Deep learning photovoltaic output prediction method and system based on meteorological feature mining

Abstract

The invention provides a deep learning photovoltaic output prediction method and system based on meteorological feature mining, and belongs to the technical field of photovoltaic power prediction. Comprising the following steps: acquiring photovoltaic power historical data of a photovoltaic station and meteorological information of a geographic position of the photovoltaic station; screening out N meteorological factors for model training based on a mutual information rule; a PDF deep learning prediction model is constructed, model training is carried out, a plurality of periodic feature modules are used for carrying out feature extraction on photovoltaic power historical data and N kinds of meteorological historical data, obtained N + 1 attention results are aggregated and input into a predictor based on a linear architecture, and model parameters are updated through back propagation; the periodic feature module is composed of a multi-period decoupling module and a double-variable modeling module; and performing photovoltaic power prediction under different meteorological conditions by using the trained PDF deep learning prediction model.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power prediction, and in particular to a deep learning photovoltaic output prediction method and system based on meteorological feature mining. Background Art

[0002] Compared to the controllable continuous output characteristics of traditional thermal and hydropower generation, photovoltaic power generation is subject to complex environmental variables such as irradiation intensity and meteorological conditions, and exhibits significant intermittent characteristics. This instability poses a serious challenge to the reliable operation of the power system: the power fluctuations generated by large-scale photovoltaic grid connection may cause grid frequency deviations, increasing the difficulty for dispatching departments to coordinate the ratio of traditional power sources and renewable energy. In this context, the importance of high-precision photovoltaic power prediction technology has become increasingly prominent. Its application value is reflected in reducing the risk of transient shocks to the power grid, optimizing the coordinated dispatch of renewable and traditional energy sources, and improving the efficiency of photovoltaic absorption, providing key technical support for the construction of new power systems.

[0003] The existing data-driven intelligent algorithm systems are mainly based on artificial neural networks (ANNS). Some experts have systematically compared the prediction performance of regression algorithms, Markov chains and various ANN architectures, and confirmed that ANN has stronger representational capabilities in handling complex nonlinear prediction tasks. However, due to differences in model topology and training mechanisms, the prediction performance of different neural networks shows obvious differentiation. Some experts have innovatively constructed a fusion model based on adversarial networks and long short-term memory networks (LSTM), achieving a breakthrough in short-term prediction accuracy in low-sample data scenarios. Its prediction error is significantly lower than that of traditional BP networks, recurrent neural networks (RNN) and SVM models. Some experts have proposed the use of LSTM architecture, which effectively improves the ability to analyze time series dynamic features and shows better prediction stability than BP and RNN networks. Existing research data shows that the LSTM network can deeply extract the spatiotemporal correlation characteristics of power generation sequences through a gating mechanism. Its unique memory unit design effectively avoids the gradient dissipation defect of RNN. The memory characteristics of LSTM make it more applicable in engineering power forecasting under complex meteorological conditions. Deep neural networks based on LSTM and its variants have become the focus of current research on intelligent prediction models.

[0004] However, recurrent neural networks, represented by LSTM and its variants, have defects in model training and the model itself. Recurrent neural networks currently face the problem of vanishing gradients during model training and excessive training time when processing long time series. The model itself uses a sliding window format and needs to store prediction results as input for the next time period. At the same time, the model has a large number of parameters and requires a lot of computing resources. There may be a risk of oscillation when processing non-stationary series. Summary of the Invention

[0005] In light of this, the present invention provides a deep learning photovoltaic output prediction method and system based on meteorological feature mining. This method considers the coupling characteristics between meteorological factors and photovoltaic output, screens meteorological factors for interpretability based on the mutual information law, and selects the most appropriate meteorological factors, taking into account the specific impacts of different meteorological factors on photovoltaic output. This method also improves on previous LSTM-based photovoltaic power prediction methods by using a periodic decoupling framework, PDF, for photovoltaic output prediction. This method uses fast Fourier decomposition to better reflect the characteristics of photovoltaic output, effectively avoiding the oscillation risks associated with processing non-stationary sequence processes, and improving prediction accuracy and model training speed.

[0006] The technical solution adopted by the embodiment of the present invention to solve the technical problem is:

[0007] A deep learning photovoltaic output prediction method based on meteorological feature mining is characterized by comprising:

[0008] Step S1, obtaining historical photovoltaic power data of the photovoltaic station and meteorological information of the geographical location of the photovoltaic station; meteorological factors include total irradiance, normal direct irradiance, horizontal plane diffuse irradiance, temperature, air pressure, and relative humidity;

[0009] Step S2, based on the mutual information law, N meteorological factors are selected for model training;

[0010] Step S3: construct a PDF deep learning prediction model and perform model training, wherein N+1 periodic feature modules are used to extract features based on photovoltaic power historical data and N types of meteorological historical data, respectively. The obtained N+1 attention results are aggregated and input into a predictor based on a linear architecture, and the model parameters are updated through back propagation. The periodic feature module is composed of a multi-period decoupling module and a bivariate modeling module. The multi-period decoupling module obtains the long-term sequence and short-term sequence of the input data through learning. The bivariate modeling module predicts the long-term sequence and the short-term sequence according to different mapping methods to obtain the attention result.

[0011] Step S4: Use the trained PDF deep learning prediction model to predict photovoltaic power under different meteorological conditions.

[0012] Preferably, the step S2 of selecting N meteorological factors for model training based on the mutual information law includes:

[0013] Calculate the mutual information I(X i ; Y):

[0014]

[0015] Where, X i is the time series of the i-th type of meteorological factors, Y is the time series of photovoltaic power, and p(x,y) represents X i The joint probability distribution function of X and Y, p(x) and p(y) are respectively i and the marginal probability distribution function of Y;

[0016] According to the order of mutual information values ​​from high to low, the meteorological factors corresponding to the first N values ​​are taken to participate in model training.

[0017] Preferably, the model training process in step S3 includes:

[0018] Step S31, establishing a data set and dividing it into a training set, a validation set, and a test set in proportion, wherein the data set includes historical photovoltaic power data and N types of historical meteorological data corresponding to N meteorological factors;

[0019] Step S32: Input the training set data into the PDF deep learning prediction model according to the historical sequence length for training:

[0020] The multi-period decoupling module decomposes the time series into multiple frequency domain models according to the frequency domain through the fast Fourier transform FFT. According to the decoupling results of FFT, K frequency features are selected in the order of frequency amplitude from large to small. For the selected frequency features, each frequency band is patched and divided. Perform vertical cutting and horizontal cutting to obtain long-term sequences and short-term series Where i represents the time series of the i-th frequency segment, i∈[1,K], j represents the j-th column of the long-term partition, r represents the r-th row of the short-term partition, {p i} represents the sequence set after patching division, where PDF stacks and aggregates the same parts of each patch during vertical cutting and stitches them into a new sequence;

[0021] The bivariate modeling module is used to model the long-term series and short-term series Perform feature extraction to obtain features and features And the results and Aggregate into the attention result Among them, the linear layer is used to After temporal encoding, the self-attention mechanism is used to send the constant sequence to the Transformer encoder. The features are then aggregated according to the original temporal order, and finally a linear layer is used as a predictor to generate the final predicted sequence. For short sequence variation feature extraction, PDF uses convolutional layers and nonlinear activation functions to capture the local features of short sequences. Each patch is then spliced ​​together in temporal order and mapped back to the original sequence using a linear layer.

[0022] Inputting the aggregated N+1 attention results into a predictor based on a linear fully connected layer to obtain photovoltaic power prediction data, wherein the photovoltaic power prediction data is the photovoltaic power prediction data adjusted based on the attention results of the N meteorological data;

[0023] Update the model parameters of the PDF deep learning prediction model through back propagation to obtain the model with minimum loss;

[0024] Update the learning rate, train the model multiple times, use the early stopping mechanism, and obtain the optimal model under the current parameters as the trained PDF deep learning prediction model;

[0025] Use the validation set to validate the model, and use the test set to test the model.

[0026] Preferably, when N=3, the mutual information is arranged from large to small, and the meteorological factors corresponding to the first three values ​​include total irradiance, normal direct irradiance, and horizontal surface scattered irradiance.

[0027] A deep learning photovoltaic output prediction system based on meteorological feature mining is used for the aforementioned method and adopts a PDF deep learning prediction model. In the PDF deep learning prediction model, N+1 periodic feature modules are used to extract features based on photovoltaic power historical data and N types of meteorological historical data, respectively. The obtained N+1 attention results are aggregated and input into a predictor based on a linear architecture, and the model parameters are updated by back propagation; the periodic feature module is composed of a multi-period decoupling module and a bivariate modeling module. The multi-period decoupling module obtains the long-term sequence and short-term sequence of the input data through learning, and the bivariate modeling module predicts the long-term sequence and the short-term sequence according to different mapping methods to obtain the attention result; the prediction result output by the PDF deep learning prediction model is the photovoltaic power prediction data obtained by adjusting the N types of meteorological factors.

[0028] As can be seen from the above technical solution, the deep learning photovoltaic output prediction method and system based on meteorological feature mining provided by the embodiment of the present invention first obtains the photovoltaic power historical data of the photovoltaic station and the meteorological information of the geographical location of the photovoltaic station; based on the mutual information law, N meteorological factors are screened for model training; a PDF deep learning prediction model is constructed and the model training is performed, wherein N+1 periodic feature modules are used to extract features based on the photovoltaic power historical data and N types of meteorological historical data, respectively, and the obtained N+1 attention results are aggregated and input into a predictor based on a linear architecture, and the model parameters are updated through back propagation; the periodic feature module is composed of a multi-period decoupling module and a two-variable modeling module; and the trained PDF deep learning prediction model is used to predict photovoltaic power under different meteorological conditions. The present invention predicts photovoltaic output based on meteorological factor screening and PDF deep learning model, effectively avoiding the risk of oscillation in the process of processing non-stationary sequence, and improving prediction accuracy and model training speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a flow chart of the deep learning photovoltaic output prediction method based on meteorological feature mining of the present invention.

[0030] Figure 2 for Figure 1 Schematic diagram of the PDF overall framework of the method shown.

[0031] Figure 3 Schematic diagram of the multi-cycle decoupling module in the PDF overall framework.

[0032] Figure 4 Schematic diagram of the bivariate modeling module in the PDF overall framework.

[0033] Figure 5 This is the PDF diagram of photovoltaic power prediction results. DETAILED DESCRIPTION

[0034] The technical solutions and technical effects of the present invention are further described in detail below with reference to the accompanying drawings of the present invention.

[0035] The present invention proposes a deep learning photovoltaic output prediction method based on meteorological feature mining. First, a computer uses the mutual information law to calculate the actual output of a photovoltaic power station and the mutual information results of the meteorological information corresponding to the photovoltaic power station. The meteorological factors are screened according to the size of the mutual information. Then, a periodicity decoupling framework (PDF) deep learning model is built on the computer. Finally, the computer inputs the screened meteorological factors into the deep learning model based on the power information of the actual photovoltaic power station and the meteorological information of the geographical location of the photovoltaic power station. After computer model training, the prediction results of the photovoltaic data are finally obtained. At the same time as the theoretical analysis, a deep learning model is established in pycharm based on the Python language using a computer to facilitate application in engineering practice.

[0036] refer to Figure 1 As shown, the deep learning photovoltaic output prediction method based on meteorological feature mining provided by the present invention includes:

[0037] Step S1, obtaining historical photovoltaic power data of the photovoltaic station and meteorological information of the geographical location of the photovoltaic station; meteorological factors include total irradiance, normal direct irradiance, horizontal plane diffuse irradiance, temperature, air pressure, and relative humidity;

[0038] Step S2, based on the mutual information law, N meteorological factors are selected for model training;

[0039] Step S3: construct a PDF deep learning prediction model and perform model training, wherein N+1 periodic feature modules are used to extract features based on historical photovoltaic power data and N types of meteorological historical data, respectively. The obtained N+1 attention results are aggregated and input into a predictor based on a linear architecture, and the model parameters are updated through backpropagation. The periodic feature module consists of a multi-period decoupling module and a bivariate modeling module. The multi-period decoupling module obtains the long-term and short-term sequences of the input data through learning, and the bivariate modeling module predicts the long-term and short-term sequences according to different mapping methods to obtain attention results.

[0040] Step S4: Use the trained PDF deep learning prediction model to predict photovoltaic power under different meteorological conditions.

[0041] Step S1 provides historical data on a variety of meteorological factors that are highly correlated with PV output forecasts. However, the impact of different meteorological factors on PV output varies in different regions, at different times, and under different weather conditions. Using all types of meteorological data for forecasting results in excessive data volume and low computational efficiency. Therefore, it is very necessary to adjust the meteorological data selection strategy in real time according to actual conditions, screening out meteorological factors with higher relevance in the current situation to adjust the PV output forecast results. The mutual information (MI) of two random variables is a measure of the interdependence between variables. In step S2, N meteorological factors are selected based on the mutual information law for model training:

[0042] Calculate the mutual information I(X i ; Y):

[0043]

[0044] Where, X i is the time series of the i-th type of meteorological factors, Y is the time series of photovoltaic power, both are discrete variables, p(x,y) represents X i The joint probability distribution function of X and Y, p(x) and p(y) are respectively i and the marginal probability distribution function of Y;

[0045] According to the order of mutual information values ​​from high to low, the meteorological factors corresponding to the first N values ​​are taken to participate in model training.

[0046] Time series prediction often leads to poor accuracy of prediction results because the time series is too long, and as the time series becomes longer, the difficulty of prediction increases sharply and the complexity of the model increases day by day. Existing time series prediction methods, especially those based on deep learning, often only input one-dimensional time series data. The one-dimensional data is too simple and limits the model's ability to capture the inherent periodicity and long-term dependence of the data. The present invention establishes a new time series prediction framework - the Periodicity Decoupling Framework (PDF). PDF can effectively process long-term series data and capture the periodicity and long-term dependence in the data (the period here not only refers to the periodicity at each frequency after frequency domain decomposition, but also refers to the periodicity divided into long-term and short-term for each frequency band). The main idea of ​​PDF is to transform one-dimensional data into a two-dimensional representation to better capture the temporal variation pattern of the data. PDF transforms the time feature from one dimension into two-dimensional information of long sequence feature extraction and short sequence feature extraction. The overall framework is as follows Figure 2As shown. This paper emphasizes that when using multiple variables to predict a single variable, the model adopts the concept of channel independence, extracting two-dimensional periodic features from multiple variables and the target variable based on a periodic decoupling framework, and aggregating all extracted multivariate features. Currently, linear-based predictors are widely used in the field of time series prediction. This paper inputs the aggregated features into a predictor based on a linear fully connected layer. The predictor's output is only one-dimensional, enabling the prediction of a single photovoltaic power.

[0047] The model training process in step S3 includes:

[0048] Step S31: Establish a data set and divide it into a training set, a validation set, and a test set in proportion. The data set includes photovoltaic power data and N types of meteorological data corresponding to N meteorological factors.

[0049] Step S32: Input the training set data into the PDF deep learning prediction model for training according to the historical sequence length:

[0050] Multi-cycle decoupling module, reference Figure 3 As shown in the figure, the time series is decomposed into multiple frequency domain models according to the frequency domain by fast Fourier transform FFT. According to the decoupling result of FFT, K frequency features are selected in the order of frequency amplitude from large to small. For the selected frequency features, each frequency band is patched and divided. Perform vertical cutting and horizontal cutting to obtain long-term sequences and short-term series Where i represents the time series of the i-th frequency segment, i∈[1,K], j represents the j-th column of the long-term partition, r represents the r-th row of the short-term partition, {p i} represents the sequence set after patching division, where PDF stacks and aggregates the same parts of each patch during vertical cutting and stitches them into a new sequence;

[0051] Bivariate Modeling Module, reference Figure 4 As shown, the long-term series and short-term series Perform feature extraction to obtain features and features And the results and Aggregation into attention results Among them, the processing of long-term sequences is linear projection, several Transformer encoder layers, aggregation layers and linear projection, and the linear layer is used to process long-term sequences. Perform time encoding, and then use the self-attention mechanism to send the normal sequence to the Transformer encoder, then aggregate the features according to the original time order, and finally use the linear layer as a predictor to generate the final predicted sequence. Formula (2) is the feature extraction of long sequences; for the feature extraction of short sequence changes, a convolution block consisting of a convolution layer and a nonlinear activation function is used. In the figure, SELU represents the scaled exponential linear unit, Concat represents the connection function, and Linear is the linear projection; since the time range is a small local range, PDF uses convolution layers and nonlinear activation functions to capture the local features of short sequences, and then splices each patch in time order and uses a linear layer to map back to the original sequence. Formula (3) is the feature extraction of short sequences;

[0052]

[0053] In the formula, BatchNorm() is the batch normalization function, MSA() is the multi-head attention mechanism that uses multiple independent self-attention heads to enhance the representation ability, and the self-attention mechanism is used to set the Q, K, and V matrices to be MLP() is a multi-layer linear mapping, Flatten() is an aggregation function, Conv1d() is a one-dimensional convolutional neural network, SELU() is a scaled exponential linear unit function, and Concat() is a connection function;

[0054] The aggregated N+1 attention results are input into the predictor based on the linear fully connected layer to obtain the photovoltaic power prediction data. The photovoltaic power prediction data is the photovoltaic power prediction data adjusted based on the attention results of N meteorological data.

[0055] Update the model parameters of the PDF deep learning prediction model through back propagation to obtain the model with minimum loss;

[0056] Update the learning rate, train the model multiple times, use the early stopping mechanism, and obtain the optimal model under the current parameters as the trained PDF deep learning prediction model;

[0057] Validate the model using the validation set;

[0058] The final trained model is output, and the trained model is applied to the test set to calculate the model evaluation index to achieve photovoltaic power prediction under different meteorological conditions.

[0059] As can be seen, the PDF deep learning prediction model extracts a multivariate correlation score matrix based on different deep learning model architectures. Based on the prediction sequence length and model hyperparameters (which can be adjusted through hyperparameter tuning methods), the obtained attention results are aggregated and input into a linear architecture-based predictor. After backpropagation updates the parameters, a model with a smaller loss value is obtained under these parameters. After updating the learning rate, the model is trained multiple times using an early stopping mechanism to obtain the optimal model under the current parameters.

[0060] Furthermore, the present invention provides a deep learning photovoltaic output forecasting system based on meteorological feature mining, which utilizes a PDF deep learning forecasting model. In this PDF deep learning forecasting model, N+1 periodic feature modules are used to extract features based on historical photovoltaic power data and N types of historical meteorological data, respectively. The resulting N+1 attention results are aggregated and input into a predictor based on a linear architecture, and the model parameters are updated through backpropagation. The periodic feature module is composed of a multi-period decoupling module and a bivariate modeling module. The multi-period decoupling module learns to derive long-term and short-term sequences of the input data, and the bivariate modeling module predicts the long-term and short-term sequences using different mapping methods to obtain attention results. The PDF deep learning forecasting model outputs photovoltaic power prediction data adjusted using N meteorological factors. Photovoltaic power forecasting is implemented using the aforementioned method.

[0061] As an optional implementation method, the N types of historical data corresponding to the meteorological factors after mutual information screening and the photovoltaic output historical data can be integrated into a whole (such as a matrix) and input into the PDF model for feature extraction and photovoltaic output prediction. However, compared with the aforementioned method of extracting single data features, extracting matrix features will cause data missing during the process of extracting matrix features, resulting in a large difference in the output prediction accuracy. More accurate prediction data can be obtained by extracting features from single data separately through the PDF model and then aggregating the attention results.

[0062] A specific embodiment is given below:

[0063] This paper selects the actual power data of photovoltaic power stations in a certain region of China with a total installed capacity of 50MW, and obtains the meteorological information of the photovoltaic power station. It mainly considers six aspects: total irradiance, normal direct irradiance, horizontal plane scattered irradiance, temperature, air pressure and relative humidity. The time sampling interval of the data is 15 minutes, that is, 96 time sampling points per day, which meets the short-term photovoltaic output needs. At the same time, the time span is from January 1, 2019, to December 26, 2020, and the long-term characteristics of photovoltaic output can also be obtained. The specific format is shown in Table 1.

[0064] Table 1 Photovoltaic power station data format

[0065]

[0066] The mutual information between the actual power of the photovoltaic power station and the other six meteorological factors is calculated according to formula (1). The calculation results are shown in Table 2.

[0067] Table 2 Meteorological characteristics screening results

[0068]

[0069]

[0070] It can be seen from Table 2 that, under the current circumstances, the three meteorological factors of total irradiance, normal direct irradiance, and horizontal scattered irradiance have a much greater impact on photovoltaic output than the other three meteorological factors. Therefore, it is necessary to consider whether meteorological factors with low mutual information values ​​will seriously affect the accuracy of the predicted data when participating in model training. It is necessary to investigate the optimal number of meteorological factors to be used through specific experiments. The present invention innovatively considers the impact of the value of N on the accuracy of the prediction results, and conducts in-depth research on specific actual data, especially considering the main factors affecting photovoltaic output under abnormal meteorological conditions. Taking the data in January 2019 as an example, the photovoltaic output on January 4 and 9 showed abnormal fluctuations. Among them, the total irradiance, normal direct irradiance, and horizontal scattered irradiance on the 4th decreased, resulting in a decrease in photovoltaic power; the total irradiance and horizontal scattered irradiance on the 9th decreased, resulting in a decrease in photovoltaic power. The change in scattered irradiance was not obvious, while the change in normal direct irradiance was drastic, resulting in a sharp drop in photovoltaic power. Compared with the 1st, the two meteorological factors of normal direct irradiance and horizontal scattered irradiance changed significantly on the 10th, while the photovoltaic output remained basically the same, verifying that photovoltaic output is related to both normal direct irradiance and horizontal scattered irradiance. When considering N=2, it is easy to ignore the large amount of information contained in the meteorological factors, resulting in a decrease in the model prediction effect. The specific experimental results are shown in Table 5. Therefore, the present invention considers the case of N=3 when selecting meteorological factors, fully extracts meteorological information, and only considers the three meteorological factors of total irradiance, normal direct irradiance, and horizontal scattered irradiance. When using multiple variables to predict photovoltaic power later, only the above three meteorological factors with large mutual information are introduced into the model to ensure the accuracy of the model.

[0071] Selecting N as 3 is an optional implementation and does not mean that the present invention has confirmed the selection of the three meteorological factors of total irradiance, normal direct irradiance, and horizontal plane scattered irradiance. When this method is applied to different regions, different seasons, and different weather conditions, it is also possible to appropriately change the value of N, and the factors with mutual information values ​​in the first N positions can also be meteorological factors other than the aforementioned three. The selection of the N value is based on the evaluation indicators of the test set. Through experiments, different N values ​​can be defined, and then the model training and testing can be carried out. The values ​​of each evaluation indicator (MSE\MAE) are calculated and compared, and the N value that performs best in all indicators is selected as the selected value.

[0072] The sequence lengths and model hyperparameters used by the prediction model are shown in Table 3.

[0073] Table 3 Sequence length and model hyperparameters

[0074]

[0075] During actual training, the computer set the historical sequence length to 96 time steps and the predicted future sequence length to 96 time steps. The model was trained using a random dropout mechanism, early stopping mechanism, and mini-batch gradient descent. The mean absolute error (MAE) was selected as the loss function, and the mean squared error (MSE) and mean squared error (MSE) were selected as the prediction indicators. Other Transformer-based models were selected as control groups, and the multivariate input before the unscreened variables was also used as the control group for the experiment. The prediction indicators of the PDF-based deep learning photovoltaic output prediction method proposed in this invention and other control groups were calculated. The calculation results are shown in Table 3.

[0076] Table 4 Test set evaluation indicators

[0077]

[0078]

[0079] Table 5 N = 2 test set MSE analysis

[0080]

[0081] It can be concluded from Table 4 that, from the perspective of data preprocessing methods, the prediction results after filtering meteorological factors are all better than the prediction structure without filtering meteorological factors, which proves that the data preprocessing method based on meteorological feature mining is better than the deep learning model that considers all meteorological factors. From the perspective of the lightweight model itself, the lightweight model PDF has the best prediction effect because it upgrades the time data dimension and uses FFT for data processing, which is consistent with the characteristics of the power system.

[0082] In order to highlight the advantages of the PDF model in photovoltaic output prediction, Figure 5 The graph shows the forecast results of photovoltaic output under different meteorological conditions. The forecast sequence fluctuates more frequently around 0.

[0083] For the above Figure 5 The PDF photovoltaic power prediction results are analyzed. The first row of the figure shows that when the historical sequence is cloudy or the photovoltaic power fluctuates significantly due to cloud cover, the predicted sequence is sunny or cloudy. The second and third rows of the figure show that both the historical and predicted sequences are sunny, but the difference is due to the different initial prediction time points caused by the sliding window. The fourth row of the figure shows that when the historical sequence is rainy, sunny, and cloudy, the predicted sequences are sudden cloudy, rainy, and sunny. For the above classification, when using multivariate prediction to predict multivariate, the trend characteristics of the sequence can be well learned. This is reflected in the excellent prediction effect of the model when predicting sunny days based on sunny days, with a difference of about 1%. However, its prediction effect decreases significantly when the sequence fluctuates significantly. This is reflected in the model using abnormal weather as the historical sequence or abnormal weather as the predicted sequence. In both cases, the model's performance deteriorates. In particular, when the amplitude of the change in photovoltaic power in the historical sequence is too large compared to the previous day, the predicted result for the next day deviates significantly from the actual value. In summary, using the lightweight deep learning model of the present invention, in the multi-step prediction multi-step state, the data of the next 96 time points can be well predicted for most of the year.

[0084] The present invention's deep learning photovoltaic output prediction method based on meteorological feature mining uses historical data for model training each time a prediction is executed. A mutual information method is used for pre-screening, and meteorological factors with a high correlation with the current photovoltaic output variable are screened out for use in subsequent model training. This effectively improves prediction efficiency and reduces the amount of computation. A periodic decoupling framework PDF deep learning model is then built on a computer. Finally, the computer inputs the screened meteorological factors into the deep learning model based on the power information of the actual photovoltaic power station and the meteorological information of the photovoltaic power station's geographical location. After computer model training, the predicted results of the photovoltaic data are finally obtained. While conducting theoretical analysis, a deep learning model is established in PyCharm using Python language, making it easy to apply to engineering practice. This effectively avoids the risk of oscillation in processing non-stationary sequence processes, improving prediction accuracy and model training speed.

[0085] The above disclosure is only a preferred embodiment of the present invention, and it is certainly not intended to limit the scope of the present invention. A person skilled in the art can understand that all or part of the processes of the above embodiment and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A deep learning photovoltaic output prediction method based on meteorological feature mining, characterized by: include: Step S1, obtaining historical photovoltaic power data of the photovoltaic station and meteorological information of the geographical location of the photovoltaic station; meteorological factors include total irradiance, normal direct irradiance, horizontal plane diffuse irradiance, temperature, air pressure, and relative humidity; Step S2, based on the mutual information law, N meteorological factors are selected for model training; Step S3: construct a PDF deep learning prediction model and perform model training, wherein N+1 periodic feature modules are used to extract features based on photovoltaic power historical data and N types of meteorological historical data, respectively. The obtained N+1 attention results are aggregated and input into a predictor based on a linear architecture, and the model parameters are updated through back propagation. The periodic feature module is composed of a multi-period decoupling module and a bivariate modeling module. The multi-period decoupling module obtains the long-term sequence and short-term sequence of the input data through learning. The bivariate modeling module predicts the long-term sequence and the short-term sequence according to different mapping methods to obtain the attention result. Step S4: Use the trained PDF deep learning prediction model to predict photovoltaic power under different meteorological conditions.

2. The deep learning photovoltaic output prediction method based on meteorological feature mining according to claim 1 is characterized in that: The step S2 selects N meteorological factors for model training based on the mutual information law, including: calculating the mutual information I(X) between the i-th meteorological factor and the photovoltaic power i ; Y): Where, X i is the time series of the i-th type of meteorological factors, Y is the time series of photovoltaic power, and p(x,y) represents X i The joint probability distribution function of X and Y, p(x) and p(y) are respectively i and the marginal probability distribution function of Y; According to the order of mutual information values ​​from high to low, the meteorological factors corresponding to the first N values ​​are taken to participate in model training.

3. The deep learning photovoltaic output prediction method based on meteorological feature mining according to claim 2 is characterized in that: The model training process in step S3 includes: Step S31, establishing a data set and dividing it into a training set, a validation set, and a test set in proportion, wherein the data set includes historical photovoltaic power data and N types of historical meteorological data corresponding to N meteorological factors; Step S32: Input the training set data into the PDF deep learning prediction model according to the historical sequence length for training: The multi-period decoupling module decomposes the time series into multiple frequency domain models according to the frequency domain through the fast Fourier transform FFT. According to the decoupling results of FFT, K frequency features are selected in the order of frequency amplitude from large to small. For the selected frequency features, each frequency band is patched and divided. Perform vertical cutting and horizontal cutting to obtain long-term sequences and short-term series Where i represents the time series of the i-th frequency segment, i∈[1,K], j represents the j-th column of the long-term partition, r represents the r-th row of the short-term partition, {p i } represents the sequence set after patching division, where PDF stacks and aggregates the same parts of each patch during vertical cutting and stitches them into a new sequence; The bivariate modeling module is used to model the long-term series and short-term series Perform feature extraction to obtain features and features And the results and Aggregate into the attention result Among them, the linear layer is used to After temporal encoding, the self-attention mechanism is used to send the constant sequence to the Transformer encoder. The features are then aggregated according to the original temporal order, and finally a linear layer is used as a predictor to generate the final predicted sequence. For short sequence variation feature extraction, PDF uses convolutional layers and nonlinear activation functions to capture the local features of short sequences. Each patch is then spliced ​​together in temporal order and mapped back to the original sequence using a linear layer. Inputting the aggregated N+1 attention results into a predictor based on a linear fully connected layer to obtain photovoltaic power prediction data, wherein the photovoltaic power prediction data is the photovoltaic power prediction data adjusted based on the attention results of the N meteorological data; Update the model parameters of the PDF deep learning prediction model through back propagation to obtain the model with minimum loss; update the learning rate, train the model multiple times, use the early stopping mechanism, and obtain the optimal model under the current parameters as the trained PDF deep learning prediction model; the validation set indicators decrease, and the next time it is worse than the previous one, use the validation set to verify the model, and use the test set to test the model.

4. The deep learning photovoltaic output prediction method based on meteorological feature mining according to claim 2 is characterized in that: When N=3, the mutual information is arranged from large to small, and the meteorological factors corresponding to the first three values ​​include total irradiance, normal direct irradiance, and horizontal surface scattered irradiance.

5. A deep learning photovoltaic output prediction system based on meteorological feature mining, characterized by: Used to execute the method described in any one of claims 1 to 4, a PDF deep learning prediction model is adopted, in which N+1 periodic feature modules are used to extract features based on photovoltaic power historical data and N types of meteorological historical data, respectively, and the obtained N+1 attention results are aggregated and input into a predictor based on a linear architecture, and the model parameters are updated by back propagation; the periodic feature module is composed of a multi-period decoupling module and a bivariate modeling module, the multi-period decoupling module obtains the long-term sequence and short-term sequence of the input data through learning, and the bivariate modeling module predicts the long-term sequence and the short-term sequence according to different mapping methods to obtain the attention result; the prediction result output by the PDF deep learning prediction model is the photovoltaic power prediction data obtained by adjusting the N types of meteorological factors.

Images