Power transfer learning prediction method for new photovoltaic stations based on coupled irradiance and temperature

Abstract

The present invention discloses a power transfer learning prediction method for a newly built photovoltaic field station coupled with irradiance and temperature, which relates to the field of new energy prediction technology. The steps include using data from photovoltaic power stations that have been in operation for a long time to perform basic network training, and saving the overall structure of the model and the parameters of the DNN network, CNN network, and LSTM network layer for extracting the features of slope irradiance and component backplane temperature. Secondly, the parameters of the convolutional layer in the CNN network used for feature extraction, the parameters of the hidden layer in the DNN network, and the parameters of the hidden layer in the LSTM network are fixed for knowledge transfer, but training can still be performed for fine-tuning. Then, the remaining parameters are randomly initialized to perform adaptive training on new data. The data of the newly built photovoltaic power station is used to train and fine-tune the network to achieve transfer learning, thereby improving the adaptability of the network to the target domain data. The method of the present invention is verified using actual operation data of photovoltaic power stations, and the results show that the error in the prediction results is reduced by more than 25%.

Description

Technical Field

[0001] The present invention relates to the technical field of new energy prediction, and in particular to a power transfer learning prediction method for a newly built photovoltaic station by coupling irradiance and temperature. Background Art

[0002] With the continuous growth of global demand for clean energy, photovoltaic power generation, as an environmentally friendly and renewable energy form, has developed rapidly in recent years. my country has actively responded to the call of the international community and proposed to build a new power system with new energy as the main body to accelerate the development of the new energy industry. In recent years, with my country's strong support for the new energy industry, photovoltaic power generation systems have developed rapidly in my country. According to statistics from the National Energy Administration, by the end of 2023, my country's cumulative grid-connected photovoltaic power generation capacity has reached 610 million kilowatts, ranking first in the world. Affected by meteorological factors such as irradiance, wind speed, wind direction, ambient temperature and relative humidity, photovoltaic power generation systems are volatile and highly uncontrollable, which affects the safe and stable operation of photovoltaic power stations. In addition, the volatility of photovoltaic power generation systems may also affect price fluctuations in the electricity market and affect the economic efficiency of photovoltaic power stations. The volatility problem of photovoltaic power generation systems is not only related to the stable operation of the power system, but also to the sustainable development of the photovoltaic power generation industry and the realization of economic benefits. High-precision photovoltaic power generation power prediction technology is one of the important means to solve such problems. With the continuous growth of my country's photovoltaic power generation grid-connected capacity, the demand for photovoltaic power generation power prediction has also increased year by year.

[0003] Although photovoltaic power generation prediction has made certain progress in theory and technology, the meteorological environment in different regions and the operation of photovoltaic power stations will have a great impact on the prediction effect of photovoltaic power generation. In the face of newly built photovoltaic power stations, it is necessary to make corresponding adjustments to the photovoltaic power generation prediction model and parameters, further study and improve the prediction method, and improve its applicability and reliability. Summary of the invention

[0004] The present invention provides a power transfer learning prediction method for a newly built photovoltaic station by coupling irradiance and temperature, which solves the problem of output prediction in a small sample scenario of a newly built photovoltaic power station.

[0005] The power transfer learning prediction method for a newly built photovoltaic station by coupling irradiance and temperature comprises the following steps:

[0006] Step 1: Define the prediction of the power generation of a photovoltaic power station in the next 15 minutes to 4 hours as a prediction task, distinguish two different photovoltaic power stations in the same area, name the photovoltaic power station that has been in operation for a long time as photovoltaic power station A, and name the newly built photovoltaic power station as photovoltaic power station B. The prediction task of photovoltaic power station B is the final prediction task.

[0007] Step 2: Obtain historical meteorological data and historical output data of different photovoltaic power stations and perform preprocessing. Use cluster analysis to divide the historical meteorological data into three different weather types: sunny, cloudy, and rainy. At the same time, obtain the inclination angle and orientation angle parameters of the photovoltaic array to calculate the slope irradiance characteristics of the photovoltaic array.

[0008] Step 3: collect the temperature data of the center point of the back plate of the photovoltaic module through the temperature probe of the meteorological environment monitoring instrument of different photovoltaic power stations, and use it to represent the back plate temperature of the photovoltaic module, check whether there is an abnormal value of excessively high temperature in the collected data, and determine whether there is a hot spot fault at the measuring point to ensure the availability of data;

[0009] Step 4: Use the DNN network to establish a photovoltaic module backplane temperature feature extraction module, train the model based on the photovoltaic module backplane temperature data collected in step 3, and verify the effectiveness of the results;

[0010] Step 5: Use the CNN network to establish a photovoltaic component slow time-varying feature extraction module, and extract the day-ahead features through the historical meteorological data and historical power generation data of the day before the predicted day to characterize the slow time-varying features of the photovoltaic array at the current moment;

[0011] Step 6: Use LSTM network to establish a dynamic time series feature extraction module for photovoltaic power generation data, and extract intraday features through historical meteorological data and historical power generation data of the time period adjacent to the time period to be predicted, so as to characterize the time series correlation of photovoltaic power generation data;

[0012] Step 7: construct the input of the photovoltaic output prediction model based on multiple feature extraction modules and NWP data. On the basis of the weather type classification in step 2, the model is pre-trained with the data of photovoltaic power station A, and the overall model structure and network parameters for slope irradiance and component backplane temperature feature extraction are saved;

[0013] Step 8: Based on step 7, the prediction model is adaptively adjusted using the data of photovoltaic power station B to complete the prediction task under the scenario of photovoltaic power station B;

[0014] Step 7, determine the weather type for a certain period in the future, input historical meteorological data, historical power generation data and NWP data into the corresponding prediction model to obtain the photovoltaic output forecast value for the next 15 minutes to 4 hours;

[0015] As a preferred method, model migration depends on the feature similarity between the two photovoltaic power plants. Using two photovoltaic power plants in the same area for model migration will obtain better prediction results.

[0016] Preferably, the slope irradiance feature extraction module is sensitive to the inclination and orientation angles of photovoltaic modules. For scenes with complex changes in mountain photovoltaic angles, a detailed analysis of the placement of photovoltaic modules in each area is required to ensure the effectiveness of feature extraction.

[0017] Preferably, the photovoltaic module inclined surface irradiance mechanism model expression is as follows:

[0018]

[0019] Among them, E Tirr represents the prediction data of effective irradiance of photovoltaic modules based on empirical knowledge, E b ×R b Indicates the direct solar irradiance on the slope, E b is the direct solar irradiance on the horizontal plane, represents the amount of direct solar radiation on the inclined surface, β is the tilt angle of the photovoltaic module, and E d is the solar diffuse irradiance on the horizontal plane, represents the ground reflected irradiance, ρ is the ground reflectivity,

[0020] θ T is the incident angle of the sun's rays on the inclined surface, θ z is the zenith angle;

[0021] Preferably, the model migration method freezes the overall network structure of the pre-trained model and the network parameters used for extracting slope irradiance and component backplane temperature features, and uses data from newly built photovoltaic power plants to adjust the parameters of the neural network used to predict photovoltaic output, so that the neural network has a certain degree of freedom to learn the features of small sample data sets, which helps the neural network to extract new data features based on the learned features while still maintaining common properties, thereby reducing the impact of data distribution differences, so as to achieve better generalization in predictions in new photovoltaic power plant scenarios.

[0022] Preferably, the photovoltaic output prediction model based on multi-feature extraction includes an input layer (composed of the four feature extraction modules), three LSTM layers, a Relu activation layer, a fully connected layer and an output layer (for outputting photovoltaic power generation prediction data) connected in sequence.

[0023] As a preference, the root mean square error (RMSE), accuracy, mean absolute error (MAE), and coefficient of determination (R 2 ) is used as the evaluation index of the model, and its expression is:

[0024]

[0025] Where: n is the number of data; is the predicted value of the i-th data; y i is the i-th data;

[0026] is the data average.

[0027] As a preferred method, RMSE measures the deviation between the predicted value and the true value, is more sensitive to outliers in the data, and can evaluate the overall prediction accuracy in photovoltaic power generation prediction. The unit is kW; MAE represents the average of the absolute errors between the predicted value and the observed value. MAE is a linear score. All individual differences have equal weights on the average value, and there is no problem of enlarging the error due to square addition. Therefore, as a loss function, its robustness to outliers is higher than RMSE. The unit is kW; accuracy measures the deviation between the prediction result in the 4th hour and the true value, and has no unit; R 2 Indicates the correlation between the predicted value and the actual value, unitless.

[0028] The present invention also provides a related device for a power transfer learning prediction method for a new photovoltaic site that couples irradiance and temperature. The device includes a processor and a memory. The memory is used to store program code and transmit the program code to the processor. The processor is used to execute a photovoltaic output prediction method driven by knowledge and data fusion according to instructions in the program code. A computer-readable storage medium is also provided. The computer-readable storage medium is used to store program code. The program code is used to execute the power prediction method for a new photovoltaic power station.

[0029] Compared with the prior art, the beneficial effects and advantages of the present invention are:

[0030] (1) The present invention establishes multiple feature extraction modules, which fully consider the various factors affecting photovoltaic power generation, including the influence of the installation angle of photovoltaic modules, the influence of the temperature change of photovoltaic modules, the influence of the slow time-varying characteristics of the photovoltaic array, and the influence of meteorological conditions.

[0031] (2) Aiming at the problem of few data samples of newly built photovoltaic power stations, the present invention proposes a method for predicting the output of photovoltaic power stations with a small sample size based on model migration, and divides historical data into three categories of sunny, cloudy and rainy days through clustering, thus realizing the output prediction of newly built photovoltaic power stations in different scenarios. Through practical verification, the prediction error (RMSE) of the method for predicting the output of photovoltaic power stations with a small sample size based on model migration is reduced by 25.14% to 53.79% compared with the traditional neural network method, and the prediction accuracy is greatly improved.

[0032] (3) The present invention solves the problem of photovoltaic output prediction of newly built photovoltaic power stations through a hybrid model of multi-feature extraction and a small sample scenario prediction method of model migration, effectively reduces abandoned light, helps to further increase the number of photovoltaic power stations, and is conducive to promoting the further development of the solar energy industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flowchart of a method for predicting power transfer learning of a newly built photovoltaic station by coupling irradiance and temperature of the present invention;

[0034] Figure 2 It is a flowchart of the photovoltaic output prediction model based on multi-feature extraction of the present invention;

[0035] Figure 3 It is a schematic diagram of the process flow of the inclined plane irradiance mechanism model of the present invention;

[0036] Figure 4 is a flow chart of the photovoltaic module back plate temperature feature extraction of the present invention;

[0037] Figure 5 is a flow chart of a slow time-varying feature extraction module of the present invention;

[0038] Figure 6 is a flow chart of a dynamic time series feature extraction module of the present invention;

[0039] Figure 7 It is the data and comparison chart that can be collected from two photovoltaic power stations in Zhejiang Province in the implementation case 1 of the present invention;

[0040] Figure 8 is a comparison chart of prediction results of a newly built photovoltaic power station under different weather scenarios in implementation case 1 of the present invention;

[0041] Fig. 9 It is a radar chart of the evaluation index of the prediction results of the newly built photovoltaic power station under different weather scenarios in implementation case 1 of the present invention;

[0042] Fig.10 It is the data and comparison chart that can be collected from two photovoltaic power stations in Hebei Province in the implementation case 2 of the present invention;

[0043] Fig.11 is a comparison chart of prediction results of a newly built photovoltaic power station under different weather scenarios in implementation case 2 of the present invention;

[0044] Fig.12 It is a radar chart of the prediction result evaluation indicators of the newly built photovoltaic power station under different weather scenarios in implementation case 2 of the present invention. DETAILED DESCRIPTION

[0045] The present invention will be described in detail below through specific embodiments in combination with the accompanying drawings so that people can understand the content and specific features of the present invention more intuitively and vividly, but the protection scope of the present invention is not limited to the following embodiments.

[0046] As attached Figure 1 As shown, the present invention relates to a power transfer learning prediction method for a newly built photovoltaic field station coupled with irradiance and temperature, including data acquisition and data preprocessing, slope irradiance mechanism model analysis based on meteorological data and geographic information, establishment of feature extraction module, cluster analysis of original power generation data, construction of photovoltaic output prediction model based on multi-feature extraction, model pre-training based on photovoltaic power station A, adaptive adjustment of network parameters based on photovoltaic power station B, and photovoltaic output prediction in the scenario of a newly built photovoltaic power station. The present invention first obtains the photovoltaic power generation historical data required for modeling and the meteorological data of different photovoltaic power stations, performs data preprocessing, and uses the k-means algorithm to divide the weather types into three categories of sunny, cloudy and rainy days based on the photovoltaic power generation power data and the meteorological data of the photovoltaic power station, constructs the input of the prediction model in combination with the established multiple feature extraction modules, uses the data pre-training model of photovoltaic power station A, and adaptively adjusts the model parameters based on the data of photovoltaic power station B, and finally obtains the power generation prediction data of photovoltaic power station B in the future period.

[0047] Example 1

[0048] Two photovoltaic power stations in Zhejiang Province are used to verify the effectiveness of the small sample photovoltaic power station power prediction method based on model migration. The two photovoltaic power stations in Zhejiang Province are a photovoltaic power station that has been in operation for a long time and a newly built photovoltaic power station. The historical data collected from the two photovoltaic power stations are as follows: Figure 7 As shown in the figure, a photovoltaic power station that has been in operation for a long time can collect data for a four-year period, and a newly built photovoltaic power station can collect historical data for two months.

[0049] The collected data include photovoltaic power station power generation (PV output), horizontal irradiance (GlobalHorizontal Irradiation, GHI), diffuse irradiance (Diffuse Horizontal Irradiation, DHI), direct normal irradiation (Direct Normal Irradiation, DNI), temperature (Air temperature at 2m, TEMP), wind speed at 10 meters (Wind speed at 10m, WS), wind direction at 10 meters (Wind direction at 10m, WD), humidity (Relative humidity, RH), potential precipitation (Precipitable water, PWAT), atmospheric pressure (Atmospheric pressure, AP), rainfall rate (Precipitation rate, PREC), and the temporal resolution of the data is 15 minutes.

[0050] The historical data of photovoltaic power generation from different photovoltaic power stations were collected, and the environmental monitoring instrument of the photovoltaic power station was used to obtain the environmental factors that affect the photovoltaic power generation, such as horizontal irradiance, direct irradiance, diffuse irradiance, ambient temperature, wind speed, wind direction, humidity, and atmospheric pressure. The data was preprocessed by removing outliers and filling in missing values, and the k-means clustering method was used to cluster the photovoltaic power station data into three types: sunny, cloudy, and rainy. The data preprocessing includes outlier removal based on the box plot method, interpolation processing based on the moving mean, and screening of effective power generation data.

[0051] Considering that photovoltaic power generation is affected by many factors, the correlation analysis of meteorological factors, slope irradiance, photovoltaic module backplane temperature and photovoltaic power generation that affect photovoltaic power generation is carried out, and the main factors affecting photovoltaic power generation are screened out. The various factors affecting photovoltaic power generation are analyzed, including horizontal irradiance, direct irradiance, scattered irradiance, solar zenith angle, temperature, ozone concentration, surface reflectivity, atmospheric pressure, wind direction, wind speed, aerosol optical depth, precipitation, PM2.5, PM10, photovoltaic module installation inclination angle, photovoltaic module orientation angle, photovoltaic module status, inverter status, and combined with the input variable form required by the prediction model, some variables that have a greater impact on photovoltaic power generation are selected for correlation analysis, and the correlation coefficient with photovoltaic power generation is obtained, which provides a basis for the establishment of subsequent models.

[0052] The Pearson correlation coefficient is used to weigh the correlation between output power and meteorological factors. The Pearson correlation coefficient can quantify the correlation between different factors. The closer the correlation coefficient is to ±1, the greater the correlation with photovoltaic power generation. The sampling sequences of the two random variables are: x = {x1, ..., x n}, y={y1,…,y n}, then the Pearson correlation coefficient between the two random variables is r p It can be expressed as:

[0053]

[0054] Where: n represents the sampling sequence size; and denote the sample means of the random variables x and y respectively.

[0055] The correlation coefficients between photovoltaic power generation and different meteorological characteristics are shown in Table 1.

[0056] Table 1 Correlation coefficients between different variables affecting photovoltaic power generation

[0057]

[0058] Based on the data in Table 1, from the absolute value analysis of the correlation coefficient, the photovoltaic power generation is highly correlated with the horizontal irradiance, direct irradiance and scattered irradiance, indicating that irradiance is one of the key factors affecting photovoltaic power generation. At the same time, the correlation coefficient between the slope irradiance and the photovoltaic power generation obtained by the slope irradiance mechanism model reaches 0.96, which is higher than the correlation coefficients of other factors, proving the effectiveness of the slope irradiance mechanism model. In addition, through correlation analysis, it can be found that the correlation coefficient between the photovoltaic module backplane temperature and the photovoltaic power generation can reach 0.86, which is 65.38% higher than the ambient temperature, proving the effectiveness of the photovoltaic module backplane temperature feature extraction module.

[0059] The k-means clustering algorithm is used to divide the historical data into three categories in combination with meteorological data, which respectively represent the three weather types of sunny, cloudy and rainy days. The data pre-training model of photovoltaic power station A, which has been in operation for a long time, is used to obtain the photovoltaic output prediction model suitable for photovoltaic power station A under different weather scenarios. The overall model structure and the neural network model parameters for slope irradiance and component backplane temperature feature extraction are fixed. The pre-trained model is adaptively adjusted through the data of the newly built photovoltaic power station B to make it suitable for the output prediction task of photovoltaic power station B. The weather type of photovoltaic power station B in the prediction period is judged and substituted into the corresponding prediction model to obtain the photovoltaic output prediction value of photovoltaic power station B.

[0060] The ultra-short-term photovoltaic output prediction model process based on multi-feature extraction is as follows: Figure 2 As shown, the input of the prediction model is constructed based on multiple feature extraction modules, and the multi-layer neural network training model is used to obtain the photovoltaic output for the next 15min-4h.

[0061] The slope irradiance calculation principle is as follows Figure 3 As shown, the slope irradiance in the normal direction of the photovoltaic module consists of three parts: direct irradiance, scattered irradiance, and ground reflected irradiance. In the figure, a) is a schematic diagram of the slope irradiance conversion, b) is a schematic diagram of the relative position change of the sun and the ground in a year, and c) is a schematic diagram of the relative position change of the sun and the photovoltaic module in a year.

[0062] The module for extracting the temperature characteristics of the backplane of the component is composed of Figure 4 The process of building a DNN model is analyzed to analyze the change mechanism of the back panel temperature of the photovoltaic module. During the operation of the photovoltaic module, part of the energy is converted into electrical energy output, while the other part of the energy is converted into heat energy and heat exchanged with the external environment. Considering the radiation heat exchange and convection heat exchange process between the photovoltaic module and the environment, combined with the correlation coefficient analysis, the key features that affect the back panel temperature of the photovoltaic module are obtained: slope irradiance, ambient temperature, wind speed, wind direction, and ambient humidity. The extracted key features are used to construct the input of the DNN model, and the output of the back panel temperature of the photovoltaic module is obtained through the hidden layer.

[0063] The slow time-varying feature extraction module is composed of Figure 5 The process of constructing the prediction model is based on the slow time-varying effect of the efficiency decay of photovoltaic modules. The photovoltaic power generation system will produce performance decay during operation, resulting in a decrease in the photoelectric conversion rate, which in turn affects the output of photovoltaic power generation. In the process of photovoltaic power generation prediction, the slow time-varying characteristics of the photovoltaic array are used as one of the input variables in the hope of improving the prediction accuracy. Since the performance decay of the photovoltaic array is a long-term process, the power generation state of the day before the prediction can be used to characterize the slow time-varying characteristics of the photovoltaic array at the current moment. The meteorological data and power generation history data of the day before the prediction are used to construct the input of the feature extraction module of the day before the prediction. The meteorological data include horizontal irradiance, direct irradiance, scattered irradiance, humidity, atmospheric pressure, component temperature, and slope irradiance. Combined with the effectiveness of CNN neural network in feature extraction, the slow time-varying characteristics of the photovoltaic array are extracted through CNN neural network.

[0064] The dynamic time series feature extraction module is composed of Figure 6The process is constructed, and the intraday feature extraction module is constructed based on the LSTM neural network. The slope irradiance feature extraction module and the component backplane temperature feature extraction module are coupled to extract the slope irradiance characteristics and component backplane temperature characteristics 4 hours before the prediction period. The historical power generation data and meteorological data are used to construct the input of the module. The meteorological data include horizontal irradiance, direct irradiance, diffuse irradiance, humidity, atmospheric pressure, component temperature, and slope irradiance. The multi-layer LSTM network outputs the feature values ​​of the vicinity of 4 hours to characterize the temporal correlation of photovoltaic power generation data.

[0065] Typical sunny, cloudy and rainy days were selected respectively, and the model with slow time-varying features (day-ahead features) was named "1-feature model"; the model with slow time-varying features (day-ahead features) and dynamic time series features (intraday features) was named "2-feature model"; the model with additional slope irradiance features and component backplane temperature features was named "4-feature model", and compared with a single LSTM model, the effectiveness of the small-sample photovoltaic power station output prediction method based on model migration was verified under different weather types.

[0066] The comparison of multi-model prediction results in the small sample scenario of photovoltaic power station B is shown in the figure below. Figure 8 As shown in the figure, it can be seen that the small sample photovoltaic power station output prediction method based on model migration can obtain better prediction results under different weather scenarios. The comparison of evaluation indicators of each model is shown in Fig. 9 As shown in the figure, it can be seen that the results of the photovoltaic output prediction model based on multi-feature extraction are better in fit with the true value.

[0067] By comparing different prediction models, in the small sample data set scenario in Zhejiang Province, the photovoltaic power generation prediction model based on the multi-feature extraction module showed higher prediction accuracy, and the prediction error (RMSE) was reduced by 4.85% to 21.08%, which verified that the small sample photovoltaic power station output prediction method based on model migration has high adaptability in the scenario of newly built photovoltaic power stations in Zhejiang Province.

[0068] The prediction results are shown in Table 2:

[0069] Table 2 Prediction results of different prediction models for different weather types

[0070]

[0071] Example 2

[0072] Two photovoltaic power stations in Hebei Province are used to verify the effectiveness of the small sample photovoltaic power station power prediction method based on model migration. The two photovoltaic power stations in Hebei Province are a photovoltaic power station that has been in operation for a long time and a newly built photovoltaic power station. The historical data collected from the two photovoltaic power stations are as follows: Fig.10 As shown in the figure, a photovoltaic power station that has been in operation for a long time can collect data for a period of one year, and a newly built photovoltaic power station can collect historical data for one month.

[0073] The collected data include photovoltaic power station power generation (PV output), horizontal irradiance (GlobalHorizontal Irradiation, GHI), diffuse irradiance (Diffuse Horizontal Irradiation, DHI), direct irradiance (Direct Normal Irradiation, DNI), temperature (Air temperature at 2m, TEMP), wind speed (Wind speed at 10m, WS), wind direction (Wind direction at 10m, WD), atmospheric pressure (Atmospheric pressure, AP), and the time resolution of the data is 15 minutes.

[0074] Taking into account the impact of data size on model training, the historical data of the two photovoltaic power stations are divided into two categories through cluster analysis, corresponding to the actual weather categories of sunny days and rainy days, respectively, so as to extract weather type characteristics while ensuring the data size.

[0075] Based on the data collected from a photovoltaic power station that has been in operation for a long time, a photovoltaic output prediction model based on multi-feature extraction is used to pre-train the prediction model. The overall model structure and network parameters for slope irradiance and component backplane temperature feature extraction are saved. The model is adaptively adjusted through the data of a newly built photovoltaic power station.

[0076] The prediction results of a newly built photovoltaic power station in Hebei Province are as follows: Fig.11 As shown in the figure, it can be seen that the small sample photovoltaic power station power prediction method based on model migration still has a good prediction effect in Hebei Province. The specific prediction results are shown in the table: Table 3 Prediction results of different prediction models for different weather types in Hebei Province

[0077]

[0078] Preferably, by comparing the prediction results of different models, it can be seen that in the small sample data set scenario in Hebei Province, the photovoltaic power generation prediction model based on the multi-feature extraction module shows a higher prediction accuracy, which is better than other models, and the prediction error (RMSE) is reduced by 6.72% to 31.76%, which verifies that the small sample photovoltaic power station output prediction method based on model migration has high adaptability in the scenario of newly built photovoltaic power stations in Hebei Province. The comparison chart of the evaluation indicators of each model is as follows Fig.12 shown.

[0079] Preferably, the power transfer learning prediction method for newly built photovoltaic power stations with coupled irradiance and temperature is based on a multi-feature extraction module, which fully considers the influence of environmental factors and the slow time-varying characteristics of photovoltaic components themselves. In the face of the prediction task of newly built photovoltaic power stations, the model migration method is used to complete the photovoltaic output prediction task, which improves the prediction accuracy and has strong adaptability. The small sample photovoltaic power station output prediction method based on model migration can provide an effective solution for the prediction task of newly built photovoltaic power stations, performs well in different weather type scenarios, and has strong adaptability.

[0080] Other technologies mentioned in the present invention, such as data preprocessing, K-means clustering, etc., are all existing technologies and are not described in detail.

[0081] The present invention is described in detail above in conjunction with the embodiments, but the contents described are only preferred embodiments of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made according to the scope of application of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A power transfer learning prediction method for a new photovoltaic station coupled with irradiance and temperature, characterized in that: The following steps are involved: Step 1: Define the prediction of the power generation of a photovoltaic power station in the next 15 minutes to 4 hours as a prediction task. Different photovoltaic power stations correspond to different prediction tasks. Step 2: Name two PV power stations in the same area as PV power station A and PV power station B. PV power station A corresponds to a PV power station that has been in operation for a long time and has sufficient data, while PV power station B corresponds to a newly built PV power station and has less data. Step 3: Obtain historical meteorological data and historical photovoltaic power generation data of different photovoltaic power stations and perform data preprocessing. Use k-means algorithm clustering to divide the historical meteorological data of different photovoltaic power stations into three weather types: sunny, cloudy, and rainy. Step 4: Obtain the inclination angle, azimuth angle information, short-circuit current, open-circuit voltage, and rated temperature parameter information of different photovoltaic modules through on-site data collection of different photovoltaic power stations; Step 5: Analyze the principle of photovoltaic modules absorbing solar irradiance, establish a slope irradiance mechanism model in the normal direction of the photovoltaic modules, and convert the direct irradiance data and scattered irradiance data collected in step 3 into slope irradiance data in the normal direction of the photovoltaic modules through the mechanism model based on the inclination and azimuth information collected in step 4; Step 6: Considering the difference between ambient temperature and module backplane temperature, the temperature characteristics of the photovoltaic module backplane are extracted by using deep neural networks (DNN) combined with irradiance, wind speed, wind direction, ambient temperature and relative humidity parameter information; Step 7: The health status, photoelectric conversion efficiency, and dust accumulation of photovoltaic modules are slow time-varying characteristics of the operation status of photovoltaic modules. In order to analyze the impact of the slow time-varying characteristics of the photovoltaic array on the photovoltaic power generation, the day-ahead characteristics of the photovoltaic power station are extracted through the Convolutional Neural Network (CNN) to characterize the current slow time-varying characteristic status of the photovoltaic array. Step 8: Photovoltaic power generation data is a typical time series data with a certain degree of autocorrelation, that is, there is a correlation between the power generation at the current moment and the power generation at the previous moment. In order to analyze the impact of the dynamic time series characteristics of photovoltaic power generation data on photovoltaic power generation, a long short-term memory neural network (LSTM) is introduced to extract the intraday characteristics of the photovoltaic array to characterize the time series correlation between the historical photovoltaic power generation data. Step 9: Obtain meteorological data for the next 15 minutes to 4 hours through Numerical Weather Prediction (NWP). The meteorological data include horizontal irradiance, direct irradiance, diffuse irradiance, ambient temperature, humidity, wind speed, wind direction, relative humidity and atmospheric pressure. Step 10: For photovoltaic power station A, a photovoltaic output prediction model based on multi-feature extraction is proposed. The input part of the neural network is constructed by combining the slope irradiance feature extraction module established in step 5, the photovoltaic module backplane temperature feature extraction module established in step 6, the day-ahead feature extraction module established in step 7, the intraday feature extraction module established in step 8, and the NWP data collected in step 9. Through multi-layer LSTM neural network training, the photovoltaic output forecast value for the next 15min-4h is obtained; Step 11: For photovoltaic power station A, according to the different weather types divided in step 3, pre-train photovoltaic output prediction models based on multi-feature extraction respectively, obtain pre-trained models for photovoltaic output prediction corresponding to different weather types, save the overall model structure of the three trained models and the parameters of the neural network module for slope irradiance and component backplane temperature feature extraction; Step 12: For photovoltaic power station B, the data of photovoltaic power station B is divided into three categories: sunny, cloudy and rainy based on the weather type characteristics, and the power generation power of photovoltaic power station B is predicted based on the pre-trained models of the corresponding weather types. In order to make the model match the characteristics of photovoltaic power station B, on the basis of freezing the structure of the pre-trained model and the parameters of the neural network module for extracting the slope irradiance and component backplane temperature features, the parameters of the LSTM neural network used for power prediction output are adjusted through the historical power generation data and meteorological data of photovoltaic power station B, and the photovoltaic power generation power prediction model adapted to the three different weather types of photovoltaic power station B is obtained.

2. The method for predicting power transfer learning of a newly built photovoltaic station by coupling irradiance and temperature according to claim 1 is characterized by: Historical meteorological data include direct irradiance DNI, diffuse irradiance DHI, horizontal irradiance GHI, ambient temperature, relative humidity, solar zenith angle, atmospheric pressure, wind speed, wind direction, and PM2.5 concentration.

3. The method for predicting power transfer learning of a newly built photovoltaic station by coupling irradiance and temperature according to claim 1 is characterized in that: Data preprocessing includes outlier processing based on box plots, interpolation processing based on moving averages, and screening of effective power generation data.

4. The method for predicting power transfer learning of a newly built photovoltaic station by coupling irradiance and temperature according to claim 1 is characterized by: The geographical location information of the photovoltaic module includes the tilt angle, azimuth, and ground reflectivity of the photovoltaic module, and the performance parameters include the module efficiency and rated electrical parameters.

5. The method for predicting power transfer learning of a newly built photovoltaic station by coupling irradiance and temperature according to claim 1 is characterized in that: The feature extraction module consists of four parts, including a slope irradiance feature extraction module based on physical modeling, a component backplane temperature feature extraction module based on a DNN network, a slow time-varying feature extraction module based on a CNN network, and a dynamic timing feature extraction module based on an LSTM network.

6. The method for predicting power transfer learning of a newly built photovoltaic station by coupling irradiance and temperature according to claim 1 is characterized by: The photovoltaic output prediction model based on multi-feature extraction includes an input layer, three LSTM layers, a Relu activation layer, a fully connected layer and an output layer connected in sequence.

7. The method for predicting power transfer learning of a newly built photovoltaic station by coupling irradiance and temperature according to claim 1 is characterized by: Whether the meteorological characteristics of two photovoltaic power stations are similar determines whether the conditions for model migration can be met. The photovoltaic power station A and photovoltaic power station B are both located in Zhejiang Province, both have subtropical monsoon climates, and both are distributed bracket photovoltaics, which meet the conditions for model migration.

8. The method for predicting power transfer learning of a newly built photovoltaic station by coupling irradiance and temperature according to claim 1 is characterized by: The parts of the pre-trained photovoltaic output prediction model to be frozen are the overall model structure and the parameters of the neural network used to extract slope irradiance and component backplane temperature features. The neural network part used to predict the output will be adaptively adjusted based on the data collected from the newly built photovoltaic power station.

Images