Photovoltaic output prediction method based on multi-model improved neural network fusion

Abstract

The invention discloses a photovoltaic output prediction method based on multi-model improved neural network fusion. The method comprises the following steps: preprocessing historical meteorological data and historical photovoltaic output data; taking historical meteorological data as input, taking actual photovoltaic output as output, and training an HI model and an Informer model until an output error is smaller than a set threshold value; taking outputs of the HI model and the Informer model as inputs, taking actual photovoltaic output as an output, and training the LSTM model until an output error is smaller than a set threshold value; obtaining weather information at a prediction moment, and preprocessing the weather information at the prediction moment; weather information at the prediction moment is input into the HI model and the Informer model, and photovoltaic output is predicted; and inputting the outputs of the HI model and the Informer model into an LSTM model, and fusing the predicted photovoltaic output by using the LSTM model to obtain the final predicted photovoltaic output.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic output prediction of power systems, and in particular to a photovoltaic output prediction method based on LSTM neural network fusion of HI-Informe. Background Art

[0002] With the increasing emphasis on sustainable development and environmental protection, the development and utilization of new energy has become a key link in the transformation of the country's energy structure. Among them, photovoltaic power generation, as one of the most promising new energy sources, is rapidly expanding its application in the power system due to its advantages of cleanliness, renewableness, and wide distribution. From rooftop distributed photovoltaics to large-scale ground photovoltaic power stations, photovoltaic power sources have sprung up like mushrooms after a rain, injecting new vitality into the energy supply of the power system and effectively promoting the optimization of the energy structure and green and low-carbon development.

[0003] However, with the rapid increase in photovoltaic installed capacity, large-scale photovoltaic grid connection has brought unprecedented challenges to the dispatching and operation of traditional power systems. The output characteristics of photovoltaic power sources are highly dependent on natural conditions, especially sunshine intensity and weather changes, which makes photovoltaic output significantly intermittent and uncertain. This characteristic not only increases the difficulty of balancing power supply and demand, but may also have an adverse impact on the stability, safety and economy of the power system, such as causing grid voltage fluctuations, frequency deviations, line overloads and other problems, and in severe cases even threaten the reliable operation of the entire power system.

[0004] Therefore, in order to effectively respond to the challenges brought by large-scale photovoltaic grid connection and ensure the safe, stable and efficient operation of the power system, it is particularly important to make high-precision predictions of photovoltaic output. Through accurate modeling and prediction of photovoltaic output status, the power system dispatching department can grasp the power generation trend of photovoltaic power sources in advance, reasonably arrange the power generation plans of conventional energy and new energy, optimize resource allocation, and effectively alleviate the grid dispatching pressure caused by photovoltaic output fluctuations.

[0005] At the same time, accurate photovoltaic output forecasts can also provide decision-making support for the stable operation of the power system. For example, by dynamically adjusting reactive power compensation equipment, implementing demand-side management and other measures, it can deal with possible voltage stability and frequency control problems, and further enhance the flexibility and resilience of the power system.

[0006] However, photovoltaic output is affected by many factors besides meteorological factors. Secondary factors such as photovoltaic shading rate and cloud dynamics cannot be ignored. Although they are not as intuitive and significant as meteorological factors, they are difficult to accurately model and predict, which also poses a challenge to the accuracy of photovoltaic output prediction and increases the complexity of power system scheduling and operation. Therefore, it is necessary to develop new methods to realize online voltage stability risk assessment under new distribution networks, so as to effectively evaluate and warn the operation risks of distribution networks. Summary of the invention

[0007] The purpose of the present invention is to overcome the defects and shortcomings of the prior art and provide a photovoltaic output prediction method based on multi-model improved neural network fusion to achieve more accurate photovoltaic output prediction, improve the perception ability of the power system, and improve the safe and stable operation and effective scheduling ability of the power system.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is as follows: A photovoltaic output prediction method based on multi-model improved neural network fusion includes the following steps: S1: pre-processing the historical meteorological data and the historical photovoltaic output data to obtain standardized historical meteorological data and historical photovoltaic output data; S2: Take the historical meteorological data as input and the actual photovoltaic output of the historical photovoltaic output data as output, and train the HI model and the Informer model until the error between the output of the HI model and the Informer model and the actual photovoltaic output is less than the set threshold; S3: Take the output of the HI model and the Informer model as input, and the actual photovoltaic output of the historical photovoltaic output data as output, and train the LSTM model until the error between the output of the LSTM model and the actual photovoltaic output is less than the set threshold; S4: Obtaining meteorological information at the forecast time from the meteorological forecast department, preprocessing the meteorological information at the forecast time, and obtaining standardized meteorological information at the forecast time; S5: inputting the meteorological information at the prediction time into the HI model and the Informer model, predicting the photovoltaic output based on the HI model and the Informer model, and obtaining the output of the HI model and the Informer model; S6: Input the outputs of the HI model and the Informer model into the LSTM model, and use the LSTM model to fuse the predicted photovoltaic outputs of the HI model and the Informer model to obtain the final predicted photovoltaic output.

[0009] Furthermore, the preprocessing methods for the historical meteorological data and the historical photovoltaic output data include: screening outliers, filling missing values, and correcting over-limit values.

[0010] Furthermore, the photovoltaic output is predicted based on the HI model, specifically: The input sequence of the HI model is: given a length at the current moment Time Series , expressed as: ; ; In the formula, Indicated in A single feature or multiple feature inputs observed or measured at any time, where ; If the number of features in the input sequence is 1, then Contains only a single feature ,Right now , if the number of features in the input sequence exceeds 1, then Expressed as , is the number of features of the input sequence; The goal of time series forecasting is to Step prediction corresponding length sequence , the output sequence, i.e. the predicted sequence, is expressed as: ; ; In the formula, Indicates the expected prediction A single feature or multiple feature prediction output at the moment, where ,That Time means advance The corresponding time when the step is predicted; if the number of features of the input sequence is 1, the output sequence It also contains only a single feature ,Right now , if the number of features in the input sequence exceeds 1, the output sequence is , the number of features of the output sequence The number of features of the input sequence corresponding; The HI model takes the input sequence To predict the length of the subsequence The sequence output is taken as the prediction result, that is, its prediction result is expressed as: ; In the formula, The value is equal to the output sequence In , Represents the input sequence middle Single feature or multiple feature input at a moment; The value is equal to the output sequence In , Represents the input sequence middle Single feature or multiple feature input at a moment; The value is equal to the output sequence In , Represents the input sequence middle Single feature or multiple feature input at a moment; The value is equal to the output sequence In , Represents the input sequence middle A single feature or multiple feature inputs at a time.

[0011] Furthermore, the photovoltaic output is predicted based on the Informer model, specifically: The input to the encoder of the Informer model consists of a length of ,have Features, i.e., size The encoder uses a multi-head probabilistic sparse self-attention mechanism. The calculation formula is: ; Where Attention(·) is the attention function; , , They are the query matrix, key matrix, and value matrix of the same size obtained after linear transformation of the feature variables of the input sequence; is the number of features of the input sequence; , , They are , , The row dimension of represents the set of real numbers; is the activation function; Express The matrix obtained after thinning by probability is is with Matrix is ​​a sparse matrix of the same size.

[0012] Furthermore, in order to compress the feature dimension and extract the dominant attention variables in the input, the encoder uses self-attention distillation technology. The self-attention distillation technology re-superimposes one-dimensional convolution on the output of the previous layer and then performs a maximum pooling operation; after multiple multi-head probabilistic sparse self-attention and self-attention distillation, the encoder finally outputs a reduced feature map; from Layer to The distillation process of the layer is expressed as: ; In the formula, and They are Moment Layer and Layer attention input; is the maximum pooling function; is the activation function, , for Input, is a hyperparameter; is a one-dimensional convolution function; Represents a multi-head probabilistic sparse self-attention mechanism.

[0013] Furthermore, in order to output all prediction results at once, the decoder of the Informer model decodes with a parallel generation decoder mechanism; it is necessary to decode the time series with a length of When predicting a subsequence, the length of the feature to be predicted in the input sequence is The moment remains unchanged, and the part that needs to be predicted is The placeholders are filled with 0, while other features of the input sequence remain unchanged; After the masked multi-head probabilistic sparse self-attention mechanism is operated, the decoder uses the feature map output by the encoder to perform the multi-head self-attention mechanism operation; finally, the dimension of the data output is adjusted through the fully connected layer to output the required prediction results at one time.

[0014] Furthermore, the HI model and the Informer model are pre-trained using historical meteorological data and historical photovoltaic output data to form the relationship between time, meteorological information and photovoltaic output, so that the HI model and the Informer model can achieve independent forecasting based on meteorological information; The predicted photovoltaic output of the HI model and the Informer model is used as input, and the actual photovoltaic output of the historical photovoltaic output data is used as output. The LSTM model is trained and the two outputs of the HI model and the Informer model are fused using the LSTM model to correct the prediction error.

[0015] Furthermore, after training the HI model and the Informer model, if the output of the HI model and the Informer model meets the prediction accuracy requirements compared with the actual photovoltaic output, the output of the HI model is used as the And the output of the Informer model The input is fed into the LSTM model, and the training output is used to predict the photovoltaic output; After training the LSTM model, if the output of the LSTM model meets the prediction accuracy requirements compared with the actual photovoltaic output, the HI model, the Informer model and the LSTM model will be put into actual prediction. In the actual prediction part, when the model prediction value has a large deviation for a long time and the number of deviations reaches the set value, the prediction model needs to be retrained.

[0016] Furthermore, the meteorological information at the forecast time is obtained from the meteorological forecast department, and the meteorological information is updated in real time using the network information of the meteorological forecast department to predict the photovoltaic output at the specified time.

[0017] Furthermore, meteorological information is the main factor affecting PV output, and secondary factors such as PV shading rate, cloud status, and dust conditions also affect PV output, resulting in deviations in the predicted PV outputs of the HI model and the Informer model; The LSTM model is used to fuse the outputs of the HI model and the Informer model, the impact of secondary factors on photovoltaic output is modeled, and the photovoltaic output affected by secondary factors is corrected to achieve accurate photovoltaic output prediction accuracy.

[0018] Compared with the prior art, the present invention first uses the HI model (Historical Inertia Model) and the Informer model to make a preliminary prediction of the photovoltaic output, and then uses the LSTM model to fuse the outputs of the HI model and the Informer model. The present invention can, on the basis of considering the main factors of meteorological information, comprehensively consider the impact of secondary factors on photovoltaic output, correct the prediction deviation of the HI model and the Informer model, achieve more accurate photovoltaic output prediction, improve the perception ability of the power system, and improve the safe and stable operation and effective dispatching ability of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the photovoltaic output prediction method based on multi-model improved neural network fusion.

[0020] Figure 2 This is a schematic diagram of the framework of the HI model.

[0021] Figure 3 This is a schematic diagram of the framework of the Informer model.

[0022] Figure 4 Train the HI-Informer-LSTM model and predict application intent.

[0023] Figure 5 This is a schematic diagram for comparing photovoltaic output predictions. DETAILED DESCRIPTION

[0024] The photovoltaic output prediction method based on multi-model improved neural network fusion of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0025] See also Figure 1 The present invention discloses a photovoltaic output prediction method based on multi-model improved neural network fusion, comprising the following steps: S1: pre-processing the historical meteorological data and the historical photovoltaic output data to obtain standardized historical meteorological data and historical photovoltaic output data; S2: Take the historical meteorological data as input and the actual photovoltaic output of the historical photovoltaic output data as output, and train the HI model and the Informer model until the error between the output of the HI model and the Informer model and the actual photovoltaic output is less than the set threshold; S3: Take the output of the HI model and the Informer model as input, and the actual photovoltaic output of the historical photovoltaic output data as output, and train the LSTM model until the error between the output of the LSTM model and the actual photovoltaic output is less than the set threshold; S4: Obtaining meteorological information at the forecast time from the meteorological forecast department, preprocessing the meteorological information at the forecast time, and obtaining standardized meteorological information at the forecast time; S5: inputting the meteorological information at the prediction time into the HI model and the Informer model, predicting the photovoltaic output based on the HI model and the Informer model, and obtaining the output of the HI model and the Informer model; S6: Input the outputs of the HI model and the Informer model into the LSTM model, and use the LSTM model to fuse the predicted photovoltaic outputs of the HI model and the Informer model to obtain the final predicted photovoltaic output.

[0026] The preprocessing methods for historical meteorological data and historical photovoltaic output data include: screening outliers, filling missing values, correcting out-of-limit values, and finally standardizing (normalization) to prevent model deviations caused by different scales of factors.

[0027] The framework of the HI model is as follows Figure 2 As shown, the input sequence is given a length at the current moment Time Series , expressed as: ; ; In the formula, Indicated in A single feature or multiple feature inputs observed or measured at any time, where ; If the number of features in the input sequence is 1, then Contains only a single feature ,Right now , if the number of features in the input sequence exceeds 1, then Expressed as , is the number of features of the input sequence.

[0028] The goal of time series forecasting is to Step prediction corresponding length sequence , the output sequence, i.e. the predicted sequence, is expressed as: ; ; In the formula, Indicates the expected prediction A single feature or multiple feature prediction output at the moment, where ,That Time means advance The corresponding time when the step is predicted; if the number of features of the input sequence is 1, the output sequence It also contains only a single feature ,Right now , if the number of features in the input sequence exceeds 1, the output sequence is , the number of features of the output sequence The number of features of the input sequence Corresponding.

[0029] The HI model takes the input sequence To predict the length of the subsequence The sequence output is taken as the prediction result, that is, its prediction result is expressed as: ; In the formula, The value is equal to the output sequence In , Represents the input sequence middle Single feature or multiple feature input at a moment; The value is equal to the output sequence In , Represents the input sequence middle Single feature or multiple feature input at a moment; The value is equal to the output sequence In , Represents the input sequence middle Single feature or multiple feature input at a moment; The value is equal to the output sequence In , Represents the input sequence middle A single feature or multiple feature inputs at a time.

[0030] The framework of the Informer model is as follows Figure 3 As shown, the input of the encoder of the Informer model includes a length of ,have Features, i.e., size The encoder uses a multi-head probabilistic sparse self-attention mechanism. The calculation formula is: ; Where Attention(·) is the attention function; , , They are the query matrix, key matrix, and value matrix of the same size obtained after linear transformation of the feature variables of the input sequence; is the number of features of the input sequence; , , They are , , The row dimension of represents the set of real numbers; is the activation function; Express The matrix obtained after thinning by probability is is with Matrix is ​​a sparse matrix of the same size.

[0031] In order to compress the feature dimension and extract the dominant attention variables in the input, the encoder adopts self-attention distillation technology. The self-attention distillation technology re-superimposes one-dimensional convolution on the output of the previous layer and then performs a maximum pooling operation; after multiple multi-head probabilistic sparse self-attention and self-attention distillation, the encoder can finally output a reduced feature map.

[0032] Among them from Layer to The distillation process of the layer is expressed as: ; In the formula, and They are Moment Layer and Layer attention input; is the maximum pooling function; is the activation function, expressed as ,in for Input, is a hyperparameter, which is set to 1 in this embodiment; Represents a multi-head probabilistic sparse self-attention mechanism.

[0033] is a one-dimensional convolution function, is a one-dimensional convolution function, for an input sequence , among which . Set the size to Window and set the corresponding weight parameters and bias parameters In the one-dimensional convolution process, for the output sequence of the convolution Each of conduct Operation to get the output after one-dimensional convolution .

[0034] In order to output all prediction results at once, the decoder of the Informer model decodes with a parallel generation decoder mechanism; it is necessary to decode the time series with a length of When predicting a subsequence, the length of the feature to be predicted in the input sequence is The moment remains unchanged, and the part that needs to be predicted is The placeholders are filled with 0, while other features of the input sequence remain unchanged.

[0035] After the masked multi-head probabilistic sparse self-attention mechanism is operated, the decoder uses the feature map output by the encoder to perform the multi-head self-attention mechanism operation; finally, the dimension of the data output is adjusted through the fully connected layer to output the required prediction results at one time.

[0036] like Figure 4 As shown, the HI model and the Informer model are pre-trained using historical meteorological data and historical photovoltaic output data to form the relationship between time, meteorological information and photovoltaic output, so that the HI model and the Informer model can achieve independent forecasting based on meteorological information.

[0037] The predicted photovoltaic output of the HI model and the Informer model is used as input, and the actual photovoltaic output of the historical photovoltaic output data is used as output. The LSTM model is trained and the two outputs of the HI model and the Informer model are fused using the LSTM model to correct the prediction error.

[0038] After training the HI model and the Informer model, if the output of the HI model and the Informer model meets the prediction accuracy requirements compared with the actual photovoltaic output, the output of the HI model is used as the And the output of the Informer model The LSTM model is used as input, and the training output is used to predict the photovoltaic output. After the LSTM model is trained, if the output of the LSTM model meets the prediction accuracy requirements compared with the actual photovoltaic output, the HI model, the Informer model, and the LSTM model are put into actual prediction; in the actual prediction part, when the model prediction value has a large deviation for a long time and the number of deviations reaches the set value, the prediction model needs to be retrained.

[0039] The HI-Informer-LSTM model prediction method adopted in the present invention weights and fuses the prediction results of the HI model and the Informer model through the LSTM model, which can significantly reduce the prediction error, improve the overall reliability and accuracy of load forecasting, increase the stability of the power system, and reduce the dependence on specific models, thereby maintaining high performance under different circumstances.

[0040] Obtain weather information at the time of prediction from the weather forecast department, use the weather forecast department's network information to update the weather information in real time, and predict the photovoltaic output at the specified time. Meteorological information is the main factor affecting photovoltaic output. Secondary factors such as photovoltaic shading rate, cloud status, and dust conditions will also affect photovoltaic output, resulting in deviations in the predicted photovoltaic output of the HI model and the Informer model. Using the LSTM model to fuse the outputs of the HI model and the Informer model can model the impact of secondary factors on photovoltaic output and correct the photovoltaic output affected by secondary factors to achieve accurate photovoltaic output prediction accuracy.

[0041] The photovoltaic output prediction method based on multi-model improved neural network fusion of the present invention is simulated and verified, and the simulation data set comes from the actual measurement values ​​of a photovoltaic power station in Taiyuan, Shanxi and the meteorological history of Taiyuan, Shanxi from the Central Meteorological Observatory. Figure 5 is the amplitude diagram of photovoltaic output prediction, Table 1 is the statistical table of photovoltaic output prediction, and the evaluation index adopts the root mean square error RMSE (Root-Mean-Square Error), as shown in the following formula: ; For a predicted output sequence, is the total length of the predicted output sequence, and Corresponding to The predicted value and actual measured value of the forecast output sequence at each moment.

[0042] Table 1 Photovoltaic output forecast comparison table

[0043] As can be seen from Table 1, the HI model is used to predict the photovoltaic output, and the root mean square error RMSE of the photovoltaic output prediction is 1.272. The Informer model is used to predict the photovoltaic output, and the root mean square error RMSE of the photovoltaic output prediction is 0.982. The photovoltaic output prediction method of the present invention is based on a multi-model improved neural network fusion, that is, the HI-Informe-LSTM prediction method is used to predict the photovoltaic output, and the root mean square error RMSE of the photovoltaic output prediction is 0.871. The present invention can significantly reduce the prediction error and improve the reliability and accuracy of the prediction.

[0044] In summary, the present invention first uses the HI model and the Informer model to make a preliminary prediction of the photovoltaic output, and then uses the LSTM model to fuse the outputs of the HI model and the Informer model. On the basis of considering the main factors of meteorological information, the present invention can comprehensively consider the impact of secondary factors on photovoltaic output, correct the prediction deviation of the HI model and the Informer model, achieve more accurate photovoltaic output prediction, improve the perception ability of the power system, and improve the safe and stable operation and effective dispatching ability of the power system.

[0045] The above description is a detailed description of the preferred feasible embodiments of the present invention, but the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modified changes completed under the technical spirit disclosed by the present invention should fall within the patent scope covered by the present invention.

Claims

1. A photovoltaic output prediction method based on multi-model improved neural network fusion, characterized in that: The following steps are involved: S1: pre-processing the historical meteorological data and the historical photovoltaic output data to obtain standardized historical meteorological data and historical photovoltaic output data; S2: Take the historical meteorological data as input and the actual photovoltaic output of the historical photovoltaic output data as output, and train the HI model and the Informer model until the error between the output of the HI model and the Informer model and the actual photovoltaic output is less than the set threshold; S3: Take the output of the HI model and the Informer model as input, and the actual photovoltaic output of the historical photovoltaic output data as output, and train the LSTM model until the error between the output of the LSTM model and the actual photovoltaic output is less than the set threshold; S4: Obtaining meteorological information at the forecast time from the meteorological forecast department, preprocessing the meteorological information at the forecast time, and obtaining standardized meteorological information at the forecast time; S5: inputting the meteorological information at the prediction time into the HI model and the Informer model, predicting the photovoltaic output based on the HI model and the Informer model, and obtaining the output of the HI model and the Informer model; S6: Input the outputs of the HI model and the Informer model into the LSTM model, and use the LSTM model to fuse the predicted photovoltaic outputs of the HI model and the Informer model to obtain the final predicted photovoltaic output.

2. The photovoltaic output prediction method based on multi-model improved neural network fusion according to claim 1 is characterized in that: The methods for preprocessing historical meteorological data and historical photovoltaic output data include: screening outliers, filling missing values, and correcting over-limit values.

3. The photovoltaic output prediction method based on multi-model improved neural network fusion according to claim 1 is characterized in that: The photovoltaic output is predicted based on the HI model, specifically: The input sequence of the HI model is: given a length at the current moment Time Series , expressed as: ； ； In the formula, Indicated in A single feature or multiple feature inputs observed or measured at any time, where ; If the number of features in the input sequence is 1, then Contains only a single feature ,Right now , if the number of features in the input sequence exceeds 1, then Expressed as , is the number of features of the input sequence; The goal of time series forecasting is to Step prediction corresponding length sequence , the output sequence, i.e. the predicted sequence, is expressed as: ； ； In the formula, Indicates the expected prediction A single feature or multiple feature prediction output at the moment, where ,That Time means advance The corresponding time during the step prediction; If the number of features in the input sequence is 1, the output sequence is It also contains only a single feature ,Right now , if the number of features in the input sequence exceeds 1, the output sequence is , the number of features of the output sequence The number of features of the input sequence corresponding; The HI model takes the input sequence To predict the length of the subsequence The sequence output is taken as the prediction result, that is, its prediction result is expressed as: ； In the formula, The value is equal to the output sequence In , Represents the input sequence middle Single feature or multiple feature input at a moment; The value is equal to the output sequence In , Represents the input sequence middle Single feature or multiple feature input at a moment; The value is equal to the output sequence In , Represents the input sequence middle Single feature or multiple feature input at a moment; The value is equal to the output sequence In , Represents the input sequence middle A single feature or multiple feature inputs at a time.

4. The photovoltaic output prediction method based on multi-model improved neural network fusion according to claim 1 is characterized in that: The photovoltaic output is predicted based on the Informer model, specifically: The input to the encoder of the Informer model consists of a length of ,have Features, i.e., size The encoder uses a multi-head probabilistic sparse self-attention mechanism. The calculation formula is: ； Where Attention(·) is the attention function; , , They are the query matrix, key matrix, and value matrix of the same size obtained after linear transformation of the feature variables of the input sequence; is the number of features of the input sequence; , , They are , , The row dimension of represents the set of real numbers; is the activation function; Express The matrix obtained after thinning by probability is is with Matrix is ​​a sparse matrix of the same size.

5. The photovoltaic output prediction method based on multi-model improved neural network fusion according to claim 4 is characterized in that: In order to compress the feature dimension and extract the dominant attention variables in the input, the encoder uses self-attention distillation technology. The self-attention distillation technology re-superimposes one-dimensional convolution on the output of the previous layer and then performs a maximum pooling operation. After multiple multi-head probabilistic sparse self-attention and self-attention distillation, the encoder finally outputs a reduced feature map. from Layer to The distillation process of the layer is expressed as: ； In the formula, and They are Moment Layer and Layer attention input; is the maximum pooling function; is the activation function, , for Input, is a hyperparameter; is a one-dimensional convolution function; Represents a multi-head probabilistic sparse self-attention mechanism.

6. The photovoltaic output prediction method based on multi-model improved neural network fusion according to claim 5 is characterized in that: In order to output all prediction results at once, the decoder of the Informer model decodes with a parallel generation decoder mechanism; it is necessary to decode the time series with a length of When predicting a subsequence, the length of the feature to be predicted in the input sequence is The moment remains unchanged, and the part that needs to be predicted is The placeholders are filled with 0, while other features of the input sequence remain unchanged; After the masked multi-head probabilistic sparse self-attention mechanism is operated, the decoder uses the feature map output by the encoder to perform the multi-head self-attention mechanism operation; finally, the dimension of the data output is adjusted through the fully connected layer to output the required prediction results at one time.

7. The photovoltaic output prediction method based on multi-model improved neural network fusion according to claim 1 is characterized in that: Using historical meteorological data and historical photovoltaic output data, the HI model and the Informer model are pre-trained to form the relationship between time, meteorological information and photovoltaic output, so that the HI model and the Informer model can achieve independent forecasting based on meteorological information; The predicted photovoltaic output of the HI model and the Informer model is used as input, and the actual photovoltaic output of the historical photovoltaic output data is used as output. The LSTM model is trained and the two outputs of the HI model and the Informer model are fused using the LSTM model to correct the prediction error.

8. The photovoltaic output prediction method based on multi-model improved neural network fusion according to claim 1 is characterized in that: After training the HI model and the Informer model, if the output of the HI model and the Informer model meets the prediction accuracy requirements compared with the actual photovoltaic output, the output of the HI model is used as the And the output of the Informer model The input is fed into the LSTM model, and the training output is used to predict the photovoltaic output; After training the LSTM model, if the output of the LSTM model meets the prediction accuracy requirements compared with the actual photovoltaic output, the HI model, the Informer model and the LSTM model will be put into actual prediction. In the actual prediction part, when the model prediction value has a large deviation for a long time and the number of deviations reaches the set value, the prediction model needs to be retrained.

9. The photovoltaic output prediction method based on multi-model improved neural network fusion according to claim 1 is characterized in that: Obtain weather information at the forecast time from the weather forecast department, use the weather forecast department's network information to update the weather information in real time, and predict the photovoltaic output at the specified time.

10. The photovoltaic output prediction method based on multi-model improved neural network fusion according to claim 1 is characterized in that: Meteorological information is the main factor affecting PV output. Secondary factors such as PV shielding rate, cloud status, and dust conditions also affect PV output, leading to deviations in the predicted PV outputs of the HI model and the Informer model. The LSTM model is used to fuse the outputs of the HI model and the Informer model, the impact of secondary factors on photovoltaic output is modeled, and the photovoltaic output affected by secondary factors is corrected to achieve accurate photovoltaic output prediction accuracy.

Images