A multi-dimensional wind power scenario generation method based on attention mechanism

Abstract

The present invention discloses a multi-dimensional wind power scenario generation method based on an attention mechanism. The method comprises the following steps: first, constructing a historical wind power data set, and then training a wind power prediction model with an encoder-decoder structure to obtain a trained wind power prediction model; then, using the encoder in the trained wind power prediction model to extract the compression features and feature weights corresponding to each sample in the historical wind power data set, thereby updating the historical wind power data set; finally, using the encoder to extract the compression features corresponding to the target sample, searching for the historical sample closest to the target sample in the historical wind power data set as the wind power scenario, using the scenario importance as the sample similarity, and describing the future wind power with the scenario and its importance. The present invention can effectively shorten the scenario generation time without sacrificing the probability performance, and can also describe the future wind power more accurately.

Description

Technical Field

[0001] The present invention relates to a wind power scenario generation method in the field of wind power uncertainty quantification, and specifically relates to a multi-dimensional wind power scenario generation method based on an attention mechanism. Background Art

[0002] Renewable energy such as wind power is random and intermittent. With the rapid growth of wind power installed capacity, ensuring real-time power balance has become the key to the safe operation of the power system. In order to effectively deal with the problems caused by the uncertainty of renewable energy such as wind power, it is crucial to find similar wind power scenarios based on the correlation of multi-dimensional wind power for power system decision-making.

[0003] The main challenge of scenario prediction is to model and learn the random process of wind power generation. Existing wind power scenario prediction methods can be summarized into two categories: data-driven methods and methods based on historical prediction errors.

[0004] Patent CN119204548A uses a multi-dimensional feature data-driven model to predict daily and hourly water inflows, and improves the system's scheduling efficiency by accelerating the prediction calculation efficiency. Patent CN118399749A predicts the size of future output power based on historical secondary side output power and adaptive phase shift angle range. Patent CN117713188A analyzes the relationship between meteorology and changes in photovoltaic output size at corresponding moments based on historical meteorological information and photovoltaic output power, and trains photovoltaic output on similar days through a data-driven model to achieve photovoltaic output power prediction. The above method is based on a large amount of historical data training model, and the model accuracy is good, but it has never considered the problem that the data-driven method has too much calculation, which leads to a decrease in model calculation efficiency. Similarly, the data-driven distributed photovoltaic power generation power prediction method takes into account that the data-driven method does not rely on the precise mechanism model of the research object, analyzes the relationship between input and output data from the perspective of spatial correlation, historical output data and meteorology, and believes that the data-driven model based on massive data has become the development direction of distributed photovoltaic output prediction. The model and data hybrid-driven distributed photovoltaic ultra-short-term power prediction method analyzes the impact of cloud shading on the output of neighboring distributed photovoltaic power stations, thereby realizing ultra-short-term prediction of distributed photovoltaic power stations. The wind power extreme scenario generation method based on conditional generative adversarial network proposes an extreme scenario generation framework for the power system scheduling and planning problems with uncertain wind power. The distribution of historical data sets is transferred in an iterative manner to solve the problem of insufficient extreme samples in historical data sets. The multi-time scale peak-shaving scheduling method based on wind power scenarios under data drive proposes a multi-time scale evaluation index for wind power peak-shaving power. According to this index, the wind power of one year is decomposed in the time domain and a typical wind power peak-shaving scenario is generated. The multi-source power system scenario generation and scheduling method based on limited data drive proposes a historical similar day selection method based on weighted Euclidean distance and weighted correlation, and analyzes it according to the numerical weather forecast and historical meteorological records of the day before. The above methods are based on massive data, and are more dependent on the quality and quantity of data, and the complexity of the model is also greatly improved.

[0005] A bidirectional optimization method for generating wind power time series scenarios is proposed to generate daily wind power sequence scenarios, namely, in the vertical direction: based on the historical daily wind power sequence data, the optimal reduction method is used to generate representative scenarios for each period; in the horizontal direction: the taboo search method is used to selectively connect the representative scenarios of each period, thereby forming the required daily wind power sequence representative scenarios. Considering the data set error, the multi-condition admittance acquisition method of new energy equipment based on data drive has a large deviation between the model prediction value and the true value due to the influence of noise. Therefore, the error between the two is used as an evaluation index to train the neural network to correct the prediction value, but the model relies too much on historical prediction errors, requires the data to cover a long time and be accurate, and the fluctuation of the data can easily lead to poor model effect. Summary of the invention

[0006] In order to solve the problems and needs existing in the background technology, the present invention proposes a multi-dimensional wind power scenario generation method based on an attention mechanism.

[0007] The technical solution of the present invention is as follows:

[0008] 1. A multi-dimensional wind power scenario generation method based on attention mechanism

[0009] Step 1: construct a historical wind power data set and a wind power prediction model, the wind power prediction model includes a connected encoder and decoder;

[0010] Step 2: Use the historical wind power data set to train the wind power prediction model to obtain a trained wind power prediction model;

[0011] Step 3: Use the encoder in the trained wind power prediction model to extract the compressed features corresponding to each sample in the historical wind power data set, and calculate the feature weight corresponding to each compressed feature, and add all compressed features and corresponding feature weights to the historical wind power data set, so as to update the historical wind power data set;

[0012] Step 4: The target sample is composed of numerical weather forecasts and observed powers of multiple wind farms at the target time. The compressed features corresponding to the target sample are extracted using an encoder. According to the compressed features corresponding to the target sample, several wind power scenarios and scenario importance corresponding to the current target sample are selected and generated in the current historical wind power data set.

[0013] The step 3 is specifically as follows:

[0014] Step 3.1: Input each sample in the historical wind power data set into the encoder of the trained wind power prediction model, and the encoder outputs multi-channel compressed features;

[0015] Step 3.2: Input the compressed features on different channels of the current sample into the trained decoder to obtain the reconstructed sub-features corresponding to different channels;

[0016] Step 3.3: Calculate the Wasserstein distance between the reconstructed sub-feature of each channel and the real wind power in the historical wind power dataset, and use the inverse of the normalized Wasserstein distance of each channel as the feature weight of the corresponding compressed feature of the current channel. After traversing the reconstructed sub-features of all channels, obtain the multi-channel compressed features of the current sample and the corresponding feature weights, and update the sample.

[0017] Step 3.4: Repeat steps 3.1 to 3.3, traverse and process other samples in the historical wind power dataset, obtain multi-channel compression features and corresponding feature weights of all samples, and update the historical wind power dataset.

[0018] In step 4, several wind power scenarios and scenario importances corresponding to the current target sample are selected and generated in the current historical wind power data set according to the compression features corresponding to the target sample, specifically:

[0019] The weighted distance between the compressed features corresponding to each sample in the historical wind power data set and the compressed features corresponding to the target sample is calculated respectively, and then the corresponding scene importance is generated. Several samples with the largest scene importance in the historical wind power data set are taken as the wind power scenes corresponding to the target samples.

[0020] The inverse of the normalized weighted distance corresponding to each sample is taken as its scene importance.

[0021] The weighted distance includes a weighted Euclidean distance.

[0022] 2. A multi-dimensional wind power scenario generation system based on attention mechanism

[0023] A wind power data acquisition unit, used for acquiring wind power data;

[0024] A historical wind power data set construction unit, used to form a historical wind power data set according to the acquired historical wind power data;

[0025] A wind power prediction unit, used to store the latest wind power prediction model, and to extract compression features and reconstruct sub-features corresponding to input samples using the wind power prediction model;

[0026] A feature weight calculation unit, used to calculate the feature weight corresponding to the multi-channel compression feature of each sample according to the historical wind power data set;

[0027] A model training unit, used to train the wind power prediction model in the wind power prediction unit using the historical wind power data set until a trained wind power prediction model is obtained;

[0028] The target wind power scenario generation unit is used to extract the compression features corresponding to the target sample using the encoder in the wind power prediction model, and select and generate several wind power scenarios and scenario importances corresponding to the current target sample in the current historical wind power data set according to the compression features corresponding to the target sample.

[0029] 3. A computer device

[0030] The device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of a method for generating a multi-dimensional wind power scenario based on an attention mechanism when executing the computer program.

[0031] 4. A computer-readable storage medium

[0032] The medium stores a computer program, which, when executed by a processor, implements the steps of a method for generating a multi-dimensional wind power scenario based on an attention mechanism.

[0033] 5. A computer program product

[0034] The product includes a computer program / instruction, which, when executed by a processor, implements the steps of a method for generating a multi-dimensional wind power scenario based on an attention mechanism.

[0035] Compared with the prior art, the beneficial results of the present invention are:

[0036] In order to accelerate the process of multi-dimensional scene generation, the present invention uses a trained encoder to generate compressed features containing spatial correlation of wind power to improve calculation efficiency.

[0037] The present invention uses a trained decoder to perform feature transformation, so that the dimension of each compressed feature is successfully aligned with the dimension of the output, and then quantifies the importance of the compressed feature by characterizing the correlation between each compressed feature and the output.

[0038] Compared with the method based on the prediction model of historical errors, the multi-dimensional wind power scenario generated by the method proposed in the present invention is more accurate in describing the future wind power.

[0039] Compared with the data-driven multi-dimensional scene generation method, the method proposed in the present invention can effectively shorten the scene generation time without sacrificing the probability performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 The present invention is a flow chart of the method.

[0041] Figure 2 This is a framework diagram of the method of the present invention.

[0042] Figure 3 Schematic diagram of data processing for wind power prediction model.

[0043] Figure 4 Comparison chart of prediction performance of different methods on different datasets.

[0044] Figure 5 is the predicted power per unit value of seven wind farms at a certain moment. DETAILED DESCRIPTION

[0045] The present invention is further described below with reference to the accompanying drawings and implementation examples.

[0046] The present invention proposes a multi-dimensional wind power scenario generation method based on attention mechanism, the flow chart is as follows Figure 1 and Figure 2 As shown, the specific steps include:

[0047] Step 1: Construct a historical wind power data set. Each sample includes the numerical weather forecast and historical observation power corresponding to multiple wind farms at each moment, as well as the power corresponding to multiple wind farms at future moments. The numerical weather forecast and historical observation power corresponding to multiple wind farms at each moment are used as the input of the encoder, and the power corresponding to multiple wind farms at future moments is used as the output true value. The numerical weather forecast includes wind speed, atmospheric pressure, etc. of the wind farm. And construct a wind power prediction model. The wind power prediction model includes the attention mechanism model of the encoder and decoder architecture, such as Figure 3 As shown in the figure, the latent space dimension of the attention mechanism model is set to 512, the number of layers of the decoder and encoder are both set to 6, and the number of attention heads is set to 8.

[0048] Step 2: Use the historical wind power data set to train the wind power prediction model to obtain a trained wind power prediction model;

[0049] Step 3: Use the encoder in the trained wind power prediction model to extract the compressed features corresponding to each sample in the historical wind power data set, and calculate the feature weights corresponding to each compressed feature, add all compressed features and corresponding feature weights to the historical wind power data set, and update the historical wind power data set; that is, each sample includes the numerical weather forecast and historical observed power, compressed features, feature weights corresponding to multiple wind farms at each moment, and the power corresponding to multiple wind farms at future moments. The features compressed by the encoder contain the inherent spatial dependencies between wind farms, effectively characterize the original high-dimensional information, and reduce the scene generation time, which can provide necessary reference information for the generation of multi-dimensional wind power scenes.

[0050] Step 3 is as follows:

[0051] Step 3.1: Input each sample in the historical wind power data set into the encoder of the trained wind power prediction model, and the encoder outputs multi-channel compressed features;

[0052] Step 3.2: Input the compressed features on different channels of the current sample into the trained decoder to obtain the reconstructed sub-features corresponding to different channels;

[0053] Step 3.3: Calculate the Wasserstein distance between the reconstructed sub-feature of each channel and the real wind power in the historical wind power dataset, and use the inverse of the normalized Wasserstein distance of each channel as the feature weight of the corresponding compressed feature of the current channel. After traversing the reconstructed sub-features of all channels, obtain the multi-channel compressed features of the current sample and the corresponding feature weights, and update the sample.

[0054] Step 3.4: Repeat steps 3.1 to 3.3, traverse and process other samples in the historical wind power dataset, obtain multi-channel compression features and corresponding feature weights of all samples, and update the historical wind power dataset.

[0055] Step 4: The target sample is composed of numerical weather forecasts and observed powers of multiple wind farms at the target time. The compressed features corresponding to the target sample are extracted using an encoder. According to the compressed features corresponding to the target sample, several wind power scenarios and scenario importance corresponding to the current target sample are selected and generated in the current historical wind power data set.

[0056] In step 4, several wind power scenarios and scenario importances corresponding to the current target sample are selected and generated in the current historical wind power data set according to the compression features corresponding to the target sample, specifically:

[0057] The weighted Euclidean distance between the compressed features corresponding to each sample in the historical wind power data set and the compressed features corresponding to the target sample is calculated respectively, and the inverse of the normalized weighted Euclidean distance corresponding to each sample is taken as its scene importance. Several samples with the largest scene importance in the historical wind power data set are taken as the wind power scenes corresponding to the target samples. That is, the historical samples closest to the target samples are found and used as wind power scenes. The scene importance is used to characterize the similarity of the samples, and the future wind power is described by the wind power scenes and their importance.

[0058] The prediction performance is based on the weighted energy score index S c , which is a scoring rule for evaluating the statistical performance of the model. Generally, the smaller the energy score, the better the distribution fit. The prediction performance of different methods and different data sets was evaluated, and the statistical performance was compared, such as Figure 4 As shown. The results show that the attention mechanism model adopted by the present invention has significantly improved the energy score compared with the traditional prediction model based on historical errors. In addition, compared with the data-driven model, the attention mechanism model proposed in the present invention has the operation of compressing features, which can reduce the calculation time and improve the calculation efficiency while maintaining the statistical performance comparable to other methods, or even obtain better statistical performance. This experiment proves that the method of the present invention can not only improve the calculation efficiency in practical applications, but also ensure the accuracy and stability of the prediction effect.

[0059] This embodiment compares the computational efficiency of different methods and different data sets from the perspective of model running time, as shown in Table 1. The four data sets in Table 1 are divided by segmenting the original data set into different time periods or selecting different features, and each data set represents a different time window. The results show that the average prediction time of the attention mechanism model of the present invention is 0.2585s, which is significantly less than the average prediction time of the prediction process driven by data but not through the attention mechanism method, and the feature dimensionality reduction effect is better. Therefore, feature dimensionality reduction can effectively accelerate the prediction process of the model.

[0060] Table 1 Comparison of running time of the method of the present invention and the data-driven method on different data sets

[0061]

[0062] Figure 5 It shows the predicted power per unit value of seven wind farms at a certain moment. Through observation, it can be found that this phenomenon shows that the multi-dimensional wind power scenario generation method based on the attention mechanism proposed in the present invention is relatively stable, and the generated prediction scenario accurately describes the future wind power.

[0063] The present invention also proposes a multi-dimensional wind power scenario generation system based on an attention mechanism, comprising:

[0064] A wind power data acquisition unit, used for acquiring wind power data;

[0065] A historical wind power data set construction unit, used to form a historical wind power data set according to the acquired historical wind power data;

[0066] A wind power prediction unit, used to store the latest wind power prediction model, and to extract compression features and reconstruct sub-features corresponding to input samples using the wind power prediction model;

[0067] A feature weight calculation unit, used to calculate the feature weight corresponding to the multi-channel compression feature of each sample according to the historical wind power data set;

[0068] A model training unit, used to train the wind power prediction model in the wind power prediction unit using the historical wind power data set until a trained wind power prediction model is obtained;

[0069] The target wind power scenario generation unit is used to extract the compression features corresponding to the target sample using the encoder in the wind power prediction model, and select and generate several wind power scenarios and scenario importances corresponding to the current target sample in the current historical wind power data set according to the compression features corresponding to the target sample.

[0070] The present invention also proposes a computer device, which includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of a multi-dimensional wind power scenario generation method based on an attention mechanism.

[0071] The present invention also proposes a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of a method for generating a multi-dimensional wind power scenario based on an attention mechanism are implemented.

[0072] The present invention also proposes a computer program product, which includes a computer program / instructions. When the computer program / instructions are executed by a processor, the steps of a method for generating a multi-dimensional wind power scenario based on an attention mechanism are implemented.

[0073] Finally, it should be noted that the above embodiments and explanations are only used to illustrate the technical solution of the present invention rather than to limit it. Those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope disclosed in the technical solution of the present invention, which should be included in the scope of protection of the claims of the present invention.

Claims

1. A multi-dimensional wind power scenario generation method based on attention mechanism, characterized in that: The following steps are involved: Step 1: construct a historical wind power data set and a wind power prediction model, the wind power prediction model includes a connected encoder and decoder; Step 2: Use the historical wind power data set to train the wind power prediction model to obtain a trained wind power prediction model; Step 3: Use the encoder in the trained wind power prediction model to extract the compressed features corresponding to each sample in the historical wind power data set, and calculate the feature weight corresponding to each compressed feature, and add all compressed features and corresponding feature weights to the historical wind power data set, so as to update the historical wind power data set; Step 4: The target sample is composed of numerical weather forecasts and observed powers of multiple wind farms at the target time. The compression features corresponding to the target sample are extracted using an encoder. According to the compression features corresponding to the target sample, several wind power scenarios and scenario importance corresponding to the current target sample are selected and generated in the current historical wind power data set; The step 3 is specifically as follows: Step 3.1: Input each sample in the historical wind power data set into the encoder of the trained wind power prediction model, and the encoder outputs multi-channel compressed features; Step 3.2: Input the compressed features on different channels of the current sample into the trained decoder to obtain the reconstructed sub-features corresponding to different channels; Step 3.3: Calculate the Wasserstein distance between the reconstructed sub-feature of each channel and the real wind power in the historical wind power dataset, and use the inverse of the normalized Wasserstein distance of each channel as the feature weight of the corresponding compressed feature of the current channel. After traversing the reconstructed sub-features of all channels, obtain the multi-channel compressed features of the current sample and the corresponding feature weights, and update the sample. Step 3.4: Repeat steps 3.1 to 3.3, traverse and process other samples in the historical wind power dataset, obtain multi-channel compression features and corresponding feature weights of all samples, and update the historical wind power dataset.

2. The method for generating a multi-dimensional wind power scenario based on an attention mechanism according to claim 1, characterized in that: In step 4, several wind power scenarios and scenario importances corresponding to the current target sample are selected and generated in the current historical wind power data set according to the compression features corresponding to the target sample, specifically: The weighted distance between the compressed features corresponding to each sample in the historical wind power data set and the compressed features corresponding to the target sample is calculated respectively, and then the corresponding scene importance is generated. Several samples with the largest scene importance in the historical wind power data set are taken as the wind power scenes corresponding to the target samples.

3. The method for generating a multi-dimensional wind power scenario based on an attention mechanism according to claim 2, characterized in that: The inverse of the normalized weighted distance corresponding to each sample is taken as its scene importance.

4. The method for generating a multi-dimensional wind power scenario based on an attention mechanism according to claim 2, characterized in that: The weighted distance includes a weighted Euclidean distance.

5. A multi-dimensional wind power scenario generation system based on attention mechanism, characterized in that: include: A wind power data acquisition unit, used for acquiring wind power data; A historical wind power data set construction unit, used to form a historical wind power data set according to the acquired historical wind power data; A wind power prediction unit, used to store the latest wind power prediction model, and to extract compression features and reconstruct sub-features corresponding to input samples using the wind power prediction model; In the wind power prediction unit, each sample in the historical wind power data set is first input into an encoder in the trained wind power prediction model, and the encoder outputs multi-channel compression features; Then, the compressed features on different channels of the current sample are input into the trained decoder to obtain the reconstructed sub-features corresponding to different channels. A feature weight calculation unit, used to calculate the feature weight corresponding to the multi-channel compression feature of each sample according to the historical wind power data set; In the feature weight calculation unit, the Wasserstein distance between the reconstructed sub-feature of each channel and the real wind power in the historical wind power data set is first calculated, and then the inverse of the normalized Wasserstein distance of each channel is used as the feature weight of the compressed feature corresponding to the current channel; A model training unit, used to train the wind power prediction model in the wind power prediction unit using the historical wind power data set until a trained wind power prediction model is obtained; The target wind power scenario generation unit is used to extract the compression features corresponding to the target sample using the encoder in the wind power prediction model, and select and generate several wind power scenarios and scenario importances corresponding to the current target sample in the current historical wind power data set according to the compression features corresponding to the target sample.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of a method for generating a multi-dimensional wind power scenario based on an attention mechanism as described in any one of claims 1 to 4 are implemented.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a method for generating a multi-dimensional wind power scenario based on an attention mechanism as described in any one of claims 1 to 4 are implemented.

8. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of a method for generating a multi-dimensional wind power scenario based on an attention mechanism as described in any one of claims 1 to 4 are implemented.

Images