performance of subsequent prediction models. This method provides a novel feature engineering idea for photovoltaic power station conversion power prediction and has good engineering application prospects.

[0175] In an optional embodiment,

[0176] After randomly shuffling the training set data, the established photovoltaic power station conversion power prediction model is trained and verified by cross-validation. Regularization and early stopping strategies are used to avoid overfitting of the photovoltaic power station conversion power prediction model. The steps of obtaining the optimal hyperparameters of the photovoltaic power station conversion power prediction model through grid search include:

[0177] The training set data of the photovoltaic power station conversion power prediction model is randomly shuffled, and the photovoltaic power station conversion power prediction model is trained and verified by K-fold cross-validation. The training set is divided into K mutually exclusive subsets, and a different subset is selected as the verification set each time. The remaining subsets constitute the training set. The photovoltaic power station conversion power prediction model is trained with the training set, and the performance of the photovoltaic power station conversion power prediction model is evaluated on the verification set.

[0178] The L2 norm penalty term of the parameter vector is introduced into the objective function of the photovoltaic power station conversion power prediction model. The size of the photovoltaic power station conversion power prediction model parameters is constrained by L2 regularization, the complexity of the photovoltaic power station conversion power prediction model is controlled, and the generalization ability of the photovoltaic power station conversion power prediction model is improved when the training data is insufficient.

[0179] After each training round, the performance of the photovoltaic power plant conversion power prediction model on the validation set is evaluated. When the performance of the validation set has not improved for many consecutive times, the early stopping strategy is used to terminate the training process in advance to prevent the photovoltaic power plant conversion power prediction model from overfitting the training data.

[0180] The hyperparameters of the photovoltaic power station conversion power prediction model are optimized through grid search, and the preset hyperparameter combinations are exhaustively searched. Cross-validation is used to evaluate the performance of the photovoltaic power station conversion power prediction model corresponding to each hyperparameter combination, and the hyperparameter combination with the best performance in the validation set is selected as the final parameter configuration of the photovoltaic power station conversion power prediction model.

[0181] For example, in order to train a photovoltaic power station conversion power prediction model with excellent performance and robustness, this paper uses a combination of cross-validation and hyperparameter optimization to train the model. Cross-validation can make full use of limited data and robustly evaluate model performance by dividing the data set multiple times and exchanging training and testing roles. Hyperparameter optimization can find the optimal hyperparameter combination by systematically searching the hyperparameter space of the model to achieve the best model performance. The specific steps of combining cross-validation and hyperparameter optimization for prediction model training are as follows:

[0182] The complete feature data set is randomly divided into a training set and a test set. The training set is used for model training and hyperparameter optimization, and the test set is used for model performance evaluation. Considering the large amount of data, the training set and the test set are divided in a ratio of 8:2. Taking the hourly conversion power data of a photovoltaic power station for one year as an example, there are 8760 data samples in total, and each sample contains 20 feature variables. It is randomly divided into 7008 samples in the training set and 1752 samples in the test set.

[0183] Perform K-fold cross validation on the training set data, divide it into K subsets, select one of the subsets as the validation set each time, and the remaining K-1 subsets as the training set, train the model and evaluate the performance on the validation set. Repeat K times so that each subset has the opportunity to be used as the validation set, and finally average the K performance evaluation results to obtain the cross-validation performance of the model.

[0184] Taking 5-fold cross validation as an example, the training set of 7008 samples is divided into 5 subsets, each with 1402 samples. The first time, the first subset is used as the validation set, and the remaining 4 subsets are used as training sets. The model is trained and the root mean square error (RMSE) calculated on the validation set is 0.08; the second time, the second subset is used as the validation set, and the remaining 4 subsets are used as training sets. The model is trained and the RMSE calculated on the validation set is 0.07; and so on, after completing 5 cross validations, the 5 RMSE values ​​are 0.08, 0.07, 0.09, 0.06, and 0.08, respectively, and the average cross validation RMSE is 0.076.

[0185] Optimize the hyperparameters of the prediction model and find the optimal hyperparameter combination in the specified hyperparameter space through grid search or random search. Taking the support vector regression machine as an example, the hyperparameters that need to be optimized include the penalty coefficient, the kernel function type and its corresponding kernel parameter.

[0186] The grid search method is used, and the penalty coefficient is selected in [0.1, 1, 10, 100], and the kernel function type is selected in ['linear', 'rbf']. Among them, linear is a linear kernel function, which means finding the optimal linear decision boundary in the original feature space, and rbf is a radial basis kernel function, which means mapping the input features to an infinite-dimensional Gaussian space and finding the optimal decision boundary through nonlinear transformation. The kernel parameter of the radial basis kernel function is selected in [0.01, 0.1, 1, 10], with a total of 4×2×4=32 hyperparameter combinations. For each hyperparameter combination, a 5-fold cross-validation is used to evaluate its performance, and the hyperparameter combination with the smallest average cross-validation RMSE is selected as the optimal hyperparameter.

[0187] Assuming that the optimal hyperparameter combination obtained by grid search is a penalty coefficient of 10, a radial basis kernel function, and a kernel parameter of 0.1, the corresponding average cross-validation RMSE is 0.065, which is better than the RMSE of 0.076 under the default hyperparameters.

[0188] Using the optimal hyperparameter combination, the prediction model is retrained on the complete training set to obtain the final model parameters. The trained model is applied to the test set, and its performance evaluation indicators on the test set are calculated, including RMSE, mean absolute error (MAE), mean absolute percentage error (MAPE), etc., to comprehensively evaluate the prediction performance of the model.

[0189] Taking the support vector regression machine as an example, the model was trained on the complete training set using the optimal hyperparameter combination of penalty coefficient 10, radial basis kernel function, and kernel parameter 0.1 to obtain the final support vector, support vector coefficient, and bias term. The trained model was applied to 1752 samples in the test set to predict their conversion power values ​​and compare them with the actual values. The calculated RMSE was 0.06, MAE was 0.05, and MAPE was 2.8%, all of which were better than the performance of the model without hyperparameter optimization.

[0190] Through the above steps, the photovoltaic power station conversion power prediction model is trained and verified using cross-validation and early stopping strategies, and the optimal hyperparameters of the prediction model are obtained by combining grid search, which can effectively improve the generalization performance and robustness of the prediction model and avoid overfitting the model to the training data. In practical applications, the number of cross-validation folds, the number of tolerance rounds of the early stopping method, and the search range of hyperparameters can be flexibly selected according to the characteristics of the data set and the complexity of the model to balance model performance and computational efficiency.

[0191] Figure 2 FIG. 1 is a schematic diagram of a conversion power prediction system considering air humidity according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0192] The first unit is used to obtain the historical meteorological data of the target photovoltaic power station and the corresponding actual conversion power data, the historical meteorological data includes the radiation intensity of the whole day, the scattered radiation intensity, the ambient temperature and the relative humidity of the air, sort the historical meteorological data according to the time series, and identify and correct the missing values ​​and abnormal values; extract the design parameters of the photovoltaic power station, the design parameters include the type of photovoltaic modules, the installation angle, the installation spacing, and the total installed capacity;

[0193] The second unit is used to construct a nonlinear mapping model between relative air humidity and photovoltaic module conversion efficiency. Through the control variable experiment, the photovoltaic module conversion efficiency data under different relative air humidity levels are obtained, and the nonlinear relationship curve between the photovoltaic module conversion efficiency and the relative air humidity is fitted by the support vector regression machine. Based on the pre-processed historical meteorological data and conversion power data of the photovoltaic power station, combined with the design parameters of the photovoltaic power station, the least squares support vector regression machine is used to construct a nonlinear mapping relationship model between the historical meteorological data and the conversion power of the photovoltaic power station. The nonlinear mapping model between relative air humidity and photovoltaic module conversion efficiency is coupled with the nonlinear mapping relationship model between historical meteorological data and the conversion power of the photovoltaic power station to form a photovoltaic power station conversion power prediction model.

[0194] The third unit is used to use the variational mode decomposition algorithm to adaptively decompose the conversion power data of the photovoltaic power station conversion power prediction model, extract the multi-scale periodic characteristics and local fluctuation characteristics of the photovoltaic power station conversion power, and use the extracted characteristic variables and meteorological parameters as the input of the least squares support vector regression machine to build a photovoltaic power station conversion power prediction model driven by multi-source heterogeneous data; after randomly shuffling the training set data, the established photovoltaic power station conversion power prediction model is trained and verified by cross-validation, and the regularization and early stopping strategies are used to avoid overfitting of the photovoltaic power station conversion power prediction model. The optimal hyperparameters of the photovoltaic power station conversion power prediction model are obtained through grid search, and the prediction sub-models obtained by training with multiple different random seeds are weighted combined using an integrated learning strategy to improve the robustness and generalization performance of the photovoltaic power station conversion power prediction model.

[0195] According to a third aspect of the embodiments of the present invention,

[0196] An electronic device is provided, comprising:

[0197] processor;

[0198] a memory for storing processor-executable instructions;

[0199] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0200] A fourth aspect of the embodiments of the present invention is:

[0201] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0202] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0203] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A conversion power prediction method considering air humidity, characterized in that: include: Obtain the historical meteorological data of the target photovoltaic power station and the corresponding actual conversion power data, the historical meteorological data including the all-day radiation intensity, scattered radiation intensity, ambient temperature and relative air humidity, sort the historical meteorological data according to the time series, and identify and correct missing values ​​and abnormal values; extract the design parameters of the photovoltaic power station, the design parameters including the type of photovoltaic modules, installation angle, installation spacing, and total installed capacity; Construct a nonlinear mapping model between relative air humidity and photovoltaic module conversion efficiency, obtain photovoltaic module conversion efficiency data under different relative air humidity levels through controlled variable experiments, and use support vector regression machine to fit the nonlinear relationship curve between photovoltaic module conversion efficiency and relative air humidity; Based on the pre-processed historical meteorological data and conversion power data of the photovoltaic power station, combined with the design parameters of the photovoltaic power station, use the least squares support vector regression machine to construct a nonlinear mapping relationship model between historical meteorological data and photovoltaic power station conversion power, and couple the nonlinear mapping model between relative air humidity and photovoltaic module conversion efficiency with the nonlinear mapping relationship model between historical meteorological data and photovoltaic power station conversion power to form a photovoltaic power station conversion power prediction model; The conversion power data of the photovoltaic power station conversion power prediction model is adaptively decomposed by the variational mode decomposition algorithm, and the multi-scale periodic characteristics and local fluctuation characteristics of the photovoltaic power station conversion power are extracted. The extracted characteristic variables and meteorological parameters are used as the input of the least squares support vector regression machine to construct a photovoltaic power station conversion power prediction model driven by multi-source heterogeneous data; the training set data is randomly shuffled and the established photovoltaic power station conversion power prediction model is trained and verified by cross-validation. Regularization and early stopping strategies are used to avoid overfitting of the photovoltaic power station conversion power prediction model. The optimal hyperparameters of the photovoltaic power station conversion power prediction model are obtained through grid search, and the prediction sub-models trained with multiple different random seeds are weighted combined using an ensemble learning strategy to improve the robustness and generalization performance of the photovoltaic power station conversion power prediction model.

2. The method according to claim 1, characterized in that The nonlinear mapping model between relative air humidity and photovoltaic module conversion efficiency is constructed, and the photovoltaic module conversion efficiency data under different relative air humidity levels are obtained through the control variable experiment. The steps of fitting the nonlinear relationship curve between photovoltaic module conversion efficiency and relative air humidity by support vector regression machine include: The conversion efficiency data of photovoltaic modules under different relative air humidity conditions were obtained through controlled variable experiments. During the experiment, the relative air humidity in the experimental chamber was adjusted by a humidifier and a dehumidifier, and multiple humidity gradients were set. The electrical parameters and environmental parameters of the photovoltaic modules were continuously monitored under each humidity gradient, and other environmental factors were kept constant to eliminate interference. The real-time conversion efficiency of the photovoltaic modules under different relative air humidity was calculated using open circuit voltage, short circuit current, maximum power point voltage and current parameters. For each type of photovoltaic module, the average conversion efficiency of multiple measurements under different relative air humidity levels is extracted to construct an experimental data set. Considering the significant nonlinear relationship between relative air humidity and photovoltaic module conversion efficiency, support vector regression machine is used for curve fitting. The support vector regression machine realizes nonlinear mapping from the original space to the high-dimensional feature space by introducing the kernel function, constructs a linear regression model in the feature space, introduces slack variables and ε-insensitive loss function, and establishes the optimization objective equation of the support vector regression machine. The Gaussian radial basis kernel function is selected as the kernel function of the support vector regression machine. The Lagrange multiplier method and the Carlo-Kuhn-Tucker condition are used to transform the optimization objective equation into a dual problem for solution. The support vector, Lagrange multiplier and bias term parameters of the support vector regression machine model are obtained through training, and the nonlinear regression function of the photovoltaic module conversion efficiency with respect to the relative humidity of the air is constructed. The regularization coefficient and kernel function bandwidth parameters of the support vector regression model are optimized by combining grid search with cross validation. The nonlinear function relationship curve between the conversion efficiency of photovoltaic modules and the relative humidity of air is fitted using the optimized support vector regression model.

3. The method according to claim 1, characterized in that Based on the pre-processed historical meteorological data and conversion power data of the photovoltaic power station, combined with the design parameters of the photovoltaic power station, the steps of using the least squares support vector regression mechanism to construct a nonlinear mapping relationship model between the historical meteorological data and the conversion power of the photovoltaic power station include: The historical operation data of the photovoltaic power station is cleaned and normalized, and the input feature vector is constructed according to the time-lag characteristics of the parameters of all-day radiation intensity, scattered radiation intensity, ambient temperature and relative humidity of the air; a nonlinear mapping function is introduced to map the input feature vector to a high-dimensional feature space, and a linear regression model is constructed in the high-dimensional feature space to establish a nonlinear mapping relationship between meteorological parameters and the conversion power of the photovoltaic power station; The time series correlation constraint term is introduced into the objective function of the least squares support vector regression machine. The similarity between samples in the feature space after mapping is measured by the kernel function, so that samples close in time and with similar features have similar weight coefficients, thereby describing the time dependence characteristics of meteorological parameters. Construct the optimization objective function of the least squares support vector regression machine with time-series correlation constraint term. The objective function consists of three parts: empirical risk term, structural risk term and time-series correlation constraint term. The whale optimization algorithm is used to search for the optimal parameter combination of the least squares support vector regression model, and the regularization parameter, time series correlation constraint weight coefficient and kernel function scale parameter of the model are optimized by updating the position of the whale population; the least squares support vector regression model is trained on all training data using the optimal parameter combination obtained by the whale optimization algorithm, and the meteorological parameters at the time to be predicted are input into the trained least squares support vector regression model to obtain the corresponding photovoltaic power station conversion power prediction value, forming a nonlinear mapping relationship model between meteorological data and photovoltaic power station conversion power; By introducing time-series correlation constraints into the objective function and using the whale optimization algorithm for parameter optimization, the ability of the least squares support vector regression model to characterize the nonlinear relationship between meteorological parameters and conversion power is improved, thereby improving the accuracy and stability of photovoltaic power station conversion power prediction.

4. The method according to claim 3, characterized in that The steps of using the whale optimization algorithm to search for the optimal parameter combination of the least squares support vector regression model and optimizing the regularization parameters, time series correlation constraint weight coefficients and kernel function scale parameters of the model by updating the position of the whale population include: The regularization parameters, time series correlation constraint weights and kernel function scale parameters in the least squares support vector regression model are encoded into the position coordinates of the whales in the whale optimization algorithm, and the initial positions of a group of whales are randomly generated to form an initial population. For each whale in the population, decode its position coordinates into least squares support vector regression model parameters, train the least squares support vector regression model and calculate the prediction error on the validation set as the fitness value of the whale; The whale with the highest fitness value in the current population is taken as the leader whale, and the position of each whale in the population is updated by adopting the siege predation mode or the spiral attack mode according to the probability parameter; in the siege predation mode, the whale calculates the coefficient vector according to the linear decreasing variable and the random vector, and updates its own position based on the position and coefficient vector of the leader whale; in the spiral attack mode, the whale updates its own position according to the distance between it and the leader whale and the spiral function to enhance the global exploration ability of the algorithm; For whales whose updated positions are out of the range, their positions are adjusted to the boundary of the parameters, and it is determined whether the termination conditions of the maximum number of iterations or the performance improvement threshold are met; if the termination conditions are not met, the fitness value calculation step is returned to continue the iterative optimization; if the termination conditions are met, the position coordinates of the leading whale are decoded into the optimal parameter combination of the least squares support vector regression machine.

5. The method according to claim 4, characterized in that The method further comprises: An adaptive hybrid search strategy is introduced to dynamically adjust the selection probability of the siege predation mode and the spiral attack mode. The adaptive weight coefficient w is designed to represent the probability of selecting the siege predation mode. The probability of selecting the spiral attack mode is 1-w. The value of w is adaptively adjusted according to the current number of iterations t and the population diversity index d. The calculation formula is as follows: w=1-t / T max ×(1-d / d max ); Among them, T max is the maximum number of iterations, d max is the maximum value of population diversity. The population diversity index d is measured by the average Euclidean distance between individuals in the population: d=(2 / N(N-1))×∑∑(i=1toN-1,j=i+1toN)||X i -X j ||; Where N represents the total number of samples in the data set, X represents the samples in the data set, i and j represent the index numbers of the samples, which are used to traverse and compare the samples pairwise. i -X j || represents the distance between the i-th sample and the j-th sample; As the iteration proceeds, the value of w gradually decreases, indicating that the probability of selecting the spiral attack mode increases, and the local search capability is enhanced; w is related to the population diversity index d. When the population diversity is low, the value of w is relatively large, indicating that the probability of selecting the siege predation mode increases, and the global exploration capability is enhanced to avoid premature convergence of the algorithm; In the siege-predation mode and spiral attack mode, an adaptive parameter control mechanism is introduced to dynamically adjust the algorithm parameters and improve the optimization performance; for the shrinkage factor a in the siege-predation mode, an adaptive update strategy is designed: a=a min +(a max -a min )×(1-t / T max ) k ; Among them, a min and a max are the minimum and maximum values ​​of a respectively, and k is a control parameter used to adjust the decay speed of a. As the iteration proceeds, the value of a changes from a to max Gradually decrease to a min , which enables the algorithm to have stronger global exploration capabilities in the early stage and stronger local development capabilities in the later stage; For the spiral shape parameter b in the spiral attack mode, an adaptive update strategy is designed: b=b max -(b max -b min )×t / T max ; Among them, b min and b max are the minimum and maximum values ​​of b respectively. As the iteration proceeds, the value of b changes from b max Gradually decrease to b min , so that the algorithm performs global search with a larger spiral step size in the early stage and local search with a smaller spiral step size in the later stage, improving the convergence speed and accuracy; A global exploration strategy based on Lévy flight is introduced. Lévy flight exploration is performed with a certain probability in each iteration. The random step size that obeys the Lévy distribution is generated by the Lévy flight function to update the current solution, so as to enhance the algorithm's ability to escape from the local optimum. During the iteration process, the siege and predation mode and the spiral attack mode are dynamically selected through an adaptive hybrid search strategy, and the algorithm parameters are adjusted using an adaptive parameter control mechanism. At the same time, the Lévy flight strategy is combined for global exploration to balance the development and utilization of the algorithm.

6. The method according to claim 1, characterized in that The conversion power data of the photovoltaic power station conversion power prediction model is adaptively decomposed using the variational mode decomposition algorithm to extract the multi-scale periodic characteristics and local fluctuation characteristics of the photovoltaic power station conversion power. The extracted characteristic variables and meteorological parameters are used as the input of the least squares support vector regression machine. The steps of constructing a photovoltaic power station conversion power prediction model driven by multi-source heterogeneous data include: The original conversion power sequence of the photovoltaic power station is preprocessed to remove outliers and missing values, and the power data is smoothed and normalized; according to the spectral characteristics of the preprocessed conversion power sequence, the number and penalty parameters of the intrinsic mode function components in the variational mode decomposition are preliminarily determined; The preprocessed power sequence is adaptively decomposed using the variational mode decomposition algorithm. The original power sequence is decomposed into several intrinsic mode function components by solving the constraint optimization problem of variational mode decomposition. Hilbert transform is performed on each intrinsic mode function component obtained by variational mode decomposition to extract the envelope and instantaneous frequency characteristics of each component, and the multi-scale periodic characteristics and local fluctuation characteristics of the conversion power sequence are reconstructed according to the frequency characteristics and energy distribution of each component. The reconstructed multi-scale periodic characteristics and local fluctuation characteristics are integrated with the meteorological parameters of the photovoltaic power station to form a multi-source heterogeneous feature set as the input of the photovoltaic power station conversion power prediction model.

7. The method according to claim 1, characterized in that After randomly shuffling the training set data, the established photovoltaic power station conversion power prediction model is trained and verified by cross-validation. Regularization and early stopping strategies are used to avoid overfitting of the photovoltaic power station conversion power prediction model. The steps of obtaining the optimal hyperparameters of the photovoltaic power station conversion power prediction model through grid search include: The training set data of the photovoltaic power station conversion power prediction model is randomly shuffled, and the photovoltaic power station conversion power prediction model is trained and verified by K-fold cross-validation. The training set is divided into K mutually exclusive subsets, and a different subset is selected as the verification set each time. The remaining subsets constitute the training set. The photovoltaic power station conversion power prediction model is trained with the training set, and the performance of the photovoltaic power station conversion power prediction model is evaluated on the verification set. The L2 norm penalty term of the parameter vector is introduced into the objective function of the photovoltaic power station conversion power prediction model. The size of the photovoltaic power station conversion power prediction model parameters is constrained by L2 regularization, the complexity of the photovoltaic power station conversion power prediction model is controlled, and the generalization ability of the photovoltaic power station conversion power prediction model is improved when the training data is insufficient. After each training round, the performance of the photovoltaic power plant conversion power prediction model on the validation set is evaluated. When the performance of the validation set has not improved for multiple consecutive times, the early stopping strategy is used to terminate the training process in advance to prevent the photovoltaic power plant conversion power prediction model from overfitting the training data. The hyperparameters of the photovoltaic power station conversion power prediction model are optimized through grid search, and the preset hyperparameter combinations are exhaustively searched. Cross-validation is used to evaluate the performance of the photovoltaic power station conversion power prediction model corresponding to each hyperparameter combination, and the hyperparameter combination with the best performance in the validation set is selected as the final parameter configuration of the photovoltaic power station conversion power prediction model.

8. A conversion power prediction system considering air humidity, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain the historical meteorological data of the target photovoltaic power station and the corresponding actual conversion power data, wherein the historical meteorological data includes the radiation intensity, scattered radiation intensity, ambient temperature and relative humidity of the air throughout the day, sort the historical meteorological data according to the time series, and identify and correct the missing values ​​and abnormal values; extract the design parameters of the photovoltaic power station, wherein the design parameters include the type of photovoltaic modules, the installation angle, the installation spacing, and the total installed capacity; The second unit is used to construct a nonlinear mapping model between relative air humidity and photovoltaic module conversion efficiency. Through the control variable experiment, the photovoltaic module conversion efficiency data under different relative air humidity levels are obtained, and the support vector regression machine is used to fit the nonlinear relationship curve between the photovoltaic module conversion efficiency and the relative air humidity. Based on the pre-processed historical meteorological data and conversion power data of the photovoltaic power station, combined with the design parameters of the photovoltaic power station, the least squares support vector regression machine is used to construct a nonlinear mapping relationship model between the historical meteorological data and the conversion power of the photovoltaic power station. The nonlinear mapping model between relative air humidity and photovoltaic module conversion efficiency is coupled with the nonlinear mapping relationship model between historical meteorological data and the conversion power of the photovoltaic power station to form a photovoltaic power station conversion power prediction model. The third unit is used to use the variational mode decomposition algorithm to adaptively decompose the conversion power data of the photovoltaic power station conversion power prediction model, extract the multi-scale periodic characteristics and local fluctuation characteristics of the photovoltaic power station conversion power, and use the extracted characteristic variables and meteorological parameters as the input of the least squares support vector regression machine to build a photovoltaic power station conversion power prediction model driven by multi-source heterogeneous data; after randomly shuffling the training set data, the established photovoltaic power station conversion power prediction model is trained and verified by cross-validation, and the regularization and early stopping strategies are used to avoid overfitting of the photovoltaic power station conversion power prediction model. The optimal hyperparameters of the photovoltaic power station conversion power prediction model are obtained through grid search, and the prediction sub-models obtained by training with multiple different random seeds are weighted combined using an integrated learning strategy to improve the robustness and generalization performance of the photovoltaic power station conversion power prediction model.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images