Distributed photovoltaic short-term power combined prediction method considering extreme events
By using dynamic time alignment and quantile anomaly detection, combined with time series generative adversarial networks and hybrid prediction models, the data shortage problem caused by frequent extreme events in distributed photovoltaic systems is solved, and the short-term power prediction accuracy and grid stability are improved.
Patent Information
- Application Number
- CN202510781047.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-12
- Publication Date
- 2025-09-26
AI Technical Summary
In existing technologies, the frequent occurrence of extreme events in distributed photovoltaic systems leads to insufficient data samples. Traditional prediction methods are unable to effectively predict future extreme fluctuations, affecting grid stability and energy management decisions.
The baseline power pattern is extracted through dynamic time alignment, the extreme event scenario is divided using the quantile anomaly detection algorithm, the sample is expanded by combining the time series generative adversarial network, and a hybrid architecture of the gradient boosting tree and the spatiotemporal attention network is constructed for prediction, and the dynamic weight smoothing mechanism is used to output the results.
The accuracy of distributed photovoltaic short-term power prediction has been improved, and the data characteristics of conventional and extreme event scenarios can be better learned, thereby improving prediction accuracy and computational efficiency.
Smart Images
Figure CN120706623A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of photovoltaic power generation prediction, and in particular relates to a distributed photovoltaic short-term power combination prediction method considering extreme events. Background Art
[0002] Against the backdrop of the "dual carbon" goals, the construction of a new power system with an increasing proportion of renewable energy is accelerating. Distributed photovoltaics are increasingly contributing to my country's energy mix, becoming a major source of energy. However, distributed photovoltaic power generation exhibits intermittent, random, and volatile characteristics, posing significant challenges to the stable operation and control of power systems. Accurate distributed photovoltaic power forecasting is crucial for improving power system dispatch efficiency and alleviating the burden of grid fluctuations.
[0003] During the operation of distributed photovoltaic systems, extreme weather events such as sandstorms, solar eclipses, heavy rain, and snowstorms, as well as natural or man-made damage to photovoltaic panels, low-precision measurement equipment, and false detections can significantly impact photovoltaic power forecasts. As climate change intensifies, the frequency and intensity of extreme weather events are increasing, making the accuracy of photovoltaic system forecasts more uncertain in the face of these changes, which in turn affects grid stability and energy management decisions.
[0004] Existing methods for defining extreme events encountered during distributed photovoltaic operation are inconsistent. Furthermore, when extreme events occur, available historical data is often limited, resulting in insufficient sample sizes. Extreme events exhibit extremely short time windows and sudden occurrences, which limits traditional forecasting methods based on historical data. In these cases, the characteristics of extreme events are often not fully incorporated into existing data training, making it impossible to effectively predict future extreme fluctuations. To fully consider the impact of extreme events on distributed photovoltaic power forecasting, in addition to standardizing the definition of extreme events, it is necessary to implement data augmentation measures during data processing and modeling, especially when extreme events occur frequently. Summary of the Invention
[0005] In order to solve the above technical problems, the present application provides a distributed photovoltaic short-term power combination prediction method considering extreme events. This method solves the problem that frequent extreme events affect data quality and power prediction accuracy, and improves the distributed photovoltaic short-term power prediction accuracy.
[0006] In order to achieve the above object, the present invention is achieved through the following technical solutions:
[0007] This application is a distributed photovoltaic short-term power combination forecasting method considering extreme events, which specifically includes the following steps:
[0008] Step 1: Extract the benchmark power pattern from the distributed photovoltaic power time series data through dynamic time alignment;
[0009] Step 2: Based on the dynamic alignment difference distribution of the baseline power mode, the quantile anomaly detection algorithm is used to divide the extreme event scenario to obtain extreme event feature data;
[0010] Step 3: Use the time series generative adversarial network to reconstruct the distribution and expand the samples of the extreme event feature data obtained in step 2;
[0011] Step 4: Build a scenario-based prediction model group, use a hybrid architecture of gradient boosting tree and spatiotemporal attention network for prediction, and output the final result through a dynamic weight smoothing mechanism.
[0012] A further improvement of the present application is that step 1 extracts a reference power pattern from distributed photovoltaic power time series data through dynamic time alignment, specifically comprising the following steps:
[0013] Step 1.1: Construct the distributed photovoltaic power sequence alignment difference matrix: For two distributed photovoltaic power sequences of length n, Q = {q1, q2, ..., q n} and C={c1,c2,...,c n}, construct the difference matrix M:
[0014] M(i,j)=||q i -c j || p +λ·max(0,|ij|-τ)
[0015] Among them, ||·|| p Indicates L p Normalized distance, p = 1 or 2, λ is the path deviation penalty coefficient, τ is the maximum allowed deviation step size, i represents the time point index in the first power sequence, and j represents the time point index in the second power sequence;
[0016] Step 1.2. Calculate the cumulative difference matrix Γ(i, j): Starting from the lower left corner of the difference matrix M, search for the optimal path in the right, upper, and upper-right directions. Specifically:
[0017] Γ(i,j)=M(i,j)+min(Γ(i+1,j),Γ(i,j+1),Γ(i+1,j+1));
[0018] Step 1.3, determine the optimal alignment path: backtrack from the upper right corner of the difference matrix M to select the path with the minimum cumulative difference and calculate the total alignment cost C total :
[0019]
[0020] Where k represents the path point index on the optimal alignment path;
[0021] Step 1.4: Select the total alignment cost C total The smallest N-segment power sequence is used as the reference power mode, and N=30 corresponds to the monthly cycle characteristics.
[0022] A further improvement of the present application is that: in step 2, based on the reference power mode, the alignment difference value sequence X={x1, x2, ..., x m After sorting in ascending order, the anomaly boundary is determined by the quantile anomaly detection algorithm, which specifically includes the following steps:
[0023] Step 2.1. Calculate the median Q2 of the alignment difference value sequence X, i.e. the second quartile:
[0024]
[0025] Step 2.2, calculate the first quartile Q1 and the third quartile Q3 of the alignment difference value sequence X:
[0026]
[0027] Wherein, τ is any positive integer;
[0028] Step 2.3: Define the abnormality determination interval, i.e., the extreme event characteristic data:
[0029] Ω=(-∞,Q1-κI QR ]∪[Q3+κI QR ,+∞)
[0030] Among them, κ = 3, I QR =Q3-Q1.
[0031] A further improvement of the present application is that step 3 specifically includes the following steps:
[0032] Step 3.1, distribution reconstruction: The encoder of the temporal generation adversarial network combines the static features s of the temporal generation adversarial network with the temporal features x of the temporal generation adversarial network 1:T Mapping to the latent space h of the temporal generative adversarial network s ,h t , the generator G synthesizes data to meet:
[0033]
[0034] Among them, h s ~p(h s ),h t ~p(ht );
[0035] The sample expansion rule is: generate K sets of synthetic data for each extreme event scenario to form an enhanced data set Where D real For existing real data sets;
[0036] Step 3.2, adversarial training: discriminator D distinguishes real data (s, x 1:T ) and generate data The loss function is:
[0037]
[0038] Step 3.3, temporal consistency constraint: supervision loss ensures the temporal dependency of the generated sequence:
[0039]
[0040] A further improvement of the present application is that the gradient boosting tree in step 4 specifically includes the following steps:
[0041] Step 4.1.1, Histogram optimization: Discretize the continuous features into k buckets and accelerate the calculation of split gain through histogram difference:
[0042]
[0043] Among them, G b ,H b is the gradient estimator of bucket b, u is the regularization coefficient; G L ,H L is the gradient estimator of the child node on the left of the split point; G R ,H R is the gradient estimator of the child node to the right of the split point.
[0044] Step 4.1.2: Leaf growth with depth limit: Only split the leaf with the largest gain each time, with a maximum depth limit of d max =8.
[0045] A further improvement of the present application is that in step 4, the spatiotemporal attention network specifically includes the following steps:
[0046] Step 4.2.1, Multi-head self-attention encoding: Use the multi-head self-attention mechanism to extract the temporal dependency of step 3.3:
[0047] H attn =Concat(head1,..,head h )W O
[0048]
[0049] Among them, H attn is the multi-head self-attention encoding output, head i represents the output of the i-th attention head, i=1,2,...h,W O is the weight transformation matrix, Concat is the matrix concatenation function, is the query, key, and value matrix, is the projection weight, h is the number of attention heads, d k =d model / h;
[0050] Step 4.2.2, gated state fusion: multi-head self-attention encoding output H attn With LSTM hidden state Fusion:
[0051]
[0052] Among them, W g is the gate weight matrix, which is used to control the fusion ratio of attention and LSTM hidden state, W m is the memory transformation weight matrix, used to generate new memory states; b g 、b m is the bias term.
[0053] A further improvement of the present application is that in step 4, the dynamic weight smoothing mechanism is specifically:
[0054] v t =α×v t-1 +(1-α)×θ t
[0055] Among them, v t 、v t-1 is the exponential moving average of the current moment and the previous moment, θ t is the original data at the current moment. After the above steps, the distributed photovoltaic short-term power prediction result considering extreme events is output.
[0056] The beneficial effects of this application are:
[0057] This application extracts the benchmark power pattern through dynamic time alignment technology and sliding window traversal matching, performs dynamic alignment difference analysis based on the benchmark pattern, and divides extreme event scenarios according to the quantile anomaly detection algorithm, fully verifying the necessity of considering the impact of extreme events on prediction accuracy.
[0058] This application reconstructs the distribution and expands the samples of the filtered extreme event feature data based on the time series generative adversarial network, effectively solving the problem of insufficient extreme event samples, helping the deep learning model to better learn the time series characteristics of the samples and improve prediction accuracy.
[0059] This application uses a combined prediction model based on gradient boosting trees and spatiotemporal attention networks for both common and extreme event scenarios. Results show that the combined prediction model can better learn the data characteristics of both common and extreme event scenarios, while also improving computational efficiency and achieving higher prediction accuracy than conventional prediction methods. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1 It is a flow chart of the prediction method of this application.
[0061] Figure 2 This is the network structure of the temporal generative adversarial network of this application.
[0062] Figure 3 This is a schematic diagram of the gradient boosting tree model of this application.
[0063] Figure 4 This is a schematic diagram of the spatiotemporal attention network model of this application.
[0064] Figure 5 This is a schematic diagram of the calculation process of dynamic time alignment in this application.
[0065] Figure 6 It is a box plot of the extreme event definition of this application.
[0066] Figure 7 This is the small sample expansion result of this application based on the temporal generation adversarial network.
[0067] Figure 8 This is the power prediction result for a conventional scenario.
[0068] Figure 9 It is the power prediction result of extreme event scenario.
[0069] Figure 10 It is the power prediction result before and after data smoothing. DETAILED DESCRIPTION
[0070] The following diagrams illustrate embodiments of the present invention. For clarity, many practical details are included in the following description. However, it should be understood that these practical details are not intended to limit the present invention. In other words, in some embodiments of the present invention, these practical details are not essential.
[0071] like Figure 1As shown, the present application is a distributed photovoltaic short-term power combination prediction method considering extreme events, which specifically includes the following steps:
[0072] Step 1: Dynamic time alignment, such as Figure 5 The calculation process shown in the figure, which extracts the reference power pattern from the distributed photovoltaic power time series data, includes the following steps:
[0073] Step 1.1: Construct the distributed photovoltaic power sequence alignment difference matrix: For two distributed photovoltaic power sequences of length n, Q = {q1, q2, ..., q n} and C={c1,c2,...,c n}, construct the difference matrix M:
[0074] M(i,j)=||q i -c j || p +λ·max(0,|ij|-τ)
[0075] Among them, ||·|| p Indicates L p Normalized distance, p = 1 or 2, λ is the path deviation penalty coefficient, τ is the maximum allowed deviation step size, i represents the time point index in the first power sequence, and j represents the time point index in the second power sequence;
[0076] Step 1.2. Calculate the cumulative difference matrix Γ(i, j): Starting from the lower left corner of the difference matrix M, search for the optimal path in the right, upper, and upper-right directions. Specifically:
[0077] Γ(i,j)=M(i,j)+min(Γ(i+1,j),Γ(i,j+1),Γ(i+1,j+1));
[0078] Step 1.3, determine the optimal alignment path: backtrack from the upper right corner of the difference matrix M to select the path with the minimum cumulative difference and calculate the total alignment cost C total :
[0079]
[0080] Where k represents the path point index on the optimal alignment path;
[0081] Step 1.4: Select the total alignment cost C total The smallest N-segment power sequence is used as the reference power mode, and N=30 corresponds to the monthly cycle characteristics.
[0082] Step 2: Based on the dynamic alignment difference distribution of the benchmark power mode, the quantile anomaly detection algorithm is used to divide the extreme event scenario and obtain the extreme event feature data. In step 2, based on the benchmark power mode, the alignment difference value sequence X={x1,x2,...,x m}, after being sorted in ascending order, the anomaly boundary is determined by the quantile anomaly detection algorithm.
[0083] The quantile anomaly detection algorithm specifically includes the following steps:
[0084] Step 2.1. Calculate the median Q2 of the alignment difference value sequence X, i.e. the second quartile:
[0085]
[0086] Step 2.2, calculate the first quartile Q1 and the third quartile Q3 of the alignment difference value sequence X:
[0087]
[0088] Wherein, τ is any positive integer;
[0089] Step 2.3: Define the abnormality determination interval, i.e., the extreme event characteristic data:
[0090] Ω=(-∞,Q1-κI QR ]∪[Q3+κI QR ,+∞)
[0091] Among them, κ = 3, I QR =Q3-Q1.
[0092] The box plot of the extreme event definition in this application is as follows Figure 6 shown.
[0093] Step 3: Use the time series generative adversarial network to reconstruct the distribution and expand the samples of the extreme event feature data obtained in step 2.
[0094] like Figure 2 As shown, the temporal generative adversarial network includes an encoder, a decoder, a generator and a discriminator.
[0095] Step 3.1, distribution reconstruction: The encoder of the temporal generation adversarial network combines the static features s of the temporal generation adversarial network with the temporal features x of the temporal generation adversarial network 1:T Mapping to the latent space h of the temporal generative adversarial network s ,h t , the generator G synthesizes data to meet:
[0096]
[0097] Among them, h s ~p(h s ),h t ~p(h t );
[0098] The sample expansion rule is: generate K sets of synthetic data for each extreme event scenario to form an enhanced data set Where D real For existing real data sets;
[0099] Step 3.2, adversarial training: discriminator D distinguishes real data (s, x 1:T ) and generate data The loss function is:
[0100]
[0101] Step 3.3, temporal consistency constraint: supervision loss ensures the temporal dependency of the generated sequence:
[0102]
[0103] The results of small sample expansion based on temporal generation adversarial network are shown in the figure below. Figure 7 shown.
[0104] Step 4: Build a scenario-based prediction model group, use a hybrid architecture of gradient boosting tree and spatiotemporal attention network for prediction, and output the final result through a dynamic weight smoothing mechanism.
[0105] like Figure 3 As shown, the gradient boosting tree in step 4 specifically includes the following steps:
[0106] Step 4.1.1, Histogram optimization: Discretize the continuous features into k = 256 buckets and accelerate the calculation of split gain through histogram difference:
[0107]
[0108] Among them, G b ,H b is the gradient estimator of bucket b, u is the regularization coefficient; G L ,H L is the gradient estimator of the child node on the left of the split point; G R ,H R is the gradient estimator of the child node to the right of the split point.
[0109] In this step, continuous features refer to numerical features such as power, temperature, and irradiance. These features are continuous real numbers in the raw data. Continuous features can be static features, such as the installation angle of a photovoltaic panel, or time series features, such as hourly power values.
[0110] Step 4.1.2: Leaf growth with depth limit: Only split the leaf with the largest gain each time, with a maximum depth limit of d max =8.
[0111] like Figure 4 The schematic diagram of the spatiotemporal attention network model shown in Figure 2 shows that the spatiotemporal attention network specifically includes the following steps:
[0112] Step 4.2.1, Multi-head self-attention encoding: Use the multi-head self-attention mechanism to extract the temporal dependency of step 3.3:
[0113] H attn =Concat(head1,..,head h )W O
[0114]
[0115] Among them, H attn is the multi-head self-attention encoding output, head i represents the output of the i-th attention head, i=1,2,...h,W O is the weight transformation matrix, Concat is the matrix concatenation function, is the query, key, and value matrix, is the projection weight, h=8 is the number of attention heads, d k =d model / h=64;
[0116] Step 4.2.2, gated state fusion: multi-head self-attention encoding output H attn With LSTM hidden state Fusion:
[0117]
[0118] Among them, W g is the gate weight matrix, which is used to control the fusion ratio of attention and LSTM hidden state, W m is the memory transformation weight matrix, used to generate new memory states; b g 、b m is the bias term.
[0119] The dynamic weight smoothing mechanism is specifically as follows:
[0120] v t =α×v t-1 +(1-α)×θ t
[0121] Among them, v t 、vt-1 is the exponential moving average of the current moment and the previous moment, θ t is the original data at the current moment, and α=0.4 is the sensitivity factor. After the above steps, the output is the distributed photovoltaic short-term power forecast result considering extreme events. The power forecast result of the conventional scenario is as follows Figure 8 As shown in the figure, the power prediction results of extreme event scenarios are as follows: Figure 9 shown.
[0122] The rate is 15 minutes. At the same time, the period of distributed photovoltaic main output from 7:00 to 17:00 every day is selected as the prediction data set, with 41 data in each sample, a total of 89 days, and a total of 3649 data.
[0123] The benchmark mode is extracted through dynamic time alignment; based on the dynamic alignment difference distribution of the benchmark power mode, the quantile anomaly detection algorithm is used to divide the extreme event scenarios; based on the extreme event division results, the time series generative adversarial network is used to reconstruct the distribution and expand the sample of the extreme event feature data; Python software is used to build a scenario prediction model group, and a hybrid architecture of gradient boosting tree and spatiotemporal attention network is used for prediction. The final result is output through the dynamic weight smoothing mechanism. The power prediction results before and after data smoothing are as follows: Figure 10 shown.
[0124] The evaluation indicators of the prediction results are as follows:
[0125]
[0126] Table 1 Evaluation indicators of prediction errors of various models in conventional scenarios
[0127]
[0128]
[0129] As shown in Table 1, the comparison of the prediction error evaluation indicators of the conventional scenarios of various models shows that the present application (i.e.
[0130] LightGBM) method due to other methods.
[0131] Table 2 Evaluation indicators of prediction errors of various models for extreme event scenarios
[0132]
[0133] Table 2 compares the error evaluation indicators of various models under extreme events. The first row of Table 2, "LSTM-Transformer", is the method adopted by this application, which reflects the superiority of this application.
[0134] Table 3 Forecast error evaluation indicators before and after data smoothing
[0135]
[0136] Table 3 is a comparison table of the dynamic weight smoothing mechanism, showing the improvement of error index performance before and after this smoothing mechanism.
[0137] Table 4 Error evaluation index of power forecast results considering extreme events
[0138]
[0139]
[0140] Table 4 compares the prediction results for extreme events. The LightGBM model considers only conventional scenarios, while the LSTM-Transformer model considers only extreme event scenarios. The LightGBM+LSTM-Transformer method of this application combines the two methods, using different methods for different scenarios. This application, the LightGBM+LSTM-Transformer combined model, addresses the issue of frequent extreme events affecting data quality and power forecast accuracy, improving the accuracy of distributed photovoltaic short-term power forecasts.
[0141] The foregoing is merely an embodiment of the present invention and is not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.
Claims
1. A distributed photovoltaic short-term power combination forecasting method considering extreme events, characterized by: The distributed photovoltaic short-term power combination prediction method specifically includes the following steps: Step 1: Extract the benchmark power pattern from the distributed photovoltaic power time series data through dynamic time alignment; Step 2: Based on the dynamic alignment difference distribution of the baseline power mode, the quantile anomaly detection algorithm is used to divide the extreme event scenario to obtain extreme event feature data; Step 3: Use the time series generative adversarial network to reconstruct the distribution and expand the samples of the extreme event feature data obtained in step 2; Step 4: Build a scenario-based prediction model group, use a hybrid architecture of gradient boosting tree and spatiotemporal attention network for prediction, and output the final result through a dynamic weight smoothing mechanism.
2. The distributed photovoltaic short-term power combination forecasting method considering extreme events according to claim 1 is characterized by: Step 1 extracts the reference power pattern from the distributed photovoltaic power time series data through dynamic time alignment, which specifically includes the following steps: Step 1.1: Construct the distributed photovoltaic power sequence alignment difference matrix: For two distributed photovoltaic power sequences of length n, Q = {q1, q2, ..., q n } and C={c1,c2,...,c n }, construct the difference matrix M: M(i,j)=||q i -c j || p +λ·max(0,|ij|-τ) Among them, ||·|| p Indicates L p Normalized distance, p = 1 or 2, λ is the path deviation penalty coefficient, τ is the maximum allowed deviation step size, i represents the time point index in the first power sequence, and j represents the time point index in the second power sequence; Step 1.2: Starting from the lower left corner of the difference matrix M, search for the optimal path in the right, upper, and upper-right directions, and calculate the cumulative difference matrix Γ(i, j): Γ(i,j)=M(i,j)+min(Γ(i+1,j),Γ(i,j+1),Γ(i+1,j+1)); Step 1.3, determine the optimal alignment path: backtrack from the upper right corner of the difference matrix M to select the path with the minimum cumulative difference and calculate the total alignment cost C total : Where k represents the path point index on the optimal alignment path; Step 1.4: Select the total alignment cost C total The smallest N-segment power sequence is used as the reference power mode, and N=30 corresponds to the monthly cycle characteristics.
3. The distributed photovoltaic short-term power combination forecasting method considering extreme events according to claim 2 is characterized by: In step 2, based on the reference power mode, the alignment difference value sequence X = {x1, x2, ..., x m After sorting in ascending order, the anomaly boundary is determined by the quantile anomaly detection algorithm, which specifically includes the following steps: Step 2.
1. Calculate the median Q2 of the alignment difference value sequence X, i.e. the second quartile: Step 2.2, calculate the first quartile Q1 and the third quartile Q3 of the alignment difference value sequence X: Wherein, τ is any positive integer; Step 2.3: Define the abnormality determination interval, i.e., the extreme event characteristic data: Ω=(-∞,Q1-κI QR ]∪[Q3+κI QR ,+∞) where κ = 3, I QR = Q3 - Q1.
4. The method for distributed photovoltaic short-term power combination prediction considering extreme events according to claim 1, characterized in that: Step 3 specifically includes the following steps: Step 3.1, distribution reconstruction: The encoder of the temporal generation adversarial network combines the static features s of the temporal generation adversarial network with the temporal features x of the temporal generation adversarial network 1:T Mapping to the latent space h of the temporal generative adversarial network s ,h t , the generator G synthesizes data to meet: Among them, h s ~p(h s ),h t ~p(h t ); The sample expansion rule is: generate K sets of synthetic data for each extreme event scenario to form an enhanced data set Where D real For existing real data sets; Step 3.2, adversarial training: discriminator D distinguishes real data (s, x 1:T ) and generate data The loss function is: Step 3.3, temporal consistency constraint: supervision loss ensures the temporal dependency of the generated sequence:
5. The method for distributed photovoltaic short-term power combination prediction considering extreme events according to claim 4, characterized in that: The gradient boosting tree in step 4 specifically includes the following steps: Step 4.1.1, Histogram optimization: Discretize the continuous features into k buckets and accelerate the calculation of split gain through histogram difference: Among them, G b ,H b is the gradient estimator of bucket b, u is the regularization coefficient; G L ,H L is the gradient estimator of the child node on the left of the split point; G R ,H R is the gradient estimator of the child node to the right of the split point; Step 4.1.2: Leaf growth with depth limit: Only split the leaf with the largest gain each time, with a maximum depth limit of d max =8.
6. The distributed photovoltaic short-term power combination forecasting method considering extreme events according to claim 5 is characterized by: In step 4, the spatiotemporal attention network specifically includes the following steps: Step 4.2.1, Multi-head self-attention encoding: Use the multi-head self-attention mechanism to extract the temporal dependency of step 3.3: H attn =Concat(head1,..,head h )W O Among them, H attn is the multi-head self-attention encoding output, head i represents the output of the i-th attention head, i=1,2,...h,W O is the weight transformation matrix, Concat is the matrix concatenation function, is the query, key, and value matrix, is the projection weight, h is the number of attention heads, d k =d model / h; Step 4.2.2, gated state fusion: multi-head self-attention encoding output H attn With LSTM hidden state Fusion: Among them, W g is the gate weight matrix, which is used to control the fusion ratio of attention and LSTM hidden state, W m is the memory transformation weight matrix, used to generate new memory states; b g 、b m is the bias term.
7. The method for distributed photovoltaic short-term power combination prediction considering extreme events according to claim 6, characterized in that: In step 4, the dynamic weight smoothing mechanism is specifically as follows: v t =α×v t-1 +(1-α)×θ t Among them, v t 、v t-1 is the exponential moving average of the current moment and the previous moment, θ t is the original data at the current moment. After the above steps, the distributed photovoltaic short-term power prediction result considering extreme events is output.