Photovoltaic power prediction method and system with seasonal law characteristics
Through high-dimensional feature screening and seasonal position coding combined with expanding causal convolution network, the problem of high-dimensional feature utilization and seasonal changes in photovoltaic power prediction is solved, and the prediction accuracy and model efficiency are improved.
Patent Information
- Application Number
- CN202410912788.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-09
- Publication Date
- 2024-11-08
AI Technical Summary
Existing PV power prediction methods are difficult to effectively utilize high-dimensional features, ignore seasonal changes, resulting in increased model complexity and difficulty in assessing contributions of prediction tasks, and high resource demand.
Through high-dimensional feature screening, seasonal position coding based on the characteristics of the cosine-cosine phase, and the expansion causal convolutional network enhancement Informer architecture, the optimal feature combination is selected and photovoltaic power prediction is carried out to improve the model's seasonal perception ability and high-dimensional feature learning. ability.
Adaptive analysis of high-dimensional long-time series features is realized, which improves prediction accuracy and curve fit, reduces calculation complexity, and enhances the ability to capture seasonal changes.
Smart Images

Figure CN118916665A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of photovoltaic system power prediction, and in particular to a photovoltaic power prediction method and system with seasonal regularity characteristics. Background Art
[0002] The increasing severity of global environmental problems has prompted the global energy structure to shift towards an energy system dominated by new energy. Solar energy has become one of the most popular renewable energy sources due to its easy capture characteristics and huge energy reserves. In the application of solar energy, photovoltaic power generation has become the main utilization method due to its low cost and large-scale deployment advantages. However, photovoltaic power generation is unstable and intermittent due to the influence of various environmental factors, which poses potential safety hazards to the stable operation of the power grid. Therefore, achieving accurate and stable photovoltaic power prediction is of great significance to improving the large-scale application of photovoltaic power generation in the future.
[0003] There are various ways to implement photovoltaic power prediction tasks, and physical, mathematical and data-driven methods are widely used. With the development of hardware equipment, data-driven deep neural network algorithms have shown certain performance advantages in sequence prediction tasks. The core idea of data-driven algorithms is to learn the distribution law of data through deep networks and fit the data according to the learned knowledge. Similar days are clustered for prediction tasks through Euclidean distance, and deep networks are designed to learn and predict clustered day data. The K-Means++ algorithm can also search for similar samples of the characteristics of target data for training prediction networks. The probabilistic integration method can perform integrated estimation of prediction variables for prediction tasks. The method of selecting similar data for target tasks has the prerequisite that the target data type is known. However, the weather conditions in the prediction task are more changeable and unpredictable, which makes the deep network trained with targeted data have poor generalization performance.
[0004] In order to meet the challenges of continuous prediction tasks, we can learn feature knowledge in long time series data and apply it to continuous and unknown prediction tasks. Segmented multi-step prediction of long time series is performed through deep recurrent networks. Bi-LSTM can also be optimized through genetic algorithms to process long-term photovoltaic output data. Current research on long sequence data faces the challenge of quadratic growth of memory usage and time complexity. Therefore, in order to reduce the resource requirements of the algorithm, it is also possible to choose to use only a single power data for model training and prediction, but this results in the method being limited to expanding feature knowledge from the power data dimension, without considering the physical influence of photovoltaic power output on multiple environmental factors, which will cause the available environment of the model to be too single.
[0005] To solve the above problems, high-dimensional features can be used as auxiliary prediction solutions. Denoising autoencoders (DAEs) can reduce the dimensionality of high-dimensional features and use the reduced data for deep network training. Channel attention is used to extract features of various environmental variables and is combined with temporal convolutional networks for power prediction. Two-stream networks can learn various meteorological factors of photovoltaic systems from a temporal and spatial perspective and use them for power prediction. However, although the use of high-dimensional features improves the interpretability of photovoltaic power, there are still some limitations. The long-term seasonal changes of photovoltaic systems can be easily overlooked. At the same time, the high-dimensional feature space increases the complexity of the network, making the complex nonlinear relationship between each feature and the target label difficult to evaluate the contribution of the feature to the prediction task.
[0006] In view of this, the present invention is proposed. Summary of the invention
[0007] In view of this, the present invention discloses a photovoltaic power prediction method with seasonal regularity characteristics. This method aims at solving the problems that the high-dimensional characteristics of the environment in the current photovoltaic power prediction field are difficult to distinguish the beneficial differences of the prediction task; the common system has low attention to the impact of photovoltaic seasonal changes; and the high-dimensional long-sequence photovoltaic data has great training limitations in the algorithm. A long-term photovoltaic power prediction method with seasonal regularity characteristics is proposed.
[0008] The photovoltaic power prediction method of the present invention mainly includes three parts. The first part performs high-dimensional feature screening, and selects the optimal feature combination based on the combination of feature bottleneck scores and optimal feature combination conditions. The second part proposes seasonal targeted position coding based on sine and cosine amplitude-phase characteristics, and uses different coding methods for the four seasons to increase the seasonal perception ability of the model. The third part designs an expanded causal convolutional network to strengthen the Informer architecture and enhance the high-dimensional feature learning ability of the model. Finally, seasonal position coding is added to the selected optimal feature combination, and is used to enhance the training of the Informer network to perform prediction tasks.
[0009] Specifically, the present invention is achieved through the following technical solutions:
[0010] The present invention provides a photovoltaic power prediction method with seasonal regularity characteristics, comprising the following steps:
[0011] Perform high-dimensional feature screening on the original feature vector, and select the optimal feature combination based on the combination of feature bottleneck score and screening conditions;
[0012] A seasonal position coding based on sine and cosine amplitude and phase characteristics is established, and the seasonal position coding is used to replace the original position coding of the original Informer network. The training of the Informer network is enhanced by combining the optimal feature combination to perform photovoltaic power prediction.
[0013] In addition to providing a photovoltaic power prediction method, the present invention also provides a photovoltaic power prediction system with seasonal regularity characteristics, including:
[0014] Screening module: used to perform high-dimensional feature screening on the original feature vector, and screen out the optimal feature combination based on the combination of feature bottleneck score and screening conditions;
[0015] Prediction module: used to establish seasonal position coding based on sine and cosine amplitude and phase characteristics, use seasonal position coding to replace the original position coding of the original informer network, and combine the optimal feature combination to enhance the training of the informer network to predict photovoltaic power.
[0016] In summary, the solution of the present invention has the following beneficial effects:
[0017] The feature adaptive analysis and screening of high-dimensional long time series are realized, and seasonal coding is performed according to the seasonal characteristics of the photovoltaic system. First, the proposed feature beneficialness evaluation index: the feature bottleneck score FBS can effectively quantify the effectiveness of each variable and output power. The optimal feature combination screened by FBS can achieve the best prediction effect. Secondly, the amplitude and phase characteristics of the sine and cosine can distinguish the data of the four seasons within a certain value range through seasonal position coding, which improves the model's ability to distinguish the characteristics of the four seasons and also improves the prediction accuracy. Finally, the Informer model using the dilated causal convolution module can better complete the feature learning and prediction tasks of long sequences through the optimal feature combination and seasonal position coding. Compared with other prediction models, it has the advantages of improved prediction accuracy and curve fitting. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Various other advantages and benefits will become apparent to those skilled in the art by reading the following detailed description of the preferred embodiments. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. The same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:
[0019] Figure 1 The overall framework of the short-term photovoltaic power forecasting method for adaptive selection of high-dimensional features and seasonal targeting;
[0020] Figure 2 is the mutual information heat map of each feature;
[0021] Figure 3 Visualization of feature bottleneck scores for each feature in each season;
[0022] Figure 4 Visualize the prediction results for different feature combinations;
[0023] Figure 5 Visualization of the position codes for the four seasons;
[0024] Figure 6 This is the four-season photovoltaic power prediction result using the FASS-DCI method;
[0025] Figure 7 A flowchart of a computer device provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0026] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are only examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.
[0027] The terms used in the present disclosure are only for the purpose of describing specific embodiments and are not intended to limit the present disclosure. The singular forms "a", "said" and "the" used in the present disclosure and the appended claims are also intended to include plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more of the associated listed items.
[0028] It should be understood that although the terms first, second, third, etc. may be used in the present disclosure to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present disclosure, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" used herein may be interpreted as "at the time" or "when" or "in response to determining",
[0029] Example
[0030] The technical solution adopted by the present invention is a high-dimensional feature adaptive selection and seasonal targeted short-term photovoltaic power prediction method (FASS-DCI), and the implementation steps of the method are as follows:
[0031] A. Adaptive selection of high-dimensional features
[0032] The high-dimensional feature vectors and labels in the photovoltaic system constitute a high-dimensional feature space. There are obvious differences in the degree to which different features contribute to the interpretability of the labels. If a simple combination of original features is directly trained and used for prediction tasks, the performance of the task will be saturated or even decline. In order to select the optimal feature combination for the prediction task, we quantify the linear or nonlinear correlation between each variable and the label in the system, and propose a new metric - "Feature Bottleneck Score (FBS)" to measure the actual prediction effect of the feature combination. Then, we combine different feature configuration sets according to the feature bottleneck score, and design a feature screening condition to select the optimal feature configuration set.
[0033] 1) Feature bottleneck score calculation method
[0034] Since there are complex linear or nonlinear relationships between the features of the photovoltaic system and between the features and the labels, and the weight of each feature's influence on the label is not determined by the degree of linearity or nonlinearity between them, we choose mutual information, which can capture the degree of information contained between data, as a measure of the correlation between features and between features and labels for quantification.
[0035] First, the mutual information between the kth and lth input features is calculated as shown below:
[0036]
[0037] Where p(x ik ,x il ) is x ik and x il The joint probability distribution of ik ) and P(x il ) is x ik and x il The marginal probability distribution of .
[0038] Next, the mutual information between the kth input feature and the lth label feature is calculated as follows:
[0039]
[0040] The target features with high correlation with the label and low redundancy with other features should be considered first to deal with the feature bottleneck phenomenon. Based on this idea, we design the feature bottleneck score calculation as follows: First, calculate I(X k ,Y l ) and I(X k ,X l). Considering that each feature has been normalized before calculation, to ensure that the difference is always positive, we subtract it from 1. Next, an exponential function is used to reward features with high correlation with the target label, while penalizing features with high redundancy with other features. The final result is taken as the feature bottleneck score of feature k. It is described as k FBS It is expressed as follows, where exp represents an exponential function with the natural logarithm e as the base.
[0041]
[0042] 2) Optimal feature combination conditions
[0043] According to the feature bottleneck score of each feature, different feature configuration sets can be combined from high to low, and then the optimal feature combination can be selected according to the selection conditions.
[0044] The conditions for high-dimensional feature screening are set as follows: The optimal feature configuration set X best , which can make X best The predictor f trained on best (·) The prediction errors in the target task are smaller than the results of training on other feature configuration sets. T It is expressed as follows.
[0045]
[0046] Where L(·) represents the loss function, such as the mean square error loss function (RMSE), the cross entropy loss function (CEL), etc., and Y i T represents the target label in the target task, n T Represents the number of samples of the target task. Through this condition, X best Screening and use for subsequent prediction tasks.
[0047] B Enhanced Informer Predictor Using Seasonal Positional Encoding and Dilated Causal Convolution
[0048] 1) Calculation of seasonal position coding based on sine and cosine amplitude and phase characteristics
[0049] (1) Phase shift factor
[0050] The integer value of a given seasonal variable is represented as where n z represents the number of seasonal variables. Then the phase shift factor φ(z i ) can be expressed as follows
[0051]
[0052] With a period of 2π, φ(z i ) can make each seasonal variable z i The corresponding position code starting point appears 2π / n z The offset distinguishes different seasonal variables while ensuring continuous position encoding.
[0053] (2) Amplitude offset factor
[0054] A single phase shift will cause aliasing of sine and cosine functions with similar frequencies at similar time steps, which will result in low discrimination of position codes for different seasons. To prevent this phenomenon, we also add an amplitude shift factor A(z i ) is shown in the following formula
[0055]
[0056] (3) Seasonal position coding
[0057] For the four-season position coding of the photovoltaic system, the seasonal position coding of the photovoltaic system can be obtained as follows: where i represents the seasonal variable z i The index of , j represents the dimension index of the model. model Represents the dimensions of the model.
[0058]
[0059] 2) Sparsity calculation of high-dimensional feature matrix After calculating the seasonal phase and amplitude offset of the position code, in order to enhance the interpretability of the output power of the photovoltaic system, we use the photovoltaic optimal feature configuration set - matrix in the prediction task is the minimum training unit. In order to better handle long-term series problems and reduce computational complexity, samples corresponding to high-dimensional features in M are extracted and sparse before performing prediction tasks.
[0060] Denote each randomly selected sample as X min , which will not change the data distribution of the sample and the corresponding features, and can effectively reduce the sequence length. Then we define a Softmax function to calculate the given sample feature and the same dimension key k matrix P sparse The probability distribution of .
[0061]
[0062] In order to obtain the sample feature matrix with high "activity" for subsequent calculations, we use P sparse With uniform distribution P uniform =1 / n SThe discrete KL divergence of is measured as follows: Where D represents the difference distance between the two distributions.
[0063] D=KL(P sparse PP uniform )(10)
[0064] Select the X corresponding to the first u largest D min , and use it and the randomly sampled u k as the sparse query matrix Q u With the sparse key-value matrix K u , into the standard attention calculation formula A u (g) The calculation is performed as follows.
[0065]
[0066] 3) Enhanced Informer using dilated causal convolution
[0067] The long-term sample characteristics of photovoltaic power data make general prediction tasks have o(L 2 ) has high computational complexity and high-dimensional features, which is a test of the general model's ability to extract feature knowledge. Therefore, in the prediction task of our framework, we use the Informer prediction architecture based on the ProbSparse attention mechanism, self-attention distillation, and generative decoder for the optimal feature configuration set selected in the early stage, and add an expanded causal convolution layer to achieve more fine-grained high-dimensional temporal feature extraction.
[0068] (1) Dilated Causal Convolution of High-Dimensional Features
[0069] In order to enhance the network's ability to learn high-dimensional temporal features, causal convolution is used to capture local features more finely based on the global feature capture of sparse self-attention, and dilated convolution is introduced to expand the receptive field and ensure computational efficiency. ddc , the calculated result y ddc It can be expressed as shown below, and the calculation process is represented by DDC.
[0070]
[0071] Where d is the dilation factor, k is the size of the convolution kernel, f(i) represents the weight of the convolution kernel, and d·i represents the dilated convolution size corresponding to time step i. To ensure the stability of training when the network depth increases, the overall temporal convolution network adopts the form of residual connection, which is shown below. Where WN represents the weight normalization layer, Relu represents the activation function, and Dropout represents the dropout layer.
[0072] TCN=Dropout(Relu(WN(DDC(x ddc)))) (13)
[0073] (2)Encoder
[0074] The input matrix of the M prediction model consists of two parts: high-dimensional photovoltaic system features and timestamps. mean Filling causes feature redundancy. The operation from layer j to layer j+1 is expressed as follows:
[0075] X minj+1 =MaxPool(Relu(Conv1d(A u (X minj )))) (14)
[0076] (3)Decoder
[0077] Different from the step-by-step dynamic decoding process of traditional time series tasks, the Decoder adopts a one-time static decoding strategy, which can accurately capture the long-term dependencies of the data while improving the computational efficiency. The input vector of the decoder is represented as follows. The length is The starting placeholder (starting token) is the target placeholder filled with 0.
[0078] X de ={X token ,X0} (15)
[0079] The decoder consists of two layers of multi-head attention. The first multi-head self-attention adds a mask mechanism and is represented by Masked A. The second multi-head attention interacts with the output of the encoder and is represented by EMA. The overall decoder architecture is a standard decoder. The forward propagation process of the decoder is shown below, where LN represents the Layer Norm operation and FC represents the fully connected layer.
[0080] Y t =FC(TCN(LN(EMA(MaskA(A u (X de ))))))(19)
[0081] (4) Loss fountain
[0082] The back-propagation phase uses MSE as the loss function of Informer. The loss value is calculated as follows, where θ I is the model parameter of Informer, Y i is the true value of the i-th sample sequence, Yi t is the predicted value of the i-th sample sequence.
[0083]
[0084] The method of the above embodiment is used to carry out the experiment, specifically as follows Figure 1 As shown in Table 1, the data is taken from the Yulara Solar System located on the north side of Ayers Rock, a famous landmark in central Australia. The No. 2 service station array was selected as the experimental object, and its specific parameters are shown in Table 1.
[0085] Table 1 System parameters of service array No. 2
[0086]
[0087] According to the system failure maintenance information provided by the data website, in order to ensure the integrity of the data as much as possible, data under 8 conditions of sunny and rainy weather conditions in the four solar terms of spring, summer, autumn and winter in 2019 were selected for the experiment. Each type of data contains 8 dimensions, and each dimension and its corresponding meaning are shown in Table 2.
[0088] Table 2 Each dimension of the dataset and its corresponding meaning
[0089]
[0090] The data is used as a cycle of 8 consecutive days for the prediction task. The environmental dimensions of the first 7 days are used as model training features, and the power dimension of the 8th day is used as a label to construct the training and test data sets.
[0091] A. Data screening stage
[0092] Step 1: To ensure that each feature has the same weight on power, all data are normalized to the interval [0,1] by feature. Separate sunny and rainy days, select the first day of training data in each season under sunny and rainy conditions to calculate the mutual information value between each feature, and normalize it by feature. The results are as follows: Figure 2 shown.
[0093] Step 2: Then calculate the feature bottleneck score of each feature in each season and list them in Table 3 and Table 4, and visualize them as Figure 3 shown.
[0094] Table 3 Bottleneck scores of various characteristics on sunny days
[0095]
[0096] Table 4 Bottleneck scores of various characteristics on rainy days
[0097]
[0098] Step 3: Combine features from low to high according to feature bottleneck scores, and find the optimal feature configuration set through prediction tasks. The combination methods and experimental results are listed in Tables 5 and 6. Figure 4 The prediction evaluation indicators are visualized in .
[0099] Table 5. Sunny weather prediction results of different characteristic combinations
[0100]
[0101] Table 6 Rainy day prediction results of different characteristic combinations
[0102]
[0103] B. Prediction stage
[0104] Step 1: Select four seasons of spring, summer, autumn and winter and two weather conditions of sunny and rainy, totaling 8 test environments. At the same time, for each type of environment, select 7 consecutive days of data in each cycle of the training set to train the FASS-DCI framework, and use the continuous and invisible eighth day power data for prediction testing. To avoid the volatility of the results, the evaluation indicators of each type of test environment are averaged over ten rounds. The specific structure of the FASS-DCI framework is shown in Table 7.
[0105] Table 7FASS-DCI network structure
[0106]
[0107]
[0108] Step 2: Based on the seasonal information of the forecast task, the seasonal variable is 4. According to equations (5) and (6), the amplitude offset A(z i ) and the phase angle offset φ(z i ) as shown in Table 8:
[0109] Table 8 Amplitude and phase angle offset of position code in different seasons
[0110]
[0111] The A(z i ) and φ(z i ) into equations (7) and (8), and visualize the first 288 dimensions of the encoding value of the length of a day (288 points) as Figure 5 shown.
[0112] Step 3: Use the calculated four-season position codes to train the corresponding model according to the seasons using the four-season data, and visualize the experimental results of one of the ten rounds of verification as follows: Figure 6 The ten rounds of mean values of the evaluation indicators are listed in Table 9.
[0113] Table 9 FASS-DCI framework prediction evaluation indicators
[0114]
[0115] The results show that the overall performance and prediction effect of the model are good under sunny conditions. Among them, the model achieved the best results in autumn, with the RMSE\MAE indicators reaching the lowest level of 1.7505 / 1.1917. R2 reached the highest level of 0.9993. The performance was the worst in summer, with the RMSE\MAE reaching the highest level of 3.7548 / 2.2724. R2 reached the lowest level of 0.9967. Overall, the average RMSE and MAE of the model on sunny days are low, and R2 is high, indicating that the model has high prediction accuracy and fit.
[0116] Due to the unpredictability of rainy day conditions and the large and irregular fluctuations of various features including power data, the predictor has a higher difficulty in learning the features of each dimension, which results in a worse overall prediction effect than that of sunny days. The model still shows the worst prediction effect in summer, with RMSE / MAE reaching 14.1251 / 7.7851 and R2 dropping to 0.9530. The prediction effect in spring is relatively good, with RMSE / MAE reaching 13.0507 / 6.4512 and R2 reaching 0.9651. The average R2 of the model as a whole can reach 0.9430, indicating that for unknown rainy day prediction tasks, the model can fit the general trend of power changes.
[0117] In addition, in addition to providing a photovoltaic power prediction method with seasonal regularity characteristics, the present invention also provides a prediction system corresponding to the above prediction method, specifically including:
[0118] Screening module: used to perform high-dimensional feature screening on the original feature vector, and screen out the optimal feature combination based on the combination of feature bottleneck score and screening conditions;
[0119] Prediction module: used to establish seasonal position coding based on sine and cosine amplitude and phase characteristics, use seasonal position coding to replace the original position coding of the original informer network, and combine the optimal feature combination to enhance the training of the informer network to predict photovoltaic power.
[0120] In specific implementation, the above modules can be implemented as independent entities, or can be arbitrarily combined and implemented as the same or several entities. The specific implementation of the above units can refer to the previous method embodiments, which will not be repeated here.
[0121] Figure 7 This is a schematic diagram of the structure of a computer device disclosed in the present invention. Figure 7 As shown, the computer device 400 includes at least a memory 402 and a processor 401; the memory 402 is connected to the processor via a communication bus 403, and is used to store computer instructions executable by the processor 401, and the processor 401 is used to read computer instructions from the memory 402 to implement the steps of the method described in any of the above embodiments.
[0122] For the above-mentioned device embodiments, since they basically correspond to the method embodiments, the relevant parts can refer to the partial description of the method embodiments. The device embodiments described above are only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without paying any creative work.
[0123] Computer readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media and memory devices, including, for example, semiconductor memory devices (e.g., EPROM, EEPROM and flash memory devices), magnetic disks (e.g., internal disks or removable disks), magneto-optical disks, and CD ROM and DVD-ROM disks. The processor and memory may be supplemented by, or incorporated in, special purpose logic circuits.
[0124] Finally, it should be noted that although this specification contains many specific implementation details, these should not be interpreted as limiting the scope of any invention or the scope of protection claimed, but are mainly used to describe the features of specific embodiments of specific inventions. Certain features described in multiple embodiments within this specification may also be implemented in combination in a single embodiment. On the other hand, the various features described in a single embodiment may also be implemented separately in multiple embodiments or in any suitable sub-combination. In addition, although features may work in certain combinations as described above and even initially claimed as such, one or more features from the claimed combination may be removed from the combination in some cases, and the claimed combination may point to a sub-combination or a variation of a sub-combination.
[0125] Similarly, although operations are depicted in a particular order in the accompanying drawings, this should not be understood as requiring that these operations be performed in the particular order shown or performed sequentially, or requiring that all illustrated operations be performed to achieve the desired results. In some cases, multitasking and parallel processing may be advantageous. In addition, the separation of various system modules and components in the above-described embodiments should not be understood as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product, or packaged into multiple software products.
[0126] Thus, specific embodiments of the subject matter have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims can be performed in a different order and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the particular order or sequential order shown to achieve the desired results. In some implementations, multitasking and parallel processing may be advantageous.
[0127] The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present disclosure shall be included in the scope of protection of the present disclosure.
Claims
1. A photovoltaic power prediction method with seasonal regularity characteristics, characterized in that: The steps include: Perform high-dimensional feature screening on the original feature vector, and select the optimal feature combination based on the combination of feature bottleneck score and screening conditions; A seasonal position coding based on sine and cosine amplitude and phase characteristics is established, and the seasonal position coding is used to replace the original position coding of the original Informer network. The training of the Informer network is enhanced by combining the optimal feature combination to perform photovoltaic power prediction.
2. The photovoltaic power prediction method according to claim 1, characterized in that: The method for calculating the feature bottleneck score comprises the following steps: Calculate the mutual information between the kth and lth input features as follows: p(x ik ,x il ) is x ik and x il The joint probability distribution of ik ) and P(x il ) is x ik and x il The marginal probability distribution of The mutual information between the kth input feature and the lth label feature is expressed as follows: Calculate I(X k ,Y l ) and I(X k ,X l ) and subtracted from 1, the final result is the feature bottleneck score as shown in the following formula:
3. The photovoltaic power prediction method according to claim 2, characterized in that: The method for obtaining the optimal feature combination comprises the following steps: According to the feature bottleneck score of each feature, different feature configuration sets are combined from high to low. Then, the optimal feature combination is selected according to the screening conditions, and the screening conditions include: Get the optimal feature configuration set X best , which can make X best The predictor f trained on best (·) The prediction errors in the target task are all smaller than the results of training on other feature configuration sets. T It is expressed as follows: Among them, L(·) represents the loss function, Y i T represents the target label in the target task, n T Represents the number of samples of the target task, and the constraints of this condition are applied to X best to filter.
4. The photovoltaic power prediction method according to claim 3, characterized in that: The method for establishing the seasonal position coding is to modify the sine and cosine parameters by introducing a phase offset factor and an amplitude offset factor, thereby performing seasonal offset; The phase shift factor is φ(z i ): The integer value of a given seasonal variable is represented as Where n z represents the number of seasonal variables, φ(z i ) is expressed as shown in formula (4): With a period of 2π, φ(z i ) so that each seasonal variable z i The corresponding position code starting point appears 2π / n z The offset of , while ensuring the continuous position coding, distinguishes different seasonal variables; The amplitude shift factor is A(z i ): A(z i ) so that each seasonal variable z i The corresponding position code has a maximum amplitude of π / n z deviation.
5. The photovoltaic power prediction method according to claim 4, characterized in that: The seasonal position encoding is obtained by setting nz in (4) and (5) to 4 and changing φ(z i ), A(z i ) is combined with the sine and cosine functions to obtain the seasonal position encoding of the photovoltaic system as shown in equations (6) and (7), where i represents the seasonal variable z i The index of the model, j represents the dimension index of the model, d model Represents the dimensions of the model:
6. The photovoltaic power prediction method according to claim 5, characterized in that: During the training process of the Informer network, causal convolution is used to capture local features more finely based on the global feature capture of sparse self-attention, and dilated convolution is introduced to calculate the result y ddc It can be expressed as follows, and the calculation process is represented by DDC: Where d is the dilation factor, k is the size of the convolution kernel, f(i) represents the weight of the convolution kernel, and d·i represents the dilated convolution size corresponding to time step i; Where WN represents the weight normalization layer, Relu represents the activation function, and Dropout represents the discard layer: TCN=Dropout(Relu(WN(DDC(x ddc )))) The input matrix of the M prediction model consists of two parts, namely, high-dimensional photovoltaic system features and timestamps. The operation from the jth layer to the j+1th layer is expressed as follows: X minj+1 =MaxPool(Relu(Conv1d(A u (X minj )))) The input vector of the decoder is represented as follows, where The length is The starting placeholder (starting token) is the target placeholder filled with 0: X de ={X token ,X0} The decoder consists of two layers of multi-head attention. The first multi-head self-attention adds a Mask masking mechanism, which is represented by MaskedA. The second multi-head attention interacts with the output of the Encoder and is represented by EMA. The overall decoder architecture is a standard decoder. The forward propagation process of the decoder is as follows, where LN represents the Layer Norm operation and FC represents the fully connected layer: Y t =FC(TCN(LN(EMA(MaskA(A u (X de )))))).
7. The prediction system of the photovoltaic power prediction method according to any one of claims 1 to 6, characterized in that: include: Screening module: used to perform high-dimensional feature screening on the original feature vector, and screen out the optimal feature combination based on the combination of feature bottleneck score and screening conditions; Prediction module: used to establish seasonal position coding based on sine and cosine amplitude and phase characteristics, use seasonal position coding to replace the original position coding of the original informer network, and combine the optimal feature combination to enhance the training of the informer network to predict photovoltaic power.
8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the photovoltaic power prediction method according to any one of claims 1 to 6 are executed.
9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the photovoltaic power prediction method according to any one of claims 1 to 6 are implemented.
Citation Information
Cited By
Short-term net load prediction method and system considering seasonal characteristics
CN121261333A
