New energy station wind-solar time series power generation method based on improved domain adversarial network

Abstract

The application discloses a new energy station wind and light time sequence power generation method and system based on an improved domain adversarial network. The historical real power data of the produced and operated station is used as a reference, and the meteorological data of the source domain and the target domain station is trained through the improved domain adversarial network to realize unsupervised domain self-adaptation. In the training process, the accuracy of the wind and light time sequence power generation is improved by continuously optimizing feature extraction. The application can adapt to meteorological power data sets of different geographical positions and generate power curves more in line with the actual wind and light output distribution law. The experimental results show that the application has high reliability and practicality in generating wind and light time sequence power and has the characteristics of strong flexibility and can adapt to the wind and light power generation task of the target domain with a small amount or lack of labels. Therefore, the new energy station wind and light time sequence power generation method based on the improved domain adversarial network has a wide application prospect in improving the wind and light energy prediction accuracy and practical application.

Description

Technical Field

[0001] This invention relates to the field of new energy technology, and in particular to a method and system for generating wind and solar time-series power at new energy power plants based on an improved domain adversarial network. Background Technology

[0002] To effectively address the energy crisis and climate change, my country has put forward the key task of vigorously developing new energy sources, comprehensively promoting the large-scale development and high-quality development of wind and solar power, and accelerating the construction of wind and solar power bases. However, wind and solar power are highly dependent on weather conditions, and their power generation is random and fluctuating. Large-scale grid connection brings many challenges to the planning and operation of new power systems. Accurately characterizing the time-series power curves of wind and solar power has become the primary key step in solving these problems.

[0003] Existing research on wind and solar time-series power curves can be broadly categorized into two types from a modeling perspective: One type is mechanism-driven, which derives physical formulas based on the fundamental principles of wind and solar power generation, inputs key meteorological information such as wind speed and radiation, and uses these formulas to obtain the power output. However, the power output simulated indirectly using mechanism-driven methods is often overly idealized, deviating from the actual power output of power plants and failing to account for complex and extreme power generation scenarios under the combined influence of multiple meteorological factors. The other type is data-driven, which analyzes historical measured power sequences from wind and solar power plants to learn the changing patterns of time-series power curves from historical data. This can be further divided into statistical and artificial intelligence methods. Deep learning models, in particular, have strong data mining capabilities and can learn to extract relevant time-series features, making them widely used in the field of new energy time-series power research. However, data-driven methods require a large amount of historical data for training, and their learning ability is limited when data is scarce or the sample size is small.

[0004] Existing research methods are applicable to power plants with complete historical data, but they become ineffective for newly built wind and solar power plants where no historical power data is available.

[0005] Prior art related to this invention

[0006] Wind and solar time-series power curves calculated based on physical methods [1,2] The specific process is as follows: Based on the basic principles of wind and solar power generation, the measured or predicted wind speed and radiation are input into the physical formula, and the power curve of the unit is converted into wind and solar power output.

[0007] Disadvantages of existing technology 1

[0008] 1. The wind and solar power calculated based on physical methods is based on fundamental principles and empirical models, and the results have certain errors compared with the actual power generation of the power plants.

[0009] 2. The wind and solar power calculated based on the physical method does not take into account the comprehensive impact of other meteorological factors, such as wind direction, humidity, temperature, and precipitation, on wind and solar power generation.

[0010] Prior art related to this invention

[0011] Wind and solar time-series power curves obtained based on model fine-tuning method [3] The specific process is as follows: a mapping model from input meteorological data to output wind and solar time-series power is trained on a source domain wind and solar power station with high-quality meteorological power data. After training, the optimal parameters of the pre-trained model are saved. The pre-trained model is then loaded into a newly built wind and solar target domain station and the network structure is fine-tuned. This allows the knowledge learned from the source domain task to be transferred to the target domain task, thereby accelerating the training process of the target domain task and improving its performance.

[0012] Disadvantages of existing technology 2

[0013] 1. Model fine-tuning methods are not suitable for unsupervised adaptive models because pre-training and fine-tuning methods require supervised data in subsequent tasks to achieve good performance. For newly established wind and solar power target sites containing unsupervised meteorological data, the effectiveness of model fine-tuning methods cannot be guaranteed.

[0014] 2. Model fine-tuning methods have limited transferability. Pre-trained models may perform well on some tasks but may not generalize well on others. Specific tasks require more fine-tuning or the addition of specific structures to improve performance. Therefore, pre-trained models may require additional adjustments and optimizations when applied to new target domains.

[0015] 3. Pre-trained models typically require significant computational resources and time for training, especially for large-scale models and datasets. Additional training time and resources are also needed during the fine-tuning phase. This limits the scalability of widely adopting pre-trained models and fine-tuning methods. Summary of the Invention

[0016] This invention addresses the shortcomings of existing technologies by providing a method for generating time-series wind and solar power at renewable energy power plants based on an improved domain adversarial network. It also considers the migration issues of multi-source meteorological data and combined wind and solar power generation, offering valuable insights for the efficient development and construction of wind and solar power plants.

[0017] To achieve the above-mentioned objectives, the technical solution adopted by the present invention is as follows:

[0018] A method for generating time-series wind and solar power at renewable energy power plants based on an improved domain adversarial network, characterized in that: the improved domain adversarial network includes: a feature extractor and a wind and solar power generator; the method includes the following steps:

[0019] The wind and solar power data in the source and target domains are normalized; the normalized data is then used for subsequent training.

[0020] The feature extractor is trained using normalized data; the feature extractor consists of a channel attention mechanism module and a nonlinear module; the maximum mean difference is used as the loss function during training.

[0021] In the feature extractor, a channel attention mechanism module is added; this module adjusts the importance of channels by learning the dynamic, non-linear dependencies of channel weights to extract more representative feature representations.

[0022] The wind and solar power generator is trained using feature representations obtained through channel attention and wind and solar power data from the source and target domains. The wind and solar power generator employs a bidirectional long short-term memory network to model the temporal relationship between wind and solar power and other variables. Mean squared error is used as the loss function during training.

[0023] The feature representations extracted by the feature extractor are used to train a domain discriminator using wind and solar power data from the source and target domains. The domain discriminator employs a network structure composed of fully connected layers to distinguish features between the source and target domains. During training, the cross-entropy loss function is used to measure the classification performance of the domain discriminator.

[0024] By inputting wind and solar power data of the target domain and feature representations extracted by a feature extractor, and combining them with a wind and solar generator, the time-series wind and solar power of the target domain is generated.

[0025] Furthermore, in the data preprocessing, historical meteorological observation data of normally operating wind and solar power plants and corresponding wind and solar power generation data for the same period are considered as labeled source domain D. S The regional meteorological data of newly built wind and solar power plants will be used as the unlabeled target domain D. T That is, given a sample size of N S Includes historical meteorological data x s and power data y s The source domain and a sample size of N. T Includes only meteorological data from the Shago Desert. t The target domain.

[0026]

[0027]

[0028] Furthermore, the channel attention mechanism module consists of two fully connected layers. The first fully connected layer reduces the dimensionality and uses the ReLU activation function, while the second fully connected layer restores the dimensionality of the weight factors to the original dimensionality. The normalized weight factors are multiplied by the input to calculate the final feature representation of the channel.

[0029] Furthermore, the wind and solar power generator performs nonlinear mapping on the extracted features, and simultaneously generates wind and solar power curves for comparison and correction with the actual power curves, using a mean squared error loss function:

[0030]

[0031] Where L s Let θ be the loss function of the wind and light generator. f ,θ s These are the network parameters for the feature extractor and the landscape generator, respectively. To generate wind and solar power in the source domain, y s This represents the actual wind and solar power in the source region.

[0032] Furthermore, the domain discriminator is a binary classifier used to distinguish whether the input features belong to the source domain or the target domain. The output discriminant value is used to calculate the cross-entropy loss function with the domain classification label:

[0033]

[0034] Where L d Let θ be the loss function of the neighborhood discriminator. f ,θ d These are the network parameters for the feature extractor and the neighborhood discriminator, respectively. This represents the discriminant value output by the neighborhood discriminator. The source domain label is represented as a vector of all zeros. The target domain label is represented by a vector of all 1s.

[0035] Furthermore, during the training of the feature extractor, the maximum mean difference is added to measure the similarity of the feature distributions of the source and target domains in the feature space, and this is used as a loss function to represent the fusion effect of the extracted features:

[0036]

[0037] Where L m Let θ be the maximum mean difference loss function. f F represents the network parameters of the feature extractor. s F t Let represent the high-dimensional isomorphic feature vectors extracted from the source and target domains, respectively; φ(·) represents mapping the feature vectors to the reproducing kernel Hilbert space RKHS; after square expansion, the inner product of RKHS is transformed into the kernel function k(·,·).

[0038] Furthermore, the overall optimization objective of the improved domain adversarial network consists of three terms, as shown in the following formula:

[0039]

[0040] Where L is the loss function of the improved domain adversarial network model, and λ m ,λ d To balance the weights of the loss function, set them as follows:

[0041]

[0042]

[0043] In the formula, γ is a coefficient, taken as a constant of 10; k represents the model training process, which is the ratio of the current training iterations to the total training iterations, k∈(0,1]; λ represents the training process throughout the entire training process. m Initialize at point 1 and gradually decrease λ. d It starts from 0 and keeps growing until it reaches 1.

[0044] Furthermore, the feature extractor consists of a channel attention mechanism module and a nonlinear module; the channel attention mechanism focuses on the feature dimension of meteorological elements, using U = [u1, u2, ..., u c ], H, W, and C represent the input quantities, and H, W, and C represent the dimensions of the input quantities. The channel attention mechanism requires two steps:

[0045] The first step is compression, which uses global average pooling to compress the global information of the spatial dimension H×W into the channel response to obtain the channel output z with global characteristics;

[0046]

[0047] Where z c F is the output value after compression. sq Indicates a squeezing operation;

[0048] The second step involves activation, learning the dynamic and non-linear dependencies between channels, and outputting channel weights s. Two fully connected layers are constructed: the first fully connected layer performs dimensionality reduction, selecting ReLU as the activation function; the second fully connected layer restores the reduced weight factor dimensionality to its original dimension, as shown in the following equation:

[0049]

[0050] Where s c F is the output value after excitation. exσ(·) represents the activation operation, W represents the fully connected layer parameters, W1 represents the first fully connected layer parameter, W2 represents the second fully connected layer parameter, σ(·) represents the Sigmoid activation function, and r is the dimensionality reduction ratio.

[0051] The normalized weighting factors are multiplied by the corresponding input values ​​to calculate the final channel features. The channel attention mechanism module uses global meteorological data to learn channel weight changes and adaptively adjusts channel importance, as shown in the following formula:

[0052]

[0053] in For the final output value of the channel attention mechanism module, F scale This indicates a product operation.

[0054] Add a linear layer and a LeakyReLU activation function after the channel attention mechanism module.

[0055] Furthermore, the wind and light generator is composed of a bidirectional long short-term memory network. The memory unit of the long short-term memory network includes an input gate, a forget gate, and an output gate. Assume that the state of the memory unit at time t is c. t The input gate is i t The forget gate is f t The output gate is 0 t The final output of the LSTM unit is h. t The calculation formulas for each variable are as follows:

[0056] i t =σ(W i [h t-1 ,F (t) ]+b i (12)

[0057] f t =σ(W f [h t-1 ,F (t) ]+b f (13)

[0058]

[0059] o t =σ(W o [h t-1 ,F (t) ]+b o (15)

[0060]

[0061] In the formula h t-1c is the output of the Long Short-Term Memory network at time t-1. t-1 F represents the state of the Long Short-Term Memory (LSTM) network memory unit at time t-1. (t) The feature sequence output by the feature extractor, tanh(·) denotes the tanh activation function, W c W i W f W o The weight matrices corresponding to the memory unit, input gate, forget gate, and output gate, respectively, b c ,b i ,b f ,b o These are the bias vectors corresponding to the memory unit, input gate, forget gate, and output gate, respectively.

[0062] Bidirectional Long Short-Term Memory (BSSM) networks add a data flow of inverse BSSM to the forward BSSM network. The hidden states of the bidirectional BSSM network are independent of each other, enabling it to comprehensively learn the temporal patterns and trends hidden in the input sequence. The bidirectional BSSM network computes the hidden layers forward at each time step. and backward computation of hidden layers Final output power after splicing

[0063]

[0064]

[0065]

[0066] In the formula Let t be the output of the forward Long Short-Term Memory network. This indicates a positive long short-term memory network. Let t be the output of the reverse long short-term memory network. This represents a reverse long short-term memory network. This represents the weight matrix of a forward long short-term memory network. Let b represent the weight matrix of the inverse long short-term memory network. y This represents the bias vector of a bidirectional long short-term memory network.

[0067] Furthermore, the domain discriminator has three fully connected layers for dimensionality reduction of the input features; a normalization layer is added after the first two fully connected layers, and the LeakyReLU activation function is used for non-linear pattern recognition and classification; the sigmoid activation function is used in the last fully connected layer to map the output to probability values ​​between 0 and 1, which matches the definition of the cross-entropy loss function.

[0068] The present invention also discloses a wind and solar time-series power generation system for new energy power plants based on an improved domain adversarial network, which can be used to implement the above-mentioned wind and solar time-series power generation method for new energy power plants;

[0069] The wind and solar time-series power generation system for new energy power plants includes:

[0070] Data preprocessing module: used to normalize the wind and solar power data of the source and target domains, and to divide the data into training and testing sets;

[0071] Feature extractor training module: Trains the feature extractor using wind and solar field data from the source and target domains;

[0072] Channel attention mechanism modeling module: Add a channel attention mechanism module after the feature extractor to adjust the importance of channels by learning the dynamic and non-linear dependencies of channel weights;

[0073] Wind and solar power generator training module: The wind and solar power generator is trained using source domain wind and solar power data and feature representations obtained through channel attention mechanism. A bidirectional long short-term memory network structure is adopted to model the temporal relationship between wind and solar power and other variables.

[0074] Domain Discriminator Training Module: The feature representation extracted by the feature extractor is used to train the domain discriminator, which is used to distinguish the features of the source domain and the target domain;

[0075] Wind and solar time-series power generation module: By inputting wind and solar power data of the target domain and feature representations extracted by the feature extractor, and combining with the wind and solar generator, the module generates the wind and solar time-series power of the target domain.

[0076] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-mentioned method for generating wind and solar time-series power at new energy power plants.

[0077] The present invention also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method for generating wind and solar time-series power at new energy power plants.

[0078] Compared with the prior art, the advantages of the present invention are as follows:

[0079] This invention considers various meteorological factors in different regions of existing and planned wind and solar power plants, and constructs a feature extractor to learn and mine features from multi-source meteorological data. This invention uses historical real power data from existing wind and solar power plants as labels for the wind and solar power time-series curves generated by the wind and solar power generator. That is, it compares and corrects the actual power generation of the wind and solar power plants, resulting in wind and solar power time-series curves of higher quality than those calculated using physical formulas.

[0080] The improved domain adversarial network established in this invention is highly suitable for unsupervised domain adaptation research. It requires no labeling information from the target station, only meteorological data from both the source and target domain stations for training, thus reducing the workload of data labeling. Furthermore, the improved domain adversarial network continuously optimizes feature extraction to obtain more accurate feature representations, improving generation accuracy and broad applicability.

[0081] This invention takes newly built wind and solar power plants without historical power data as target power plants and existing wind and solar power plants with complete historical data as source power plants. By improving the domain adversarial network model to learn the nonlinear mapping between input meteorological information and output power generation, and increasing the maximum mean difference to help improve the feature transfer effect, the time-series power data of the target power plants that conforms to the actual wind and solar power output distribution pattern is generated.

[0082] In summary, the advantages of this invention are:

[0083] 1. High-quality wind and solar time-series power generation: Using historical real power data from already operational power plants as a reference, the generated wind and solar time-series power curves are of higher quality and more accurate than those calculated using traditional physical formulas.

[0084] 2. Unsupervised domain adaptation: The system does not require any label information from the target station, and only uses meteorological data from the source and target domain stations for training, which reduces the workload of data labeling.

[0085] 3. Continuous optimization of feature extraction: Improved domain adversarial networks continuously optimize feature extraction to obtain more accurate feature representations, thereby improving generation accuracy.

[0086] 4. Wide range of applications: Applicable to meteorological power datasets in different geographical locations, it can be used to study the relationship between wind and solar power characteristics and meteorological information.

[0087] 5. Provides reliability and practicality: The generated wind and solar time-series power curves are more consistent with the actual wind and solar power output distribution, and have higher reliability and practicality.

[0088] 6. High flexibility: It is suitable for regions with a small number or lack of new energy power generation labels and can perform target domain migration generation. Attached Figure Description

[0089] Figure 1 This is a schematic diagram of the improved domain adversarial network model structure according to an embodiment of the present invention;

[0090] Figure 2 This is a schematic diagram of the feature extractor network structure according to an embodiment of the present invention;

[0091] Figure 3This is a comparison chart of the photovoltaic power generation curve (a) and the physical formula calculation curve (b) of a certain power station in Qinghai during summer according to an embodiment of the present invention. Detailed Implementation

[0092] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and examples.

[0093] This embodiment establishes an improved domain adversarial network model for generating time-series power data from newly constructed wind and solar power plants. The model structure comprises three parts: a feature extractor, a neighborhood discriminator, and a wind / solar power generator. Figure 1 As shown, where G f θ f G s θ s G d θ d These represent the feature extractor, landscape generator, and neighborhood discriminator, respectively, along with their parameters; L s L represents the loss function of the wind and solar generator; d L represents the loss function of the neighborhood discriminator. m This represents the maximum mean difference loss function.

[0094] The first step is to treat the historical meteorological observation data of normally operating wind and solar power plants and the corresponding wind and solar power generation data during the same period as labeled source domain D. S The regional meteorological data of newly built wind and solar power plants will be used as the unlabeled target domain D. T That is, given a sample size of N. S Includes historical meteorological data x s and power data y s The source domain and a sample size of N. T Includes only meteorological data from the Shago Desert. t The target domain.

[0095]

[0096]

[0097] The source and target domain meteorological datasets constitute the sample space X. Inputting these datasets into the feature extractor yields source and target domain features. The wind and solar power generator and the neighborhood discriminator share the extracted features. The wind and solar power generator performs a non-linear mapping on the extracted features and simultaneously generates wind and solar power curves for comparison and correction with the actual power curves, employing the Mean Squared Loss (MSE) loss function.

[0098]

[0099] in To generate wind and solar power in the source domain, y s This represents the true wind and solar power in the source domain. The domain discriminator is a binary classifier used to distinguish whether the input features belong to the source or target domain. The output discriminant value is used to calculate the cross-entropy loss function with the domain classification label:

[0100]

[0101] in This represents the discriminant value output by the neighborhood discriminator. The source domain label is represented as a vector of all zeros. The target domain label is represented by a vector of all one values. The adversarial approach involves the domain discriminator learning to correctly distinguish feature sources, while the feature extractor learns domain-adaptive feature representations that improve generation accuracy to confuse the domain discriminator. The feature extractor and the domain discriminator are connected by a Gradient Reversal Layer (GRL) to satisfy their diametrically opposed optimization objectives during the game.

[0102] The Maximum Mean Discrepancy (MMD) is added to measure the similarity of feature distributions between the source and target domains in the feature space, and it is used as a loss function to represent the fusion effect of the extracted features.

[0103]

[0104] Where F s F t Let represent the high-dimensional isomorphic feature vectors extracted from the source and target domains, respectively. φ(·) represents mapping the feature vectors to the reproducing kernel Hilbert space (RKHS). After square expansion, the RKHS inner product is transformed into the kernel function k(·,·) for easier computation. This demonstrates that the choice of kernel function significantly impacts the measurement of the maximum mean difference. Here, a mixture kernel function constructed from five Gaussian kernel functions is used for calculation.

[0105] In summary, the overall optimization objective of the improved domain adversarial network model consists of three factors:

[0106]

[0107] Where the loss function balance weight coefficient λ m ,λ d The settings are as follows:

[0108]

[0109]

[0110] In the formula, γ is a coefficient, taken as a constant of 10; k represents the model training progress, which is the ratio of the current training iterations to the total number of training iterations, k∈(0,1). Throughout the training process, λ... m Initialize at point 1 and gradually decrease λ. d Starting from 0 and gradually increasing to 1, this setup ensures that the model primarily relies on the maximum mean difference loss function to supervise the feature extractor's learning behavior in the early stages of training. This not only quantifies the difference in feature distributions between the source and target domains but also narrows the distance between their feature distributions. As the feature distributions between the domains converge, the performance of the domain discriminator becomes more important, strengthening its classification ability and enabling the model to achieve better generalization capabilities across different data distributions.

[0111] Given that each component module in the improved domain adversarial network model has a different role and function, the model structure is rationally designed according to the characteristics of various neural network applications. The source / target domain meteorological samples include elements such as wind speed, temperature, precipitation, direct radiation, diffuse radiation, air density, cloud cover, and snowfall at different altitudes. Let the total sample size be N, the time resolution be a hours, and the number of meteorological element types be C. To enhance the expressive power of meteorological features, the feature extractor consists of a channel attention mechanism module (SqueezeandExcitation Networks, SENet) and a nonlinear module. The channel attention mechanism focuses on the feature dimension of meteorological elements, using U = [u1, u2, ..., u...]. c ], Let W = 24 / a and H = Na / 24 represent the input, where W = 24 / a and H = Na / 24. The channel attention mechanism mainly performs two steps. The first step is squeezing, which uses global average pooling (GAP) to compress the spatial dimension H×W global information into the channel response, obtaining a channel output z with global features. Compared to the local operations of convolution, it can accept a wider range of perspectives.

[0112]

[0113] The second step is excitation, which flexibly learns the dynamic and non-linear dependencies between channels and outputs channel weights s. Two fully connected layers are constructed. The first fully connected layer performs dimensionality reduction, selecting ReLU as the corresponding activation function. The second fully connected layer restores the reduced weight factor dimensionality to its original dimension.

[0114]

[0115] Where σ(·) represents the Sigmoid activation function, and r is the dimensionality reduction ratio. The normalized weight factors are multiplied by the corresponding inputs to calculate the final channel features. The channel attention mechanism uses global information from meteorological data to learn channel weight changes to adaptively adjust channel importance, boosting key features and suppressing features that are not useful for the current task.

[0116]

[0117] Adding a linear layer and a LeakyReLU activation function after the channel attention mechanism module allows the obtained feature representation to not only adapt to the importance weights of each channel but also to be better embedded into subsequent neural networks. The complete network structure of the feature extractor is as follows: Figure 2 As shown.

[0118] The wind and solar power generator is constructed using a bidirectional long short-term memory (BiLSTM) network, which offers significant advantages in processing time-series power data for wind and solar power. A long short-term memory network is a recurrent neural network with both long and short-term information memory. It can model and analyze the temporal relationships between wind and solar power and other variables, capturing the temporal dependencies between relevant features and power, thereby revealing their potential correlations. The memory unit of the long short-term memory network includes an input gate, a forget gate, and an output gate. Assuming the state of the memory unit at time t is c... t The input gate is i t The forget gate is f t The output gate is 0 t The final output of the LSTM unit is h. t The calculation formulas for each variable are as follows:

[0119] i t =σ(W i [h t-1 ,F (t) ]+b i (12)

[0120] f t =σ(W f [h t-1 ,F (t) ]+b f (13)

[0121]

[0122] o t =σ(W o [h t-1 ,F(t) ]+b o (15)

[0123]

[0124] In the formula h t-1 For the hidden state of the memory unit in the Long Short-Term Memory network at time t-1, F (t) is the feature sequence output by the feature extractor, W corresponds to the weight matrix of different gates, and b corresponds to the bias vector.

[0125] Bidirectional Long Short-Term Memory (BSSM) networks add a reverse BSSM data flow to the forward BSSM network. The hidden states of the bidirectional BSSM network are independent of each other, allowing for comprehensive learning of the temporal patterns and trends hidden in the input sequence. The bidirectional BSSM network computes the hidden layers forward at each time step. and backward computation of hidden layers Final output power after splicing

[0126]

[0127]

[0128]

[0129] The domain discriminator has three fully connected layers for dimensionality reduction of the input features. To reduce the risk of overfitting, a normalization layer is added after the first two fully connected layers, and a LeakyReLU activation function is used for non-linear pattern recognition and classification. The final fully connected layer uses a Sigmoid activation function to map the output to probability values ​​between 0 and 1, matching the definition of the cross-entropy loss function.

[0130] Finally, the effectiveness of this invention is verified based on two real scenic spots in Province A and four real scenic spots in Province B, comparing the performance of the three methods on different datasets. The first method is the pre-training and fine-tuning method mentioned earlier, which designs a network with the same structure as the feature extractor and scenic spot generator. After pre-training in the source domain, the parameters are fixed, and only the final fully connected layer of the network is fine-tuned in the target domain. The second method is the original domain adversarial network model. The third method is the improved domain adversarial network model proposed in this invention. Furthermore, Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) are used as indicators to measure the accuracy of the generated power curve, calculated using the following formulas:

[0131]

[0132]

[0133] Table 1 shows the accuracy comparison of the three models tested on different datasets.

[0134] Table 1 Test results of migration between different wind and solar power stations

[0135]

[0136] Analysis of the test results shows that the average absolute error and root mean square error of the wind and solar time-series power generation method for new energy power plants based on an improved domain adversarial network, as proposed in this embodiment of the invention, are improved by an average of 22.67% and 21.20%, respectively. Compared with the domain adversarial network, the improvements are 9.84% and 8.50%, respectively. The maximum accuracy improvement of 13.37% and 12.84% was achieved in the migration example from wind and solar power plant No. 1 in Province A to No. 1 in Province B. The test results demonstrate the advantages of combining the domain adversarial training method with the maximum mean difference function. That is, the method proposed in this invention can adaptively extract domain-invariant features from meteorological datasets from different sources and associate them with the corresponding wind and solar power, achieving end-to-end learning of input meteorological data and output power. It also proves the generalization ability of the improved domain adversarial network in the target domain, thus making it better applicable to feature transfer tasks.

[0137] The photovoltaic time-series power curve obtained by the method of this invention is compared with the photovoltaic time-series power curve simulated by the physical method, such as... Figure 3 As shown, the power curves calculated by physical formulas are often smooth and patterned, considering only the influence of solar radiation intensity while ignoring the combined effects of other meteorological factors. The photovoltaic daily power curve generated using the model of this invention exhibits significant volatility and marked uncertainty in its intraday variation trend, thus more accurately reflecting power fluctuations under complex actual meteorological conditions.

[0138] In another embodiment of the present invention, a wind and solar time-series power generation system for new energy power plants based on an improved domain adversarial network is provided. This system can be used to implement the above-described wind and solar time-series power generation method for new energy power plants, specifically including:

[0139] Data preprocessing module: This module is used to normalize the wind and solar power data in the source and target domains and divide the data into training and testing sets.

[0140] Feature Extractor Training Module: The feature extractor is trained using wind and solar power data from the source domain, and spatial and temporal features of the wind and solar power data are extracted through a convolutional neural network structure.

[0141] Channel attention mechanism modeling module: Add a channel attention mechanism module after the feature extractor to adjust the importance of channels by learning the dynamic and non-linear dependencies of channel weights.

[0142] Wind and solar power generator training module: The wind and solar power generator is trained using source domain wind and solar power data and feature representations obtained through channel attention mechanism. A bidirectional long short-term memory network structure is adopted to model the temporal relationship between wind and solar power and other variables.

[0143] Domain Discriminator Training Module: The domain discriminator is trained using wind and solar power data from the source and target domains, as well as feature representations extracted by the feature extractor, to distinguish features between the source and target domains.

[0144] Model optimization module: Regularization is performed using batch normalization and Dropout techniques, and model optimization is performed using residual connections and L2 regularization to improve the performance of the feature extractor and the neighborhood discriminator.

[0145] Testing and Evaluation Module: Uses a test set to evaluate the generation effect of the model. The quality of the generation effect is measured by calculating the mean square error and correlation coefficient between the generated wind and solar power curves and the real wind and solar power data in the target domain.

[0146] Wind and solar time-series power generation module: By inputting wind and solar power data of the target domain and feature representations extracted by the feature extractor, and combining with the wind and solar generator, the module generates the wind and solar time-series power of the target domain.

[0147] In another embodiment of the present invention, a terminal device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used in the operation of a wind and solar time-series power generation method for new energy power plants, including the following steps:

[0148] 1) Data preprocessing: First, the wind and solar power data in the source and target domains are normalized to a value range between 0 and 1. Then, the data is divided into training and testing sets for model training and evaluation.

[0149] 2) Feature Extractor Training: The feature extractor is trained using wind and solar power data from the source domain. The feature extractor employs a convolutional neural network (CNN) structure, extracting spatial and temporal features of the wind and solar power data through convolutional and pooling layers. During training, Maximum Mean Discrepancy (MMD) is used as the loss function to ensure that the feature representations of the source and target domains are as close as possible.

[0150] 3) Modeling the Channel Attention Mechanism: A channel attention mechanism module is added to the feature extractor. This mechanism adjusts channel importance by learning the dynamic, non-linear dependencies of channel weights. This module consists of two fully connected layers. The first layer reduces dimensionality and uses the ReLU activation function, while the second layer restores the dimensionality of the weight factors to their original dimensions. The normalized weight factors are multiplied by the input to calculate the final feature representation of each channel.

[0151] 4) Training the wind and solar power generator: The wind and solar power generator is trained using source domain wind and solar power data and feature representations obtained through a channel attention mechanism. The generator employs a bidirectional long short-term memory (BiLSTM) network to model the temporal relationship between wind and solar power and other variables. During training, mean squared error (MSE) is used as the loss function to ensure that the generated wind and solar power curves approximate the real wind and solar power data in the target domain.

[0152] 5) Training the Domain Discriminator: The domain discriminator is trained using wind and solar power data from the source and target domains, along with feature representations extracted by a feature extractor. The domain discriminator employs a network structure composed of fully connected layers to distinguish features between the source and target domains. The cross-entropy loss function is used during training to measure the classification performance of the domain discriminator.

[0153] 6) Overall Model Optimization: To improve the performance of the feature extractor and neighborhood discriminator, batch normalization (BN) and dropout techniques are used for regularization. To address the vanishing gradient and overfitting issues, residual connections and L2 regularization are used for model optimization.

[0154] 7) Testing and Evaluation: A test set is used to evaluate the model's generation performance. The quality of the generation is measured by calculating the mean square error (MSE) and correlation coefficient between the generated wind and solar power curves and the actual wind and solar power data in the target domain.

[0155] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory). This computer-readable storage medium is a memory device in a terminal device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and extended storage media supported by the terminal device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device.

[0156] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the wind and solar time-series power generation method for new energy power plants in the above embodiments; one or more instructions in the computer-readable storage medium are loaded and executed by the processor in the following steps:

[0157] 1) Data preprocessing: First, the wind and solar power data in the source and target domains are normalized to a value range between 0 and 1. Then, the data is divided into training and testing sets for model training and evaluation.

[0158] 2) Feature Extractor Training: The feature extractor is trained using wind and solar power data from the source domain. The feature extractor employs a convolutional neural network (CNN) structure, extracting spatial and temporal features of the wind and solar power data through convolutional and pooling layers. During training, Maximum Mean Discrepancy (MMD) is used as the loss function to ensure that the feature representations of the source and target domains are as close as possible.

[0159] 3) Modeling the Channel Attention Mechanism: A channel attention mechanism module is added to the feature extractor. This mechanism adjusts channel importance by learning the dynamic, non-linear dependencies of channel weights. This module consists of two fully connected layers. The first layer reduces dimensionality and uses the ReLU activation function, while the second layer restores the dimensionality of the weight factors to their original dimensions. The normalized weight factors are multiplied by the input to calculate the final feature representation of each channel.

[0160] 4) Training the wind and solar power generator: The wind and solar power generator is trained using source domain wind and solar power data and feature representations obtained through a channel attention mechanism. The generator employs a bidirectional long short-term memory (BiLSTM) network to model the temporal relationship between wind and solar power and other variables. During training, mean squared error (MSE) is used as the loss function to ensure that the generated wind and solar power curves approximate the real wind and solar power data in the target domain.

[0161] 5) Training the Domain Discriminator: The domain discriminator is trained using wind and solar power data from the source and target domains, along with feature representations extracted by a feature extractor. The domain discriminator employs a network structure composed of fully connected layers to distinguish features between the source and target domains. The cross-entropy loss function is used during training to measure the classification performance of the domain discriminator.

[0162] 6) Overall Model Optimization: To improve the performance of the feature extractor and neighborhood discriminator, batch normalization (BN) and dropout techniques are used for regularization. To address the vanishing gradient and overfitting issues, residual connections and L2 regularization are used for model optimization.

[0163] 7) Testing and Evaluation: A test set is used to evaluate the model's generation performance. The quality of the generation is measured by calculating the mean square error (MSE) and correlation coefficient between the generated wind and solar power curves and the actual wind and solar power data in the target domain.

[0164] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0165] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0166] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0167] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0168] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the implementation methods of the present invention, and should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of the present invention.

Claims

1. A method for generating time-series wind and solar power at renewable energy power plants based on an improved domain adversarial network, characterized in that: The improved domain adversarial network includes a feature extractor and a landscape generator; the method includes the following steps: The wind and solar power data in the source and target domains are normalized; the normalized data is then used for subsequent training. The feature extractor is trained using normalized data; the feature extractor consists of a channel attention mechanism module and a nonlinear module; the maximum mean difference is used as the loss function during training. In the feature extractor, a channel attention mechanism module is added; the channel attention mechanism module adjusts the importance of channels by learning the dynamic, non-linear dependencies of channel weights. The wind and solar power generator is trained using feature representations obtained through channel attention and wind and solar power data from the source and target domains. The wind and solar power generator employs a bidirectional long short-term memory network to model the temporal relationship between wind and solar power and other variables. Mean squared error is used as the loss function during training. The feature representations extracted by the feature extractor are used to train a domain discriminator using wind and solar power data from the source and target domains. The domain discriminator employs a network structure composed of fully connected layers to distinguish features between the source and target domains. During training, the cross-entropy loss function is used to measure the classification performance of the domain discriminator. By inputting wind and solar power data of the target domain and feature representations extracted by a feature extractor, and combining them with a wind and solar generator, the time-series wind and solar power of the target domain is generated.

2. The method for generating wind and solar time-series power at new energy power plants according to claim 1, characterized in that: In data preprocessing, historical meteorological observation data of normally operating wind and solar power plants and corresponding wind and solar power generation data for the same period are considered as labeled source domain D. S The regional meteorological data of newly built wind and solar power plants will be used as the unlabeled target domain D. T That is, given a sample size of N S Includes historical meteorological data x s and power data y s The source domain and a sample size of N. T Includes only meteorological data from the Shago Desert. t The target domain; 3. The method for generating wind and solar time-series power at new energy power plants according to claim 1, characterized in that: The channel attention mechanism module consists of two fully connected layers. The first fully connected layer reduces the dimensionality and uses the ReLU activation function. The second fully connected layer restores the dimensionality of the weight factors after dimensionality reduction to the original dimensionality. The normalized weight factors are multiplied by the input to calculate the final feature representation of the channel.

4. The method for generating wind and solar time-series power at new energy power plants according to claim 1, characterized in that: The wind and solar power generator performs nonlinear mapping on the extracted features and simultaneously generates wind and solar power curves for comparison and correction with the actual power curves, using a mean squared error loss function. Where L s Let θ be the loss function of the wind and light generator. f ,θ s These are the network parameters for the feature extractor and the landscape generator, respectively. To generate wind and solar power in the source domain, y s This represents the actual wind and solar power in the source region.

5. The method for generating wind and solar time-series power at new energy power plants according to claim 1, characterized in that: The domain discriminator is a binary classifier used to distinguish whether input features belong to the source domain or the target domain. The output discriminant value is compared with the domain classification label to calculate the cross-entropy loss function. Where L d Let θ be the loss function of the neighborhood discriminator. f ,θ d These are the network parameters for the feature extractor and the neighborhood discriminator, respectively. This represents the discriminant value output by the neighborhood discriminator. The source domain label is represented as a vector of all zeros. The target domain label is represented by a vector of all 1s.

6. The method for generating wind and solar time-series power at new energy power plants according to claim 1, characterized in that: In the training of the feature extractor, the maximum mean difference is added to measure the similarity of the feature distributions of the source and target domains in the feature space, and it is used as a loss function to represent the fusion effect of the extracted features: Where L m Let θ be the maximum mean difference loss function. f F represents the network parameters of the feature extractor. s F t Let represent the high-dimensional isomorphic feature vectors extracted from the source and target domains, respectively; φ(·) represents mapping the feature vectors to the reproducing kernel Hilbert space RKHS; after square expansion, the inner product of RKHS is transformed into the kernel function k(·,·).

7. The method for generating wind and solar time-series power at new energy power plants according to claim 1, characterized in that: The overall optimization objective of the improved domain adversarial network consists of three terms, as shown in the following formula: Where L is the loss function of the improved domain adversarial network model, and λ m ,λ d To balance the weights of the loss function, set them as follows: In the formula, γ is a coefficient, taken as a constant of 10; k represents the model training process, which is the ratio of the current training iterations to the total training iterations, k∈(0,1]; λ represents the training process throughout the entire training process. m Initialize at point 1 and gradually decrease λ. d It starts from 0 and keeps growing until it reaches 1.

8. The method for generating wind and solar time-series power at new energy power plants according to claim 1, characterized in that: The feature extractor consists of a channel attention mechanism module and a nonlinear module; the channel attention mechanism focuses on the feature dimensions of meteorological elements, using... H, W, and C represent the input quantities, and H, W, and C represent the dimensions of the input quantities. The channel attention mechanism requires two steps: The first step is compression, which uses global average pooling to compress the global information of the spatial dimension H×W into the channel response to obtain the channel output z with global characteristics; Where z c F is the output value after compression. sq Indicates a squeezing operation; The second step involves activation, learning the dynamic and non-linear dependencies between channels, and outputting channel weights s. Two fully connected layers are constructed: the first fully connected layer performs dimensionality reduction, selecting ReLU as the activation function; the second fully connected layer restores the reduced weight factor dimensions to their original dimensions, as shown in the following equation: Where s c F is the output value after excitation. ex σ(·) represents the activation operation, W represents the fully connected layer parameters, W1 represents the first fully connected layer parameter, W2 represents the second fully connected layer parameter, σ(·) represents the Sigmoid activation function, and r is the dimensionality reduction ratio. The normalized weighting factors are multiplied by the corresponding input values ​​to calculate the final channel features. The channel attention mechanism module uses global meteorological data to learn channel weight changes and adaptively adjusts channel importance, as shown in the following formula: in For the final output value of the channel attention mechanism module, F scale Indicates the product operation; Add a linear layer and a LeakyReLU activation function after the channel attention mechanism module.

9. The method for generating wind and solar time-series power at new energy power plants according to claim 1, characterized in that: The landscape generator consists of a bidirectional long short-term memory network. The memory unit of the long short-term memory network includes an input gate, a forget gate, and an output gate. Assume the state of the memory unit at time t is c. t The input gate is i t The forget gate is f t The output gate is 0 t The final output of the LSTM unit is h. t The calculation formulas for each variable are as follows: i t =σ(W i [h t-1 ,F (t) ]+b i ) (12) f t =σ(W f [h t-1 ,F (t) ]+b f ) (13) the t =σ(W o [h t-1 ,F (t) ]+b o ) (15) In the formula h t-1 c is the output of the Long Short-Term Memory network at time t-1. t-1 F represents the state of the Long Short-Term Memory (LSTM) network memory unit at time t-1. (t) The feature sequence output by the feature extractor, tanh(·) denotes the tanh activation function, W c W i W f W o The weight matrices corresponding to the memory unit, input gate, forget gate, and output gate, respectively, b c ,b i ,b f ,b o These are the bias vectors corresponding to the memory unit, input gate, forget gate, and output gate, respectively. Bidirectional Long Short-Term Memory (BSSM) networks add a data flow of inverse BSSM to the forward BSSM network. The hidden states of the bidirectional BSSM network are independent of each other, enabling it to comprehensively learn the temporal patterns and trends hidden in the input sequence. The bidirectional BSSM network computes the hidden layers forward at each time step. and backward computation of hidden layers Final output power after splicing In the formula Let t be the output of the forward Long Short-Term Memory network. This indicates a positive long short-term memory network. Let t be the output of the reverse long short-term memory network. This represents a reverse long short-term memory network. This represents the weight matrix of a forward long short-term memory network. Let b represent the weight matrix of the inverse long short-term memory network. y This represents the bias vector of a bidirectional long short-term memory network.

10. The method for generating wind and solar time-series power at new energy power plants according to claim 1, characterized in that: The domain discriminator has three fully connected layers for dimensionality reduction of the input features. A normalization layer is added after the first two fully connected layers, and the LeakyReLU activation function is used for non-linear pattern recognition and classification. The sigmoid activation function is used in the last fully connected layer to map the output to probability values ​​between 0 and 1, which matches the definition of the cross-entropy loss function.

11. A time-series power generation system for wind and solar power plants based on an improved domain adversarial network, characterized in that: This system can be used to implement the wind and solar time-series power generation method for new energy power plants as described in any one of claims 1 to 10; The wind and solar time-series power generation system for new energy power plants includes: Data preprocessing module: used to normalize the wind and solar power data of the source and target domains, and to divide the data into training and testing sets; Feature extractor training module: Trains the feature extractor using wind and solar field data from the source and target domains; Channel attention mechanism modeling module: Add a channel attention mechanism module after the feature extractor to adjust the importance of channels by learning the dynamic and non-linear dependencies of channel weights; Wind and solar power generator training module: The wind and solar power generator is trained using source domain wind and solar power data and feature representations obtained through channel attention mechanism. A bidirectional long short-term memory network structure is adopted to model the temporal relationship between wind and solar power and other variables. Domain Discriminator Training Module: The feature representation extracted by the feature extractor is used to train the domain discriminator, which is used to distinguish the features of the source domain and the target domain; Wind and solar time-series power generation module: By inputting wind and solar power data of the target domain and feature representations extracted by the feature extractor, and combining with the wind and solar generator, the module generates the wind and solar time-series power of the target domain.

12. A computer-readable storage medium, characterized in that: It stores a computer program that, when executed by a processor, implements the wind and solar time-series power generation method for new energy power stations as described in any one of claims 1 to 10.

Images