Grid-connected photovoltaic system power prediction method adopting hybrid meta learning method

Abstract

The invention discloses a grid-connected photovoltaic system power prediction method adopting a hybrid meta learning method. The method comprises the following steps: extracting meteorological data and historical power generation data; carrying out interval interpolation on the extracted data by using a nearest neighbor interpolation method, and preprocessing the data; calculating coplane irradiance considering dust; generating combined characteristics of cloud amount and coplane irradiance considering dust; generating statistical characteristics according to the power generation modes in different time periods; inputting the combined features and the statistical features into a plurality of models for training, and performing dimensionality reduction on the features by using principal component analysis; stacking a plurality of model prediction results and a deterministic model prediction result together, and combining with dimension-reduced data to form an extended feature vector; and learning the extended feature vector by using a long short-term memory network to generate a final power prediction result. According to the method, high-precision prediction of the photovoltaic power generation power in all time periods within 24 hours in the future can be realized, and the method has a relatively high practical application value.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power generation, and in particular relates to a power prediction method for a grid-connected photovoltaic system using a hybrid meta-learning method. Background Art

[0002] With the continuous increase in global energy demand and the improvement of environmental protection awareness, photovoltaic (PV) systems have been widely used as a renewable energy source. The power generation capacity of PV systems is directly affected by weather conditions, especially the changes in solar irradiance. Therefore, accurate prediction of the output power of PV systems is crucial for grid stability and energy management. However, the power output of PV systems is highly dependent on complex meteorological conditions, especially the changes in cloud cover and solar irradiance. Therefore, it is of great significance to develop a method that can accurately predict the power of PV systems.

[0003] Existing photovoltaic system power prediction methods mainly focus on grid-connected systems, and most of them rely on a single data source and simple models, which are difficult to cope with complex meteorological changes and different types of photovoltaic systems. In addition, traditional models often ignore the interaction between weather data and the impact of time series. Therefore, it is an urgent need to study a prediction method that can comprehensively consider multiple meteorological factors and is applicable to different photovoltaic systems.

[0004] Current research mainly uses the following methods to predict the power of photovoltaic systems:

[0005] 1. Methods based on statistics and physical models: These methods rely on historical data and basic physical laws, such as linear regression, time series analysis, and basic physical models. However, these methods often cannot cope with nonlinear and complex weather changes.

[0006] 2. Machine learning methods: With the improvement of computing power, machine learning models, such as random forest (RF) and extreme gradient boosting (XGB), have begun to be applied to power prediction of photovoltaic systems. Although these methods can capture nonlinear relationships in data, they still face problems of model complexity and data dependence.

[0007] 3. Deep learning methods: In recent years, deep learning methods such as deep neural networks (DNN) and recurrent neural networks (RNN) have been introduced into the field of photovoltaic power prediction. These methods can process complex time series data, but usually require a large amount of training data and have high computational complexity.

[0008] 4. Hybrid models: Some studies have attempted to combine physical models with machine learning models and proposed hybrid prediction methods. However, these hybrid models are usually complex in design and have limited applicability to different types of photovoltaic systems.

[0009] It can be seen that the defects of the existing photovoltaic power prediction methods make the existing photovoltaic power prediction have the problems of low accuracy or complex algorithms and large workload. Summary of the invention

[0010] Purpose of the invention: In response to the problems in the background technology, the present invention provides a power prediction method for a grid-connected photovoltaic system using a hybrid meta-learning method, which can be tested and verified on multiple data sets, has stronger generalization ability, and overcomes the limitations caused by a single data set. A power prediction method for a grid-connected photovoltaic system using a hybrid meta-learning method can achieve high-precision prediction of photovoltaic power generation for the entire period of the next 24 hours.

[0011] Technical solution: The present invention provides a method for predicting power of a grid-connected photovoltaic system using a hybrid meta-learning method, comprising the following steps:

[0012] Step 1: Extract meteorological data, including wind speed, cloud cover, temperature, humidity, air pressure, precipitation, and UV index; and historical power generation data, including power generation power, load power, state power, and battery system power data;

[0013] Step 2: Use the nearest neighbor interpolation method to interpolate the meteorological data and historical power generation data at intervals to make the time intervals of the meteorological data and historical power generation data consistent, and preprocess the data;

[0014] Step 3: Consider the impact of dust on the surface of photovoltaic modules on photovoltaic power generation, and calculate the coplanar irradiance considering dust;

[0015] Step 4: Generate the combined features of cloud cover and coplanar irradiance considering dust to capture the power generation law of the photovoltaic system under different conditions; generate statistical features according to the power generation mode in different time periods, including the minimum, maximum, average and standard deviation of the photovoltaic system power generation;

[0016] Step 5: Input the combined features and statistical features into the random forest, extreme gradient boosting, deep neural network, gated recurrent unit and long short-term memory network algorithm models for training, and use principal component analysis to reduce the feature dimension of meteorological data and historical power generation data;

[0017] Step 6: Combine the prediction results of the random forest, extreme gradient boosting, deep neural network, gated recurrent unit and long short-term memory network algorithm models with the prediction results of the deterministic model using the stacking method to form a new feature set, and combine it with the meteorological data and historical power generation data reduced by principal component analysis to form an extended feature vector;

[0018] Step 7: Use the LSTM network to learn the feature vector expanded in step 6. By meta-learning the stacked prediction results, the LSTM network performs weighted processing on the predictions of each model to generate the final power prediction result.

[0019] Furthermore, in step 2, preprocessing the data includes eliminating noise and processing missing values; using a noise filtering method to smooth the data curve and replace outliers.

[0020] Furthermore, in step 3, the coplanar irradiance considering dust is calculated

[0021]

[0022] in, represents the solar irradiance measured by a radiometer; SF is the contamination factor, which has a value of 1 for a clean panel or module and 0.95 for a panel covered with dust or other particles.

[0023] Furthermore, in step 4, a combined feature of solar irradiance and cloud cover taking into account dust is generated; the relationship between cloud cover and solar irradiance is nonlinear, and the two are combined to better learn and adapt to the nonlinear relationship and capture the impact of solar irradiance on the power output of the photovoltaic system under specific cloud cover conditions.

[0024] Furthermore, in step 5, the model is trained as follows:

[0025] The preprocessed data and combined features and statistical features are input into the model to train the random forest RF model. The extracted features are input for training, multiple decision trees are constructed, and the prediction results are averaged to generate power predictions for the photovoltaic system. The extreme gradient boosting XGB model is trained to improve the prediction accuracy by gradually optimizing the residual, and parallel computing is used to speed up the training process. The deep neural network DNN model is trained to use multiple hidden layers to perform complex nonlinear transformations on the input features, and the prediction error is minimized by the gradient descent method. The gated recurrent unit GRU model and the long short-term memory network LSTM model are trained to capture the long-term dependence of the photovoltaic system using the gated recurrent unit GRU and the long short-term memory network LSTM model for time series data.

[0026] Furthermore, in step 6, the deterministic model is:

[0027]

[0028] in, Describe the theoretical output power of the photovoltaic system, is the rated capacity of the module under STC; It is reduced under real operating conditions due to wire loss, pollution, aging, packaging, etc. Scaling or derating factor, 0.9; Sol STC is the solar irradiance under standard conditions (1000W / m 2 ); is the temperature coefficient power, which is 0.5% / ℃; Indicates battery temperature; is the battery temperature under STC; is the power regulation efficiency of DC-AC conversion, which is 0.9.

[0029] Beneficial effects:

[0030] 1. When processing the extracted meteorological data and historical power generation data, the present invention uses the nearest neighbor interpolation method to interpolate at intervals so that the time intervals of the meteorological data and the historical power generation data are consistent. The present invention generates a combination feature of cloud cover and coplanar irradiance taking into account dust to capture the power generation law of the photovoltaic system under different conditions. This feature can describe the actual working environment of the photovoltaic system, provide physically meaningful data input for the prediction model, and improve the power prediction of the system. The relationship between cloud cover and solar irradiance is nonlinear, and this nonlinear relationship is difficult to capture with a single feature. By combining the two, the model can better learn and adapt to these nonlinear relationships, thereby improving the prediction performance, so that the impact of solar irradiance on the power output of the photovoltaic system under specific cloud cover conditions can be captured, and the prediction error can be reduced.

[0031] 2. The present invention adopts a hybrid meta-learning method for power prediction, combining random forest, extreme gradient boosting, deep neural network, gated recurrent unit and long short-term memory network algorithm model for prediction training, which can capture the long-term dependence of the photovoltaic system and avoid the gradient vanishing problem in the traditional algorithm. In addition, the prediction accuracy can be improved by gradually optimizing the residual, and the training process can be accelerated by parallel computing. The prediction results of the random forest, extreme gradient boosting, deep neural network, gated recurrent unit and long short-term memory network algorithm model and the prediction results of the deterministic model are combined using the stacking method, and the stacked feature vectors are learned using the long short-term memory network to make high-precision predictions of photovoltaic power generation in all time periods within the next 24 hours. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a basic flow chart of the present invention for predicting the power of a grid-connected photovoltaic system;

[0033] Figure 2 is the solar irradiance map taking dust into account;

[0034] Figure 3 It is a cloud cover map;

[0035] Figure 4 is the PV power output diagram of the deterministic model;

[0036] Figure 5 It is a graph of photovoltaic power output using random forest, extreme gradient boosting, deep neural network, and gated recurrent unit;

[0037] Figure 6 It is a comparison chart between the power prediction of the hybrid meta-learning method and the measured power prediction. DETAILED DESCRIPTION

[0038] The following is a clear and complete description of the technical solutions in the embodiments of the present invention 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.

[0039] like Figure 1 As shown, the power prediction method of the grid-connected photovoltaic system based on the hybrid prediction model of the present invention makes the established hybrid prediction model closer to the actual situation, thereby achieving the purpose of accurately predicting the power of the actual grid-connected photovoltaic power generation system.

[0040] The present invention discloses a method for predicting power of a grid-connected photovoltaic system using a hybrid meta-learning method, comprising the following steps:

[0041] Step 1: Extract meteorological data (wind speed, cloud cover, temperature, humidity, air pressure, precipitation, and UV index) and historical power generation data (generated power, load power, state power, and battery system power data).

[0042] Step 2: The extracted meteorological data and historical power generation data are interpolated at intervals of 15 minutes using the nearest neighbor interpolation method to make the time intervals of the meteorological data consistent with those of the historical power generation data. The data are preprocessed, including eliminating noise and processing missing values. The noise filtering method is used to smooth the data curve and replace outliers.

[0043] Step 3: Figure 2 This is a solar irradiance diagram considering dust, where the irradiance comes from Solcast, AccuWeather and SolarAnywhere weather data API respectively. The influence of dust on the surface of photovoltaic modules on photovoltaic power generation is considered, and the coplanar irradiance considering dust is calculated using formula (1):

[0044]

[0045] in, represents the solar irradiance measured by a radiometer; SF is the contamination factor, which has a value of 1 for a clean panel or module and 0.95 for a panel covered with dust or other particles.

[0046] Step 4: Figure 3 The figure shows a cloud cover map, where the cloud cover comes from Solcast, AccuWeather and SolarAnywhere meteorological data API. Since the output power of the photovoltaic system is highly dependent on weather conditions, changes in cloud cover will significantly affect the level of solar irradiance. Therefore, a combined feature of cloud cover and co-planar irradiance taking into account dust is generated to capture the power generation law of the photovoltaic system under different conditions. This feature can describe the actual working environment of the photovoltaic system, provide physically meaningful data input for the prediction model, and improve the power prediction of the system. According to the power generation mode in different time periods, statistical features are generated, including the minimum, maximum, average and standard deviation of the power generation of the photovoltaic system.

[0047] The relationship between cloud cover and solar irradiance is nonlinear, and this nonlinear relationship is difficult to capture with a single feature. By combining the two, the model can better learn and adapt to these nonlinear relationships, thereby improving the prediction performance. This can capture the impact of solar irradiance on the power output of the photovoltaic system under specific cloud cover conditions and reduce prediction errors.

[0048] Step 5: Input the preprocessed data and new data-driven features into the model, train the random forest (RF) model, input the extracted features for training, build multiple decision trees, and average the prediction results to generate power predictions for the PV system; train the extreme gradient boosting (XGB) model to improve the prediction accuracy by gradually optimizing the residual, and use parallel computing to speed up the training process; train the deep neural network (DNN) model, use multiple hidden layers to perform complex nonlinear transformations on the input features, and minimize the prediction error through the gradient descent method; train the gated recurrent unit (GRU) model and the long short-term memory network (LSTM) model, and use the gated recurrent unit (GRU) and long short-term memory network (LSTM) models for training time series data. These models can capture the long-term dependencies of the PV system and avoid the gradient vanishing problem in traditional algorithms. Figure 4 This is the output power diagram calculated by the deterministic model based on Solcast, AccuWeather and SolarAnywhere data. Figure 5 It is a photovoltaic power diagram output by the random forest, extreme gradient boosting, deep neural network, and gated recurrent unit algorithm model.

[0049] Step 6: The prediction results of the random forest, extreme gradient boosting, deep neural network, gated recurrent unit and long short-term memory network algorithm models and the prediction results of the deterministic model are combined using the stacking method to form a new feature set, and combined with the meteorological data and historical power generation data reduced by principal component analysis to form an extended feature vector, where the deterministic model is expressed by formula (2):

[0050]

[0051] Among them PD m t Describe the theoretical output power of the photovoltaic system, is the rated capacity of the module under STC; It is reduced under real operating conditions due to wire loss, pollution, aging, packaging, etc. Scaling or derating factor, 0.9; Sol STC is the solar irradiance under standard conditions (1000W / m 2 ); is the temperature coefficient power, which is 0.5% / ℃; Indicates battery temperature; is the battery temperature under STC; is the power regulation efficiency of DC-AC conversion, which is 0.9.

[0052] Step 7: Use the LSTM network to learn the stacked feature vectors. By meta-learning the stacked prediction results, the LSTM network will weight the predictions of each model to generate the final power prediction results. Figure 6 The figure shows a comparison between the power prediction of the hybrid meta-learning method and the measured power prediction, which can make high-precision predictions of photovoltaic power generation in all periods within the next 24 hours.

[0053] In summary, the hybrid meta-learning method proposed in this paper has demonstrated stronger generalization ability through testing and verification on multiple data sets, overcoming the limitations caused by a single data set. It can effectively combine the advantages of different models and comprehensively consider deterministic and random factors, thereby significantly improving the accuracy and robustness of photovoltaic system power prediction.

[0054] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A power prediction method for a grid-connected photovoltaic system using a hybrid meta-learning method, characterized in that: The following steps are involved: Step 1: Extract meteorological data, including wind speed, cloud cover, temperature, humidity, air pressure, precipitation, and UV index; and historical power generation data, including power generation power, load power, state power, and battery system power data; Step 2: Use the nearest neighbor interpolation method to interpolate the meteorological data and historical power generation data at intervals to make the time intervals of the meteorological data and historical power generation data consistent, and preprocess the data; Step 3: Consider the impact of dust on the surface of photovoltaic modules on photovoltaic power generation, and calculate the coplanar irradiance considering dust; Step 4: Generate the combined features of cloud cover and coplanar irradiance considering dust to capture the power generation law of the photovoltaic system under different conditions; generate statistical features according to the power generation mode in different time periods, including the minimum, maximum, average and standard deviation of the photovoltaic system power generation; Step 5: Input the combined features and statistical features into the random forest, extreme gradient boosting, deep neural network, gated recurrent unit and long short-term memory network algorithm models for training, and use principal component analysis to reduce the feature dimension of meteorological data and historical power generation data; Step 6: Combine the prediction results of the random forest, extreme gradient boosting, deep neural network, gated recurrent unit and long short-term memory network algorithm models with the prediction results of the deterministic model using the stacking method to form a new feature set, and combine it with the meteorological data and historical power generation data reduced by principal component analysis to form an extended feature vector; Step 7: Use the LSTM network to learn the feature vector expanded in step 6. By meta-learning the stacked prediction results, the LSTM network performs weighted processing on the predictions of each model to generate the final power prediction result.

2. The method for predicting power of a grid-connected photovoltaic system using a hybrid meta-learning method according to claim 1, characterized in that: In step 2, preprocessing the data includes eliminating noise and processing missing values; using a noise filtering method to smooth the data curve and replace outliers.

3. The method for predicting power of a grid-connected photovoltaic system using a hybrid meta-learning method according to claim 1, characterized in that: In step 3, the coplanar irradiance Sol considering dust is calculated. i t : in, represents the solar irradiance measured by a radiometer; SF is the contamination factor, which has a value of 1 for a clean panel or module and 0.95 for a panel covered with dust or other particles.

4. The method for predicting power of a grid-connected photovoltaic system using a hybrid meta-learning method according to claim 1, characterized in that: In step 4, a combined feature of solar irradiance and cloud cover considering dust is generated; the cloud cover comes from Solcast, AccuWeather and SolarAnywhere meteorological data API. The relationship between cloud cover and solar irradiance is nonlinear. The two are combined to better learn and adapt to the nonlinear relationship and capture the impact of solar irradiance on photovoltaic system power output under specific cloud cover conditions.

5. The method for predicting power of a grid-connected photovoltaic system using a hybrid meta-learning method according to claim 1, characterized in that: In step 5, the training of the model is: The preprocessed data, combined features, and statistical features are input into the model, the random forest RF model is trained, the extracted features are input for training, multiple decision trees are constructed, and the prediction results are averaged to generate the power prediction of the PV system; Train the extreme gradient boosting XGB model to improve prediction accuracy by gradually optimizing the residuals and use parallel computing to speed up the training process; Train the deep neural network DNN model, use multiple hidden layers to perform complex nonlinear transformations on the input features, and minimize the prediction error through the gradient descent method; train the gated recurrent unit GRU model and the long short-term memory network LSTM model, and use the gated recurrent unit GRU and the long short-term memory network LSTM model to train time series data to capture the long-term dependence of the photovoltaic system.

6. The method for predicting power of a grid-connected photovoltaic system using a hybrid meta-learning method according to claim 1, characterized in that: In step 6, the deterministic model is: Among them, PD m t Describe the theoretical output power of the photovoltaic system, is the rated capacity of the module under STC; It is reduced under real operating conditions due to wire loss, pollution, aging, packaging, etc. Scaling or derating factor; Sol STC is the solar irradiance under standard conditions (1000W / m 2 ); is the temperature coefficient power, which is 0.5% / ℃; Indicates battery temperature; is the battery temperature under STC; is the power regulation efficiency of DC-AC conversion.

Images