Photovoltaic power intelligent prediction method and system integrating mechanism model and field measurement

Abstract

The present invention belongs to the field of photovoltaic power prediction technology, and specifically relates to a photovoltaic power intelligent prediction method and system that integrates a mechanism model and field station measurements. The photovoltaic power intelligent prediction method that integrates a mechanism model and field station measurements of the present invention collects and organizes historical field station measured data and pattern data, uses pattern data to add physical constraints to the training model, improves the accuracy of the training model through field station measured data, and finally integrates the output results to achieve the purpose of complementary advantages of the two types of data, thereby achieving the effect of improving the accuracy of photovoltaic power generation prediction; and by distinguishing between stable and severe weather types, using a CNN+Transformer model under stable weather conditions and a CNN+BiLSTM model under severe weather conditions, further improving the accuracy of photovoltaic power generation prediction, saving computing resources, and improving prediction efficiency.

Description

Technical field:

[0001] The present invention belongs to the technical field of photovoltaic power prediction, and specifically relates to a photovoltaic power intelligent prediction method and system integrating a mechanism model and actual measurements at a site. Background technology:

[0002] According to investigation, with the rapid development of renewable energy around the world, new energy has gradually transformed from supplementary energy to supporting energy in the power system, and solar energy has become one of the most ideal alternative energy sources. Photovoltaic power generation technology is crucial to improving the stability of the power system. The accurate prediction of its power generation capacity depends on high-precision meteorological forecast results, which plays a very important role in the stable and safe operation of the power system.

[0003] Traditional photovoltaic power generation prediction includes simple physical model prediction method and complex physical model prediction method. The first method uses satellite detection and meteorological observation data, combined with power parameters, to predict photovoltaic power generation; the second method combines photovoltaic power generation principles and numerical weather prediction models, considers multiple weather factors, and improves prediction accuracy. However, traditional photovoltaic power generation prediction methods also have certain defects: in terms of physical model methods, the accuracy is limited, which is reflected in the fact that traditional physical models are difficult to accurately capture all factors affecting photovoltaic power generation and rely on the quality of data; the harshness is reflected in the need to consider many factors such as light radiation, wind, weather conditions, etc. In photovoltaic power generation prediction, the precise weight of each meteorological element is uncertain, which poses a challenge to the establishment of an accurate photovoltaic power generation prediction model. In terms of historical data methods, there is a lack of real-time performance, and it is impossible to effectively reflect the current environmental changes and weather conditions in a timely manner, and ignore the impact of external factors on photovoltaic power generation.

[0004] Although existing technologies have used deep learning technology to discover deep features for power prediction and have achieved certain results, if deep learning technology only uses the measured data of the station and only considers the relationship between the data, it will abandon the traditional classical meteorological dynamic constraints; but if only meteorological model data is used, the accuracy of photovoltaic output prediction for the station area will not meet the standards due to the inaccuracy of meteorological data. Summary of the invention:

[0005] The purpose of the present invention is to provide a method and system for intelligent prediction of photovoltaic power that integrates mechanism models and field station measurements. By collecting and organizing historical field station measured data and pattern data, using pattern data to add physical constraints to the training model, and using field station measured data to improve the accuracy of the training model, the output results are finally integrated to achieve the purpose of complementing the advantages of the two data and realize the effect of improving the accuracy of photovoltaic power prediction.

[0006] The first aspect of the present invention provides a photovoltaic power intelligent prediction method integrating a mechanism model and field station measurements, comprising the following steps:

[0007] Obtain historical site measured data and model data;

[0008] The historical station measured data are divided into the historical station measured data under stable weather and the historical station measured data under severe weather, and the model data are divided into the model data under stable weather and the model data under severe weather;

[0009] Preprocess the data;

[0010] Training models to obtain stable weather measured models, severe weather measured models, stable weather physical models, and severe weather physical models;

[0011] Obtain weather model forecast data to determine weather stability;

[0012] If the weather stability meets the stable weather conditions, the stable weather measured model and the stable weather physical model are used for prediction respectively, and the first photovoltaic output sequence and the second photovoltaic output sequence are obtained by inverse normalization; the first photovoltaic output sequence and the second photovoltaic output sequence are fused by the variational method to obtain the photovoltaic output prediction sequence; and

[0013] If the weather stability meets the severe weather conditions, the severe weather measured model and the severe weather physical model are used for prediction respectively, and the third photovoltaic output sequence and the fourth photovoltaic output sequence are obtained by inverse normalization; the third photovoltaic output sequence and the fourth photovoltaic output sequence are fused through the variational method to obtain the photovoltaic output prediction sequence.

[0014] The second aspect of the present invention provides a photovoltaic power intelligent prediction system integrating a mechanism model and field station measurements, comprising:

[0015] Data acquisition module, used to obtain historical station measured data and model data;

[0016] A data splitting module is used to split the historical station measured data into the historical station measured data under stable weather and the historical station measured data under severe weather, and split the model data into the model data under stable weather and the model data under severe weather;

[0017] A data preprocessing module, used for preprocessing data;

[0018] A model training module is used to train the model to obtain a stable weather measured model, a severe weather measured model, a stable weather physical model, and a severe weather physical model;

[0019] The weather stability judgment module is used to obtain weather model forecast data and judge the weather stability;

[0020] The stable weather condition prediction module is used to use the stable weather measured model and the stable weather physical model to make predictions respectively when the weather stability meets the stable weather conditions, and to obtain the first photovoltaic output sequence and the second photovoltaic output sequence by denormalization; the first photovoltaic output sequence and the second photovoltaic output sequence are merged by the variational method to obtain the photovoltaic output prediction sequence;

[0021] The severe weather condition prediction module is used to use the severe weather measured model and the severe weather physical model to make predictions respectively when the weather stability meets the severe weather conditions, and to obtain the third photovoltaic output sequence and the fourth photovoltaic output sequence by denormalization; the third photovoltaic output sequence and the fourth photovoltaic output sequence are merged by the variational method to obtain the photovoltaic output prediction sequence.

[0022] A third aspect of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is loaded into the processor, the method for intelligently predicting photovoltaic power based on the fusion mechanism model and actual measurements at the site as described above is implemented.

[0023] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a photovoltaic power intelligent prediction method based on the fusion mechanism model and actual measurements at the site as described above.

[0024] The present invention has the following beneficial effects: the fusion mechanism model and the photovoltaic power intelligent prediction method measured by the site of the present invention collect and organize historical site measured data and pattern data, use the pattern data to add physical constraints to the training model, improve the accuracy of the training model through the site measured data, and finally fuse the output results to achieve the purpose of complementing the advantages of the two types of data, thereby achieving the effect of improving the accuracy of photovoltaic power prediction; and by distinguishing between stable and severe weather types, using the CNN+Transformer model under stable weather conditions and the CNN+BiLSTM model under severe weather conditions, the accuracy of photovoltaic power prediction is further improved, computing resources are saved, and prediction efficiency is improved. Description of the drawings:

[0025] Figure 1 It is a flow chart of the invented fusion mechanism model and the photovoltaic power intelligent prediction method measured at the site;

[0026] Figure 2 This is a schematic diagram of sliding window

[0027] Figure 3 It is a scatter plot of surface solar radiation and historical photovoltaic actual output before data cleaning;

[0028] Figure 4 It is a scatter plot of surface solar radiation and historical photovoltaic actual output after SVR cleaning;

[0029] Figure 5 This is a schematic diagram of the Transformer time series prediction architecture;

[0030] Figure 6 This is the loss function drop graph of CNN+BiLSTM;

[0031] Figure 7 This is the performance of CNN+BiLSTM on the test set. Specific implementation method:

[0032] See also Figure 1 The first aspect of the present invention provides a photovoltaic power intelligent prediction method integrating a mechanism model and actual measurements at a site, comprising the following steps:

[0033] The first step is to obtain historical site measured data and model data.

[0034] In this embodiment, historical station measured data and model data every 15 minutes can be collected and organized, where the station measured data includes historical station meteorological measured data and photovoltaic power generation data, and the model data can use ECMWF (European Centre for Medium-Range Weather Forecasts) data, with longitude selected from 119.22 to 119.24 east longitude and latitude selected from 34.06 to 34.10 north latitude, integrated into single grid data.

[0035] Optionally, the model data can be matched with the actual output during the modeling period, and the matching method is the "nearest" nearest neighbor method.

[0036] The second step is to split the data, splitting the historical station measured data into historical station measured data under stable weather and historical station measured data under severe weather, and splitting the model data into model data under stable weather and model data under severe weather.

[0037] Specific:

[0038] The Sandberg index SI is used as the atmospheric stability criterion, the formula is: SI = T 500 -T S , where T 500 The ambient temperature is 500hPa, T S It is the temperature of the air mass at 850hPa when the moist air mass rises along the dry adiabatic line and reaches the lifting condensation height, and then rises along the moist adiabatic line to 500hPa. If SI>0, the air layer is preliminarily judged to be stable, and if SI<0, the air layer is judged to be unstable.

[0039] If the time nodes with SI < 0 on that day account for more than 25% of the total time nodes on that day, the historical station measured data and model data on that day will be classified as historical station measured data under severe weather and model data under severe weather respectively; otherwise

[0040] The historical station measured data and model data for that day are classified as historical station measured data under stable weather and model data under stable weather, respectively.

[0041] The third step is to preprocess the data, including:

[0042] Missing value processing: visualize data distribution characteristics, check missing values, and use linear interpolation to process missing values;

[0043] Normalization processing: Normalize the characteristics of each meteorological element and the power generation power by column and then align the data. The formula is: where x min is the minimum value of the column, x max is the maximum value of the column;

[0044] Feature dimensionality reduction processing: Get the contribution ranking of each meteorological element to the output power of photovoltaic power generation, and realize feature dimensionality reduction according to the ranking result. Specifically, the LightGBM model can be used to implement the GBDT algorithm. Boosting obtains a set of weak classifiers through training and combines them into a strong classifier to improve regression performance. The main idea of ​​the GBDT algorithm is to use weak classifiers (decision trees) for iterative training to obtain the optimal model. The LightGBM model is an efficient gradient boosting decision tree (GBDT) framework. The commonly used LightGBM regressor parameters in this model include: "objective" is used to specify the learning task, "num_leaves" controls the complexity of the tree, "learning_rate" adjusts the step size, "n_estimators" determines the number of trees, and "early_stopping_round" prevents overfitting. These parameters jointly affect the complexity, learning speed and stopping criteria of the model. The model uses LGBMRegressor (num_leaves = 31) to create a regressor object, performs GridSearchCV cross-validation through the parameter grid containing "learning_rate", "n_estimators" and "early_stopping_round", and fits the model on the training set and evaluation set, and finally finds the best parameter combination and prints it out. The model supports efficient parallel training, obtains the contribution ranking of various meteorological elements to photovoltaic power output power, realizes feature dimensionality reduction based on the ranking results, saves computing costs, and tends to more important features.

[0045] Time feature column construction: Calculate the total number of seconds for each data point starting from midnight of the day, divide it by the total number of seconds in a day, and record it in the "time ratio" column. This can record the time of the day when the data point is located, which helps the model capture the periodic information of photovoltaic power generation;

[0046] Sliding windowing: Use the power data at the current moment and then further process the feature data. The method used is sliding windowing. The specific operation is to replace the feature of each time in the array with the feature sequence of the past 24 hours (including the past 96 times of this time), keep the power output value (label) corresponding to each position unchanged, and finally obtain the processed pattern and the measured feature label pair of the station; see Figure 2 , which is a schematic diagram of sliding windowing.

[0047] Data cleaning: Since power generation output is related to human factors, such as equipment updates and corporate decisions, the gap between actual output and theoretical output may be too large. In view of the fact that photovoltaic power generation output data is theoretically highly periodic and strongly correlated with solar radiation. In this embodiment, surface solar radiation (SSR) can be selected as a typical radiation variable, and a support vector regression algorithm (SVR) can be used to perform a preliminary regression prediction of the output based on the surface solar radiation (SSR); an error threshold (for example, 160) with the actual output is set, and the output values ​​and their corresponding feature data that are higher than the error threshold are eliminated.

[0048] Figure 3 It is a scatter plot of surface solar radiation and historical photovoltaic actual output before data cleaning; Figure 4 It is a scatter plot of surface solar radiation after SVR cleaning and historical actual photovoltaic output.

[0049] The fourth step is to train the model to obtain the stable weather measured model, the severe weather measured model, the stable weather physical model and the severe weather physical model.

[0050] Specifically, the method for training the model includes:

[0051] The pre-processed historical station measured data under stable weather conditions, historical station measured data under severe weather conditions, model data under stable weather conditions, and model data under severe weather conditions are divided into training sets and test sets at a ratio of 7:3. To avoid the problems of training set leakage and data imbalance, the existing data can be classified and aggregated, with the first seven days as the training set and the last three days as the test set.

[0052] The CNN+Transformer model is trained and tested using historical station measured data under stable weather and model data under stable weather to obtain a stable weather measured model and a stable weather physical model; and

[0053] The pre-processed historical station measured data under severe weather and pattern data under severe weather were used to train and test the CNN+BiLSTM model to obtain the severe weather measured model and severe weather physical model. Figure 6 It is the loss function drop graph of the severe weather measured model.

[0054] In the embodiment, preferably, a priori analysis is performed on the predicted sequence to distinguish between stable and severe weather types, so as to select a model with better prediction results for long / short sequences. Specifically, the CNN+Transformer and CNN+BiLSTM models designed in this embodiment, in which CNN is used for feature dimensionality reduction; the advantage of the Transformer model is that it can capture long sequence features in the data without the need for cyclic connections, which makes the model more advantageous than some recurrent neural network-based models such as GRU and LSTM. After comparison, although the LSTM model is designed to process long time series, experiments have shown that it cannot effectively preserve long time series or capture long-term dependencies, thereby reducing the prediction performance; the Transformer model performs better in capturing long-term dependencies, surpassing the LSTM model; and the self-attention mechanism of the Transformer model is too complicated when processing short sequences, resulting in unnecessary computational overhead. Therefore, this solution trains two parallel neural network models.

[0055] In this embodiment, optionally, the CNN structure is as follows: the input layer accepts input data of a shape of (batch_size, time_steps, input_dims); the number of input channels of convolution layer A is equal to the input feature dimension, the number of output channels is 64, and the convolution kernel size is set to 1, which is used to extract input data features and realize feature dimensionality reduction; batching layer A batches the output of convolution layer A to improve stability and accelerate training; pooling layer A performs pooling operation on the output of batching layer A to reduce computational complexity; then enters convolution layer B with an input channel number of 64 and an output channel number of 128, and the size of each convolution kernel is 1 for further feature extraction; then passes through batching layer B and pooling layer B; enters the Dropout layer, and the ratio is set to 0.2 to prevent overfitting; finally, the forward transmission enters the time series prediction part: Transformer / LSTM.

[0056] In this embodiment, see Figure 5 , the Transformer module can adopt the traditional Transformer network self-attention mechanism; the BiLSTM model can adopt the traditional bidirectional long short-term memory network BiLSTM.

[0057] In this embodiment, optionally, the prediction error may be reduced by adjusting model parameters, optimizing network structure, and the like.

[0058] Regarding model parameter adjustment: For LSTM networks, the optimizer can be adjusted to SGD or Adam. The "sequence_length" sequence length is adjusted according to actual needs and computing resources to indicate the length of historical information considered. The "learning_rate" learning rate, "hidden_size" hidden layer size, and "batch_size" batch size of the input data are optimized to make the model converge to the optimal value. A dynamic learning rate strategy is adopted. When the loss does not improve within a certain epoch, the learning rate is reduced to jump out of the local optimum. The solution uses TPE (Tree-structured Parzen Estimator) to determine the optimal learning rate of 0.04 and the number of LSTM units of 38. For Transformer, similar parameters also need to be adjusted. At the same time, since Transformer does not have a cyclic structure, the number of attention heads, the number of stacked layers of encoders and decoders, and parameters related to position encoding also need to be adjusted.

[0059] Regarding the adjustment of the network structure, it is proposed to introduce GRU gated recurrent units to improve the multi-layer LSTM model. The GRU unit model has fewer parameters and higher computational efficiency. Replacing some layers of LSTM with GRU units can speed up the computational efficiency while combining the advantages of LSTM in capturing long-term dependencies. At the same time, the Informer model, a variant of Transformer, can also be used to improve the prediction. The model introduces sampling and sparsification strategies to calculate the mutual information between the query (Q) and the key (K) to evaluate the sparsity of attention and select elements with high mutual information for attention calculation, effectively reducing the computational complexity of self-attention from O(n / 2). 2 ) is reduced to O(n×log(n)).

[0060] In this embodiment, optionally, the evaluation index of the loss function may be root mean square error (RMSE) or mean absolute error (MAE); preferably, the mean absolute error (MAE).

[0061] The fifth step is to obtain weather model forecast data to determine weather stability.

[0062] Optionally, the Sandberg index SI can be used as an atmospheric stability criterion, the formula is: SI = T 500 -T S , where T 500 The ambient temperature is 500hPa, T SIt is the temperature of the air mass at 850hPa when the moist air mass rises along the dry adiabatic line and reaches the lifting condensation height, and then rises along the moist adiabatic line to 500hPa. If SI>0, the air layer is preliminarily judged to be stable, and it is more suitable to use Transformer to capture long-term dependencies and make predictions. If SI<0, the air layer is judged to be unstable. The larger the negative value, the more unstable the air layer is, and it is suitable to use LSTM to capture short-term dependencies and make time series predictions.

[0063] Step 6: If the weather stability meets the stable weather conditions, the stable weather measured model and the stable weather physical model are used for prediction respectively, and the first photovoltaic output sequence and the second photovoltaic output sequence are obtained by denormalization; the first photovoltaic output sequence and the second photovoltaic output sequence are fused by the variational method to obtain the photovoltaic output prediction sequence; and

[0064] If the weather stability meets the severe weather conditions, the severe weather measured model and the severe weather physical model are used for prediction respectively, and the third photovoltaic output sequence and the fourth photovoltaic output sequence are obtained by inverse normalization; the third photovoltaic output sequence and the fourth photovoltaic output sequence are fused through the variational method to obtain the photovoltaic output prediction sequence.

[0065] In one application scenario, optionally, the weather model forecast data for the next 48 hours is obtained to determine the weather stability and perform preprocessing; according to the type of weather conditions, the data is input into the corresponding model and denormalized to output two photovoltaic output sequences; the denormalization formula is:

[0066] Y=Y'×(x max -x min )+x min ;

[0067] where x min is the minimum power value in the historical photovoltaic power generation data, x max is the maximum power value in the historical photovoltaic power data, Y' is the model output power value, and Y is the denormalized power value;

[0068] The method of fusing two output sequences by variational method can be:

[0069] The one-dimensional variational fusion output power value is used, the second photovoltaic output sequence or the fourth photovoltaic output sequence is used as the background field, and the first photovoltaic output sequence or the third photovoltaic output sequence is used as the observation field accordingly. The cost function

[0070] J(x)=(xx b ) T B -1 (xx b )+(H(x)-y) T R-1 (H(x)-y);

[0071] where x is the analytical field, x b is the background field, B is the background field error covariance matrix, H is the observation operator, y is the observation data, and R is the observation error covariance matrix.

[0072] On the basis of the above-mentioned embodiments, the second aspect of the present invention further provides a photovoltaic power intelligent prediction system integrating a mechanism model and field station measurements, including:

[0073] Data acquisition module, used to obtain historical station measured data and model data;

[0074] A data splitting module is used to split the historical station measured data into the historical station measured data under stable weather and the historical station measured data under severe weather, and split the model data into the model data under stable weather and the model data under severe weather;

[0075] A data preprocessing module, used for preprocessing data;

[0076] A model training module is used to train the model to obtain a stable weather measured model, a severe weather measured model, a stable weather physical model, and a severe weather physical model;

[0077] The weather stability judgment module is used to obtain weather model forecast data and judge the weather stability;

[0078] The stable weather condition prediction module is used to use the stable weather measured model and the stable weather physical model to make predictions respectively when the weather stability meets the stable weather conditions, and to obtain the first photovoltaic output sequence and the second photovoltaic output sequence by denormalization; the first photovoltaic output sequence and the second photovoltaic output sequence are merged by the variational method to obtain the photovoltaic output prediction sequence;

[0079] The severe weather condition prediction module is used to use the severe weather measured model and the severe weather physical model to make predictions respectively when the weather stability meets the severe weather conditions, and to obtain the third photovoltaic output sequence and the fourth photovoltaic output sequence by denormalization; the third photovoltaic output sequence and the fourth photovoltaic output sequence are merged by the variational method to obtain the photovoltaic output prediction sequence.

[0080] In this embodiment, the working method of each module refers to the above embodiment and will not be repeated here.

[0081] Based on the above embodiments, the third aspect of the present invention further provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is loaded into the processor, it implements a photovoltaic power intelligent prediction method based on the fusion mechanism model and actual measurements at the site as described above.

[0082] Based on the above embodiments, the fourth aspect of the present invention further provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements a photovoltaic power intelligent prediction method based on the fusion mechanism model and actual measurements of the site as described above.

Claims

1. A photovoltaic power intelligent prediction method integrating mechanism model and field station measurement, characterized in that: The following steps are involved: Acquire historical station measured data and model data; the model data is weather forecast data; The historical station measured data are divided into the historical station measured data under stable weather and the historical station measured data under severe weather, and the model data are divided into the model data under stable weather and the model data under severe weather; Preprocess the data; Training models to obtain a stable weather measured model, a severe weather measured model, a stable weather physical model, and a severe weather physical model; the method for training models comprises: using pre-processed historical station measured data under stable weather and pattern data under stable weather to train CNN+Transformer models to obtain a stable weather measured model and a stable weather physical model; and using pre-processed historical station measured data under severe weather and pattern data under severe weather to train CNN+BiLSTM models to obtain a severe weather measured model and a severe weather physical model; Obtain weather model forecast data to determine weather stability; in If the weather stability meets the stable weather conditions, the stable weather measured model and the stable weather physical model are used for prediction respectively, and the first photovoltaic output sequence and the second photovoltaic output sequence are obtained by inverse normalization; the first photovoltaic output sequence and the second photovoltaic output sequence are fused by the variational method to obtain the photovoltaic output prediction sequence; and If the weather stability meets the severe weather conditions, the severe weather measured model and the severe weather physical model are used for prediction respectively, and the third photovoltaic output sequence and the fourth photovoltaic output sequence are obtained by inverse normalization; the third photovoltaic output sequence and the fourth photovoltaic output sequence are fused through the variational method to obtain the photovoltaic output prediction sequence.

2. The photovoltaic power intelligent prediction method according to claim 1, characterized in that: The historical site measured data includes historical site meteorological measured data and historical photovoltaic power generation data.

3. The photovoltaic power intelligent prediction method according to claim 1, characterized in that: The method of splitting the historical station measured data into the historical station measured data under stable weather and the historical station measured data under severe weather, and splitting the model data into the model data under stable weather and the model data under severe weather comprises: Use the Sandberg Index SI as a criterion for atmospheric stability; If the time nodes with SI < 0 on that day account for more than 25% of the total time nodes on that day, the historical station measured data and model data on that day will be classified as historical station measured data under severe weather and model data under severe weather respectively; otherwise The historical station measured data and model data for that day are classified as historical station measured data under stable weather and model data under stable weather, respectively.

4. The photovoltaic power intelligent prediction method according to claim 1, characterized in that: The data preprocessing includes: missing value processing, normalization processing, time feature column construction, sliding windowing, and feature dimension reduction processing.

5. The photovoltaic power intelligent prediction method according to claim 1, characterized in that: The feature dimension reduction process includes: Obtain the contribution ranking of each meteorological element to the output power of photovoltaic power generation, and realize feature dimensionality reduction based on the ranking results.

6. The photovoltaic power intelligent prediction method according to claim 4, characterized in that: The preprocessing of data includes data cleaning, which includes: Surface solar radiation is selected as a typical radiation variable, and the support vector regression algorithm is used to make a preliminary regression prediction of the output based on the surface solar radiation. Set an error threshold with the actual output, and remove output values ​​and their corresponding feature data that are higher than the error threshold.

7. A photovoltaic power intelligent prediction system integrating mechanism model and field station measurement, characterized in that: include: A data acquisition module is used to acquire historical station measured data and model data; the model data is weather forecast data; A data splitting module is used to split the historical station measured data into the historical station measured data under stable weather and the historical station measured data under severe weather, and split the model data into the model data under stable weather and the model data under severe weather; A data preprocessing module, used for preprocessing data; A model training module is used to train the model to obtain a stable weather measured model, a severe weather measured model, a stable weather physical model and a severe weather physical model; the method for training the model comprises: using the pre-processed historical station measured data under stable weather and the pattern data under stable weather to respectively train the CNN+Transformer model to obtain a stable weather measured model and a stable weather physical model; and using the pre-processed historical station measured data under severe weather and the pattern data under severe weather to respectively train the CNN+BiLSTM model to obtain a severe weather measured model and a severe weather physical model; The weather stability judgment module is used to obtain weather model forecast data and judge the weather stability; The stable weather condition prediction module is used to use the stable weather measured model and the stable weather physical model to make predictions respectively when the weather stability meets the stable weather conditions, and to obtain the first photovoltaic output sequence and the second photovoltaic output sequence by denormalization; the first photovoltaic output sequence and the second photovoltaic output sequence are merged by the variational method to obtain the photovoltaic output prediction sequence; The severe weather condition prediction module is used to use the severe weather measured model and the severe weather physical model to make predictions respectively when the weather stability meets the severe weather conditions, and to obtain the third photovoltaic output sequence and the fourth photovoltaic output sequence by denormalization; the third photovoltaic output sequence and the fourth photovoltaic output sequence are merged by the variational method to obtain the photovoltaic output prediction sequence.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into the processor, the method for intelligently predicting photovoltaic power based on the fusion mechanism model and actual measurements at the site according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for intelligently predicting photovoltaic power based on the fusion mechanism model and actual measurements at a site according to any one of claims 1 to 6 is implemented.

Images