Distributed photovoltaic ultra-short-term output prediction method based on multiple models

Abstract

The invention discloses a multi-model-based distributed photovoltaic ultra-short-term output prediction method, which can improve the precision of photovoltaic output ultra-short-term prediction through a multi-model distributed photovoltaic output method, and comprises the following steps of: firstly, processing acquired data to obtain a high-quality data set; then, a photovoltaic output ultra-short-term prediction model is trained, the model evaluates the importance of each feature by using an attention mechanism, extracts a hierarchical structure of more abstract features by using a convolutional layer, and performs feature fusion on the results of the two features, so that the model can effectively improve the photovoltaic output ultra-short-term prediction result and train a weather prediction model in the future 4 hours; the method comprises the following steps: generating an accurate photovoltaic output by using predicted future meteorological data, and finally, correcting a photovoltaic output ultra-short-term prediction result according to a meteorological prediction result in the future 4 hours and a photovoltaic output short-term prediction result to obtain a more accurate value.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation, and in particular to a distributed photovoltaic ultra-short-term output prediction method based on multiple models. Background Art

[0002] Photovoltaic power generation, as a clean and renewable energy form, has attracted widespread attention due to its low-carbon and environmentally friendly characteristics. Since photovoltaic power generation output is affected by many factors such as solar irradiance, temperature, relative humidity, cloud thickness, wind speed, geographical location, etc., it has strong volatility and uncontrollability, which will greatly affect the stable operation of the power system. Therefore, accurate photovoltaic output prediction is of great significance to improving the stability of power system operation and the rationality of grid management.

[0003] The ultra-short-term photovoltaic output prediction technology is not yet mature. In the actual photovoltaic output prediction, it is necessary to combine the future predicted meteorological data to predict the photovoltaic output. However, the minimum time scale of the meteorological forecast data is 1 hour, and the ultra-short-term photovoltaic output prediction is mostly based on a 15-minute prediction cycle. There is an inconsistency in the time scale between the two, which increases the difficulty of information fusion. Secondly, even if the model works well in a single prediction, predicting the photovoltaic output in the 4th hour requires the model to go through 16 iterations, which will cause serious error propagation, making the result error of the 4th hour very large. In response to the above problems, the present invention proposes a distributed photovoltaic ultra-short-term output prediction method based on multiple models. Summary of the invention

[0004] The purpose of the present invention is to provide a distributed photovoltaic ultra-short-term output prediction method based on multiple models, which can predict the output data of the next 4 hours through historical output data, and correct the ultra-short-term prediction results of photovoltaic output in combination with the short-term photovoltaic output prediction results, thereby reducing the advantage of errors, improving the accuracy of ultra-short-term prediction of photovoltaic output, and promoting its implementation in the power system and energy field, so as to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a distributed photovoltaic ultra-short-term output prediction method based on multiple models, comprising the following steps:

[0006] Step 1: Obtain a data set, collect the original data set, and preprocess the original data set to obtain a sample data set;

[0007] Step 2: Model training: construct a training model and train the PV output ultra-short-term prediction model, the next 4-hour weather prediction model, and the PV output short-term prediction model based on the data set;

[0008] Step 3: Model prediction results: According to the weather forecast results for the next 4 hours and the short-term forecast results for photovoltaic output, the preliminary ultra-short-term forecast results for photovoltaic output are corrected to obtain the final ultra-short-term forecast results for photovoltaic output and output them.

[0009] Exemplarily, the acquiring of the data set includes:

[0010] Collect data, obtain historical photovoltaic output data sets from photovoltaic power plants and historical meteorological data sets from the Meteorological Bureau;

[0011] Data processing, removing duplicate data, deleting useless information, repairing all missing data, and eliminating abnormal data.

[0012] Exemplarily, the ultra-short-term photovoltaic output prediction model includes generating training samples, constructing and training the ultra-short-term photovoltaic output prediction model, and selecting weights for the ultra-short-term photovoltaic output prediction model.

[0013] Exemplarily, when generating training samples, the photovoltaic output data set and the meteorological data set are integrated, and the photovoltaic output, meteorological data, and time data information are integrated to obtain training samples. According to the prediction model, the training samples are converted into the data shape required by the prediction model, and the sample data is divided into training set, validation set, and test set according to a certain ratio.

[0014] Exemplarily, when constructing and training an ultra-short-term photovoltaic output prediction model, an ultra-short-term photovoltaic output prediction model is constructed based on machine learning and deep learning networks, including XGBoost, LSTM, Transformer and their variants. Single-step ultra-short-term photovoltaic output prediction training is performed first, and then the photovoltaic output is predicted for the next 4 hours.

[0015] Exemplarily, when selecting the weights of the ultra-short-term photovoltaic output prediction model, the ultra-short-term photovoltaic output prediction model is evaluated using a validation set and tested using a test set, and the model weight with the best performance is selected as the model weight for subsequent ultra-short-term photovoltaic output predictions.

[0016] Exemplarily, the weather forecast for the next 4 hours includes generating training samples, constructing a weather forecast model for the next 4 hours and training and selecting weights for the weather forecast model for the next 4 hours;

[0017] Generate training samples: Process the meteorological data set to obtain training samples, convert the training samples into the data shape required by the prediction model according to the prediction model, and divide the sample data into training set, validation set and test set according to a certain ratio;

[0018] Build and train a 4-hour weather forecast model: Build a 4-hour weather forecast model based on machine learning and deep learning networks, including NowcastNet, TimesNet, MTS-Mixer and their variants, and train the 4-hour weather forecast model based on training samples;

[0019] Weight selection of weather forecast model for the next 4 hours: The weather forecast model for the next 4 hours is evaluated using the validation set and tested using the test set, and the model weight with the best performance is selected as the model weight for subsequent weather forecasts for the next 4 hours.

[0020] Exemplarily, the photovoltaic output short-term prediction model training includes generating training samples, constructing and training the photovoltaic output short-term prediction model, and selecting weights of the photovoltaic output short-term prediction model.

[0021] Exemplarily, when generating training samples, the photovoltaic output data set and the meteorological data set are integrated, and the photovoltaic output, meteorological data, and time data information are integrated to obtain training samples. According to the prediction model, the training samples are converted into the data shape required by the prediction model, and the sample data is divided into training set, validation set, and test set according to a certain ratio.

[0022] Exemplarily, when building and training a photovoltaic output short-term prediction model, a photovoltaic output short-term prediction model is built based on machine learning and deep learning networks, including XGBoost, LSTM, MTS-Mixer and variants thereof, and the photovoltaic output short-term prediction model is trained based on training samples;

[0023] When selecting the weights of the photovoltaic output short-term prediction model, the photovoltaic output short-term prediction model is evaluated using the validation set, and tested using the test set, and the model weight with the best performance is selected as the model weight for subsequent photovoltaic output short-term prediction.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] The present invention can improve the accuracy of ultra-short-term prediction of photovoltaic output by means of a multi-model distributed photovoltaic output method. First, the acquired data is processed to obtain a high-quality data set after processing, and then the ultra-short-term prediction model of photovoltaic output is trained and constructed by using an Attention-CNN-LSTM model. The model uses an attention mechanism to evaluate the importance of each feature, uses a convolutional layer to extract a hierarchical structure of more abstract features, and performs feature fusion on the results of the two, and inputs them into the LSTM for ultra-short-term prediction. The model can effectively improve the ultra-short-term prediction result of photovoltaic output, and trains a meteorological prediction model for the next 4 hours. The model is constructed using NowcastNet, which can effectively process and predict weather phenomena of different scales, can quickly respond to weather changes, and generate relatively accurate meteorological data in the future within a short period of time. The photovoltaic output short-term prediction model is trained and constructed by using an Attention-MTS-Mixer model. Through the attention mechanism, the model can pay more attention to the important part of the input data, and generate an accurate photovoltaic output by using the predicted future meteorological data. Finally, the ultra-short-term prediction result of photovoltaic output is corrected according to the meteorological prediction result of the next 4 hours and the short-term prediction result of photovoltaic output to obtain a more accurate value.

[0026] Other features and advantages of the present invention will be described in the following description, and partly become obvious from the description, or be understood by implementing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A flow chart of distributed photovoltaic output prediction according to the present invention;

[0028] Figure 2 This is a schematic diagram of the CNN-LSTM model structure based on the attention mechanism of the present invention;

[0029] Figure 3 This is the framework diagram of the NowcastNet model of the present invention;

[0030] Figure 4 This is the structural diagram of the attention-based MTS-Mixer model of the present invention. DETAILED DESCRIPTION

[0031] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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.

[0032] The present invention provides a distributed photovoltaic ultra-short-term output prediction method based on multiple models, comprising the following steps:

[0033] Step 1: Obtain a data set, collect the original data set, and preprocess the original data set to obtain a sample data set;

[0034] Extract the data needed for each model from the original data, including time, historical output, historical meteorological data, etc.;

[0035] Step 2: Model training: construct a training model and train the PV output ultra-short-term prediction model, the next 4-hour weather prediction model, and the PV output short-term prediction model based on the data set;

[0036] Using the trained ultra-short-term photovoltaic output forecasting model, based on historical photovoltaic output data and historical meteorological data, preliminary ultra-short-term photovoltaic output forecasting results are obtained;

[0037] Use the trained 4-hour weather forecast model to obtain the 4-hour weather forecast results based on historical weather data;

[0038] Using the trained photovoltaic output short-term prediction model, the photovoltaic output short-term prediction results are obtained according to the predicted meteorological data for the next 4 hours;

[0039] Step 3: Model prediction results: According to the weather forecast results for the next 4 hours and the short-term forecast results for photovoltaic output, the preliminary ultra-short-term forecast results for photovoltaic output are corrected to obtain the final ultra-short-term forecast results for photovoltaic output and output them.

[0040] Obtaining the dataset includes:

[0041] Collect data, obtain historical photovoltaic output data sets from photovoltaic power plants and historical meteorological data sets from the Meteorological Bureau;

[0042] Data processing, removing duplicate data, deleting useless information, repairing all missing data, and eliminating abnormal data.

[0043] The ultra-short-term photovoltaic output prediction model includes generating training samples, constructing and training the ultra-short-term photovoltaic output prediction model, and selecting the weights of the ultra-short-term photovoltaic output prediction model.

[0044] When generating training samples, the photovoltaic output data set and the meteorological data set are integrated, and the photovoltaic output, meteorological data, and time data information are integrated to obtain training samples. According to the prediction model, the training samples are converted into the data shape required by the prediction model, and the sample data are divided into training set, validation set, and test set according to a certain ratio.

[0045] When building and training the ultra-short-term photovoltaic output prediction model, the ultra-short-term photovoltaic output prediction model is built based on machine learning and deep learning networks, including XGBoost, LSTM, Transformer and their variants. First, a single-step ultra-short-term photovoltaic output prediction training is performed, and then the photovoltaic output prediction for the next 4 hours is performed.

[0046] When selecting the weight of the ultra-short-term photovoltaic output prediction model, the ultra-short-term photovoltaic output prediction model is evaluated using the validation set, and tested using the test set, and the model weight with the best performance is selected as the model weight for subsequent ultra-short-term photovoltaic output prediction.

[0047] The weather forecast for the next 4 hours includes generating training samples, building and training the weather forecast model for the next 4 hours, and selecting the weights of the weather forecast model for the next 4 hours;

[0048] Generate training samples: Process the meteorological data set to obtain training samples, convert the training samples into the data shape required by the prediction model according to the prediction model, and divide the sample data into training set, validation set and test set according to a certain ratio;

[0049] Build and train a 4-hour weather forecast model: Build a 4-hour weather forecast model based on machine learning and deep learning networks, including NowcastNet, TimesNet, MTS-Mixer and their variants, and train the 4-hour weather forecast model based on training samples;

[0050] Weight selection of weather forecast model for the next 4 hours: The weather forecast model for the next 4 hours is evaluated using the validation set and tested using the test set, and the model weight with the best performance is selected as the model weight for subsequent weather forecasts for the next 4 hours.

[0051] The photovoltaic output short-term prediction model training includes generating training samples, constructing and training the photovoltaic output short-term prediction model, and selecting the photovoltaic output short-term prediction model weights.

[0052] When generating training samples, the photovoltaic output data set and the meteorological data set are integrated, and the photovoltaic output, meteorological data, and time data information are integrated to obtain training samples. According to the prediction model, the training samples are converted into the data shape required by the prediction model, and the sample data are divided into training set, validation set, and test set according to a certain ratio.

[0053] When building and training a photovoltaic output short-term prediction model, a photovoltaic output short-term prediction model is built based on machine learning and deep learning networks, including XGBoost, LSTM, MTS-Mixer and their variants, and the photovoltaic output short-term prediction model is trained based on training samples;

[0054] When selecting the weights of the photovoltaic output short-term prediction model, the photovoltaic output short-term prediction model is evaluated using the validation set, and tested using the test set, and the model weight with the best performance is selected as the model weight for subsequent photovoltaic output short-term prediction.

[0055] Embodiment 1:

[0056] The photovoltaic ultra-short-term prediction model training in this embodiment includes the following steps:

[0057] Step 1: Data acquisition and preprocessing: Integrate the historical photovoltaic output data obtained from the photovoltaic power station with the meteorological data obtained from the Meteorological Bureau, and integrate the data required for the model, including detailed time, installed capacity, historical photovoltaic output, temperature, humidity, wind direction, wind speed, rainfall, radiation intensity, cloud cover and other data that can affect photovoltaic output. Preprocess the data, including removing duplicate data, deleting useless information, repairing all missing data, and eliminating abnormal data.

[0058] Step 2, data conversion, according to the ultra-short-term photovoltaic output prediction model, convert the training samples into the data shape required by the prediction model, and divide the sample data into training set, verification set and test set according to a certain ratio. The ratio in this embodiment is 6:2:2.

[0059] Step 3: Constructing an ultra-short-term photovoltaic output prediction model 1 In this embodiment, a CNN-LSTM model based on the attention mechanism is used to construct an ultra-short-term photovoltaic output prediction model. The schematic diagram of the model structure is shown in FIG. Figure 1 As shown in the figure, the CNN-LSTM model based on the attention mechanism consists of three parts: the CNN unit based on the attention mechanism, the LSTM unit and the output unit. The CNN unit module based on the attention mechanism is based on the standard CNN module, and a parallel attention mechanism branch is added to extract significant features. The CNN module is stacked by multiple layers of one-dimensional networks. Each layer contains a convolution layer, a batch normalization layer, and a nonlinear layer. At the same time, sampling aggregation is achieved through the pooling layer. The stacking of convolution layers creates a hierarchical structure that gradually extracts more abstract features. The parallel attention mechanism will independently evaluate the importance of the input features and assign an attention weight to each feature. The more important the data feature, the higher its attention weight will be. Then, feature fusion is performed, and the output features of the CNN module are multiplied element by element by the attention weight output by the corresponding attention mechanism module to obtain a result that can reflect the importance of the feature. Finally, the fusion result is used as the input of the LSTM unit, and the LSTM is used for time series modeling to obtain the final prediction result and output it.

[0060] Step 4: Train the ultra-short-term prediction model for photovoltaic output, with a time resolution of 15 minutes and a prediction time of 15 minutes to 4 hours. First, perform single-step ultra-short-term prediction training for photovoltaic output. According to the photovoltaic output and meteorological data at the current time t, the photovoltaic output at the next time t+1 is predicted in combination with the historical photovoltaic output and meteorological data of the previous n time steps. Then, train the photovoltaic output for the next 4 hours. First, use the ultra-short-term prediction of photovoltaic output in a single step to predict the photovoltaic output at time t+1. Then, use the predicted photovoltaic output at time t+1 to predict the photovoltaic output at time t+2 in a single step, and so on, until the photovoltaic output of the 4th hour is predicted.

[0061] Step 5: Weight selection, set training hyperparameters, and select a suitable optimization method to train the model. In this embodiment, the loss function uses the root mean square error (RMSE) loss function, the optimization method uses Adam, the number of training cycles is 200, the batch size is 5000, and the initial learning rate is set to 0.001. The photovoltaic output ultra-short-term prediction model is evaluated using the validation set, and tested using the test set, and the model weight with the best performance is selected as the model weight for subsequent photovoltaic output ultra-short-term prediction.

[0062] Embodiment 2:

[0063] The weather forecast model training for the next 4 hours in this embodiment is as follows:

[0064] Step 1: Obtain data, collect historical meteorological data, and pre-process the data, including removing duplicate data, deleting useless information, repairing all missing data, and eliminating abnormal data;

[0065] Step 2: Data conversion: According to the weather forecast model for the next 4 hours, the training samples are converted into the data shape required by the forecast model, and the sample data is divided into a training set, a validation set, and a test set according to a certain ratio. The ratio in this embodiment is 6:2:2.

[0066] Step 3: Build a weather forecast model for the next 4 hours. This embodiment uses the NowcastNet model, which mainly includes an evolution network, a generative network, a physics-conditional mechanism, a neural evolution operator, and a temporal discriminator. Discriminator), the evolutionary network integrates physical mechanisms, takes historical data as input, uses the U-Net structure to predict momentum and residual, and uses the neural evolution operator to simulate the physical evolution in the precipitation process to generate future states that conform to physical laws. The generation network further refines the prediction results based on the output of the evolutionary network. The forecast encoder and forecast decoder use the output of the evolutionary network as the evolutionary condition for GAN training. NowcastNet uses the physical condition mechanism to minimize the prediction error. This mechanism ensures that the prediction results of the model not only meet the data-driven characteristics, but also follow the laws of physics, such as momentum conservation. In the training process of the generative adversarial network, the temporal discriminator is used to distinguish between real radar data and fake data generated by the model. The purpose of the discriminator is to train the generation network so that the prediction results it generates are consistent with the real data in terms of statistical characteristics;

[0067] Step 4: train a weather forecast model for the next 4 hours, input historical weather data into the weather forecast model for the next 4 hours, and obtain the forecast weather results for the next 4 hours;

[0068] Step 5: Weight selection. Adjust the hyperparameters of the evolved network and the generated network, select a suitable optimization method to train the model, use the validation set to evaluate the weather forecast model for the next 4 hours, and use the test set to test it. Select the model weight with the best performance as the model weight for the subsequent short-term photovoltaic output forecast.

[0069] Embodiment three:

[0070] The photovoltaic short-term prediction model training of this embodiment includes the following steps:

[0071] Step 1: Data acquisition and preprocessing: Integrate historical photovoltaic output data and meteorological data, and preprocess the data, including removing duplicate data, deleting useless information, repairing all missing data, and eliminating abnormal data.

[0072] Step 2, data conversion, according to the photovoltaic output short-term prediction model, the training samples are converted into the data shape required by the prediction model, and the sample data are divided into a training set, a validation set and a test set according to a certain ratio. The ratio in this embodiment is 6:2:2.

[0073] Step 3: Construct a photovoltaic output short-term prediction model. This embodiment constructs a photovoltaic output short-term prediction model based on the MTS-Mixer model of attention. The specific implementation formula is as follows:

[0074] Formula (1) is the time series data X h Creates an enhanced input representation Formula (2) is used to capture time series data Time dependence Formula (3) is used to extract and model channel dependencies in time series data

[0075]

[0076] Among them, norm() represents the normalization operation, which converts the original time series data into a unified scale. PE() represents the position encoding operation, which adds position information to the time series data. Here, only the Encoder part of the Transformer is used, and the Decoder part is removed, and it is directly changed to a fully connected mapping;

[0077] Step 4: training a photovoltaic output short-term prediction model, inputting the future meteorological data for the next 4 hours into the photovoltaic output short-term prediction model, and obtaining the photovoltaic output short-term prediction results for the next 4 hours;

[0078] Step 5: Weight selection, set training hyperparameters, and select appropriate optimization methods to train the model. Use the validation set to evaluate the photovoltaic output short-term prediction model, and use the test set to test it. Select the model weight with the best performance as the model weight for subsequent photovoltaic output short-term prediction.

[0079] Embodiment 4:

[0080] The photovoltaic output prediction in this embodiment includes the following steps:

[0081] Step 1: Obtain historical meteorological data and historical photovoltaic output data.

[0082] Step 2: Input historical meteorological data and historical photovoltaic output data into a pre-trained photovoltaic output ultra-short-term prediction model to obtain preliminary photovoltaic output ultra-short-term prediction results.

[0083] Step three, input the historical meteorological data into the pre-trained 4-hour weather forecast model to obtain the weather forecast results for the next 4 hours.

[0084] Step 4: Input the predicted future meteorological data into the pre-trained photovoltaic output short-term prediction model to obtain the photovoltaic output short-term prediction result.

[0085] Step 5: According to the short-term forecast result of photovoltaic output and the meteorological forecast result for the next 4 hours, the preliminary ultra-short-term forecast result of photovoltaic output is corrected to obtain the ultra-short-term forecast result of photovoltaic output.

[0086] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A distributed photovoltaic ultra-short-term output prediction method based on multiple models, characterized in that: The steps include: Step 1: Obtain a data set, collect the original data set, and preprocess the original data set to obtain a sample data set; Step 2: Model training: construct a training model and train the PV output ultra-short-term prediction model, the next 4-hour weather prediction model, and the PV output short-term prediction model based on the data set; Step 3: Model prediction results: According to the weather forecast results for the next 4 hours and the short-term forecast results for photovoltaic output, the preliminary ultra-short-term forecast results for photovoltaic output are corrected to obtain the final ultra-short-term forecast results for photovoltaic output and output them.

2. The method for ultra-short-term distributed photovoltaic output prediction based on multiple models according to claim 1 is characterized by: The acquiring of the data set comprises: Collect data, obtain historical photovoltaic output data sets from photovoltaic power plants and historical meteorological data sets from the Meteorological Bureau; Data processing, removing duplicate data, deleting useless information, repairing all missing data, and eliminating abnormal data.

3. The method for ultra-short-term distributed photovoltaic output prediction based on multiple models according to claim 1 is characterized in that: The ultra-short-term photovoltaic output prediction model includes generating training samples, constructing and training the ultra-short-term photovoltaic output prediction model, and selecting the weight of the ultra-short-term photovoltaic output prediction model.

4. The method for ultra-short-term distributed photovoltaic output prediction based on multiple models according to claim 3 is characterized by: When generating training samples, the photovoltaic output data set and the meteorological data set are integrated, and the photovoltaic output, meteorological data, and time data information are integrated to obtain training samples. According to the prediction model, the training samples are converted into the data shape required by the prediction model, and the sample data are divided into training set, validation set, and test set according to a certain ratio.

5. The method for ultra-short-term distributed photovoltaic output prediction based on multiple models according to claim 4 is characterized by: When building and training the ultra-short-term photovoltaic output prediction model, the ultra-short-term photovoltaic output prediction model is built based on machine learning and deep learning networks, including XGBoost, LSTM, Transformer and their variants. First, a single-step ultra-short-term photovoltaic output prediction training is performed, and then the photovoltaic output prediction for the next 4 hours is performed.

6. The method for ultra-short-term distributed photovoltaic output prediction based on multiple models according to claim 5 is characterized by: When selecting the weight of the ultra-short-term photovoltaic output prediction model, the ultra-short-term photovoltaic output prediction model is evaluated using the validation set, and tested using the test set, and the model weight with the best performance is selected as the model weight for subsequent ultra-short-term photovoltaic output prediction.

7. The method for ultra-short-term distributed photovoltaic output prediction based on multiple models according to claim 1 is characterized by: The weather forecast for the next 4 hours includes generating training samples, constructing and training a weather forecast model for the next 4 hours, and selecting weights for the weather forecast model for the next 4 hours; Generate training samples: Process the meteorological data set to obtain training samples, convert the training samples into the data shape required by the prediction model according to the prediction model, and divide the sample data into training set, validation set and test set according to a certain ratio; Build and train a 4-hour weather forecast model: Build a 4-hour weather forecast model based on machine learning and deep learning networks, including NowcastNet, TimesNet, MTS-Mixer and their variants, and train the 4-hour weather forecast model based on training samples; Weight selection of weather forecast model for the next 4 hours: The weather forecast model for the next 4 hours is evaluated using the validation set and tested using the test set, and the model weight with the best performance is selected as the model weight for subsequent weather forecasts for the next 4 hours.

8. The method for ultra-short-term distributed photovoltaic output prediction based on multiple models according to claim 1 is characterized by: The photovoltaic output short-term prediction model training includes generating training samples, constructing and training the photovoltaic output short-term prediction model, and selecting the photovoltaic output short-term prediction model weights.

9. The method for ultra-short-term distributed photovoltaic output prediction based on multiple models according to claim 8 is characterized by: When generating training samples, the photovoltaic output data set and the meteorological data set are integrated, and the photovoltaic output, meteorological data, and time data information are integrated to obtain training samples. According to the prediction model, the training samples are converted into the data shape required by the prediction model, and the sample data are divided into training set, validation set, and test set according to a certain ratio.

10. The method for ultra-short-term distributed photovoltaic output prediction based on multiple models according to claim 8, characterized in that: When building and training a photovoltaic output short-term prediction model, a photovoltaic output short-term prediction model is built based on machine learning and deep learning networks, including XGBoost, LSTM, MTS-Mixer and their variants, and the photovoltaic output short-term prediction model is trained based on training samples; When selecting the weights of the photovoltaic output short-term prediction model, the photovoltaic output short-term prediction model is evaluated using the validation set, and tested using the test set, and the model weight with the best performance is selected as the model weight for subsequent photovoltaic output short-term prediction.

Images