A turning weather-based wind power prediction method based on a hybrid neural network
By improving the successive variational mode decomposition and hybrid neural network model, the accuracy and stability issues of wind power prediction under transitional weather conditions are solved, achieving high-precision short-term wind power prediction and supporting wind power grid-connected dispatch and power system security.
Patent Information
- Application Number
- CN202611276354.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-21
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies for wind power forecasting under transitional weather conditions suffer from large prediction errors, making it difficult to meet the needs of refined scheduling and wind-storage coordinated control. Furthermore, traditional methods suffer from problems such as mode decomposition mismatch and insufficient model generalization ability.
An improved successive variational mode decomposition (SVMD) is combined with a temporal convolutional network (TCN), a bidirectional gated recurrent unit (BiGRU), and an attention mechanism to construct a hybrid prediction model. The decomposition parameters and hyperparameters are then optimized collaboratively using the Newton-Raphson optimization algorithm (NRBO) to form a TCN-BiGRU-AM hybrid prediction model.
It improves the accuracy of short-term wind power forecasting under transitional weather conditions, reduces the degree of mode aliasing, enhances the response capability to long-term time series dependencies and local context changes, reduces the workload of manual parameter tuning, and is suitable for wind power grid-connected dispatch and safe operation of the power system.
Smart Images

Figure CN123121499A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of new energy power generation prediction technology, and relates to a wind power prediction method based on hybrid neural networks under changing weather conditions. Background Technology
[0002] As the penetration rate of new energy sources such as wind power in the power system continues to increase, wind power forecasting results have become an important basis for new energy consumption, dispatch planning, reserve capacity allocation, and source-load balance control. Under normal meteorological conditions, wind power sequences usually exhibit relatively continuous variation patterns, and existing statistical models, machine learning models, or conventional recurrent neural networks can meet short-term forecasting needs to a certain extent.
[0003] However, during cold waves, typhoons, severe convection, systematic precipitation, and other transitional weather events, meteorological elements such as wind speed, wind direction, temperature, air pressure, and humidity can change rapidly within a short period. Wind power output is prone to shifting from a stable state to a sharp drop, a sudden climb, or a violent oscillation, causing the power sequence to exhibit stronger non-stationarity, abrupt changes, and input-output mapping drift. If forecast errors are large during these periods, it can easily increase the risk of generator disconnection, scheduling plan deviations, and system frequency instability.
[0004] Existing wind power forecasting technologies under transitional weather conditions mainly include data augmentation, scene recognition, segmented forecasting, and probability interval forecasting. Data augmentation methods can alleviate the problem of insufficient samples for low-probability weather to some extent, but there may still be distributional differences between the generated samples and the actual transition process. Scene recognition and segmented forecasting methods rely on weather type classification and manual rules, which limits the model's generalization ability. Probability interval forecasting can describe uncertainty, but it often has the problem of being too wide near extreme transition points, making it difficult to directly meet the high-precision point forecasting requirements of refined scheduling and wind-storage coordinated control.
[0005] Furthermore, traditional decomposition methods typically require pre-setting the number of modes or relying on manual experience to select decomposition parameters, which can easily lead to over-decomposition, under-decomposition, or mode aliasing. Conventional deep learning prediction models struggle to simultaneously account for long-range temporal dependencies, local contextual abrupt changes, and contributions from key time steps under transitional weather conditions. Setting decomposition parameters and neural network hyperparameters separately can also easily result in mismatches between the data decomposition module and the prediction model. Therefore, there is an urgent need for a short-term wind power prediction method that is designed for transitional weather processes, capable of effectively decomposing strong non-stationary wind power sequences, jointly learning multi-scale temporal features, and collaboratively optimizing key parameters. Summary of the Invention
[0006] In view of this, the purpose of this invention is to provide a method for predicting wind power under transitional weather conditions based on a hybrid neural network.
[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting wind power under transitional weather conditions based on hybrid neural networks includes the following steps: S1: Collect wind power data and meteorological data of the target wind farm during and near-term weather transition periods, and construct a sample dataset according to the preset time resolution; S2: Preprocess the sample dataset; S3: The original wind power sequence is decomposed by improved successive variational mode decomposition (SVMD) to obtain multiple mode components with different frequency characteristics. S4: Recombine the power mode components obtained from the decomposition with meteorological variables to form the input feature matrix; S5: Construct a TCN-BiGRU-AM hybrid prediction model, including a temporal convolutional network, a bidirectional gated recurrent unit, and an attention mechanism, which respectively undertake the functions of long-range dependency extraction, local context modeling, and key moment weighting. S6: Establish a unified optimization function for decomposition parameters and prediction model parameters, and use the Newton-Raphson optimization algorithm (NRBO) to optimize the unified optimization function to obtain the optimal decomposition parameters and the optimal hybrid prediction model hyperparameters. S7: Based on the input feature matrix, the optimized improved successive variational mode decomposition parameters, and the hybrid neural network hyperparameters, train the TCN-BiGRU-AM hybrid prediction model and output the wind power prediction value at the target prediction time.
[0008] Furthermore, the meteorological data mentioned in step S1 includes at least one of wind speed, wind direction, temperature, air pressure, and humidity, and a sample dataset is constructed according to a preset time resolution.
[0009] Furthermore, step S2 specifically includes: The sample dataset is subjected to time alignment, outlier handling, normalization, and sliding window processing to construct a historical input window; let the length of the historical input window be... L At any moment t The input feature sequence includes historical power sequences and corresponding meteorological variables, and the prediction target is future... h Step wind power output; Let the length of the history input window be... L At any moment t The observable input feature sequence is represented as ;in, Includes historical wind power output as well as meteorological variables such as wind speed, wind direction, temperature, air pressure, and humidity, expressed as follows: ; The forecasting task is to estimate the wind power output h steps in the future, given a historical window. The overall prediction mapping relationship is shown below:
[0010] In the formula, Indicates the model parameters The nonlinear mapping function that determines the outcome.
[0011] Furthermore, the improved successive variational mode decomposition (SVMD) described in step S3 introduces a spectral isolation constraint term based on the stepwise mode extraction, ensuring that the currently extracted mode remains separated from the previously extracted modes in the frequency domain, thereby reducing mode aliasing. The improved objective function is shown below:
[0012] In the formula, u L ( t ) indicates the number of digits to be extracted. L One modal component; f L ( t ) indicates extracting the first L The residual components obtained after each modal component; u i ( t ) indicates the previously extracted number i One modal component; f ( t () represents the original wind power sequence to be decomposed; ω L This represents the center frequency of the i-th modal component; J 1 indicates a modal bandwidth constraint term; J 2 represents the reconstruction error constraint term; J 3 indicates a spectrum isolation constraint term; α Represents the modal bandwidth constraint term J A penalty weight of 1; L This indicates the sequence number of the current mode to be extracted. i This indicates the sequence number of the previously extracted mode.
[0013] During the solution process, the improved SVMD iteratively updates the current modal components according to the above formula; the corresponding modal update relationship is shown in the following formula:
[0014] In the formula, û L n+1 ( ω ) indicates the first n The +1 iteration yields the first... kFrequency domain representation of each modal component; ( ω This represents the frequency domain representation of the original wind power sequence; û i ( ω ) indicates the first i Frequency domain representation of the extracted modal components; λ ( ω ) represents the frequency domain representation of the Lagrange multiplier; ω Represents frequency variables; ω L and ω i Representing the current modal component and the th modal component respectively i The center frequencies of the extracted modal components; α Indicates the bandwidth penalty factor; β Indicates the spectral isolation constraint coefficient; n Indicates the number of iterations; i and k Indicates the modal number.
[0015] The spectral isolation term is used to constrain the frequency domain overlap between the new mode and the extracted modes, so that each mode forms a clearer distribution boundary in the frequency domain.
[0016] Furthermore, step S5, which involves constructing the TCN-BiGRU-AM hybrid prediction model, includes: Temporal Convolutional Networks (TCNs) extract multi-scale long-range temporal dependencies through causal convolution, dilated convolution, and residual connections. Causal convolution ensures the model uses only current and historical information at the prediction time. Dilated convolution expands the receptive field by introducing gaps between kernel elements, enabling the model to capture long-scale dependencies with fewer network layers. Residual connections mitigate gradient degradation during deep network training. The computational relationship of dilated causal convolution is shown below:
[0017] In the formula, F ( t () indicates time t The output of the dilated causal convolution; K Indicates the kernel length; x s-d·k Indicates that the index in the input sequence is sd·k Historical samples; d Indicates the expansion factor; i This indicates the index of the convolution kernel element.
[0018] The Bi-Gated Recurrent Unit (BiGRU) performs forward and reverse encoding within the historical input window to extract local contextual dynamic features. Its bi-directional state concatenation relationship is shown in the following equation:
[0019] In the formula, x t express t Input feature vector at any time; and These represent the forward GRU and the reverse GRU respectively. t The hidden state at any given moment; This indicates the hidden state from the previous moment; Indicates the hidden state in the next time step; h t This represents the result of concatenating bidirectional hidden states; the symbol || represents vector concatenation.
[0020] BiGRU uses both forward and backward GRU to encode the historical input window, and concatenates the hidden states in both directions at the same time step. The update and reset gates inside the GRU unit determine the retention ratio of the historical state to the current input, and the specific calculation process is as follows:
[0021] In the formula, z t Indicates an update to the door; r t Indicates that the door is being reset; t Indicates the candidate hidden state; h t express t Always hide your status; h t-1 This indicates the hidden state in the previous moment; x t This represents the current input feature vector; W z , W r and W These represent the weight matrices used in the calculation of the update gate, reset gate, and candidate hidden state, respectively. σ represents the Sigmoid activation function; tanh represents the hyperbolic tangent activation function; ⊙ represents element-wise multiplication.
[0022] The formula updates the degree of retention of historical information of the gate control and resets the dependence of the candidate state of the gate control on the past state; The attention mechanism assigns weights to the hidden states at different time steps, highlighting key observational information near weather turning points, power ramp-up points, or wind direction abrupt changes. Its attention score, weight normalization, and weighted output process are shown in the following equation:
[0023] In the formula, α i Indicates the first i Attention weights for each time step; h i Indicates the first i The hidden state at each time step; q Represents the attention query vector; s( h i ,q) indicates that by h i and q The calculated attention scoring function; exp(·) represents the exponential function; j This represents the time step index in the normalized summation. The attention weights reflect the relative contribution of each time step to the prediction target.
[0024] Furthermore, in step S6, the penalty factor in the improved successive variational mode decomposition and the hyperparameters of the TCN-BiGRU-AM hybrid prediction model are used as joint optimization variables. The hyperparameters include the number of bidirectional hidden layer nodes, the initial learning rate, the L2 regularization coefficient, and the learning rate adjustment factor. NRBO is introduced as a collaborative optimization strategy, consisting of the Newton-Raphson search rule and a trap avoidance operator. The Newton-Raphson search rule enhancement algorithm is used to refine the search near excellent candidate solutions. The trap avoidance operator expands the search range through random perturbation to prevent premature convergence of parameter search. NRBO optimizes the SVMD penalty factor with permutation entropy as the objective. The key hyperparameters of TCN-BiGRU-AM, including the number of bidirectional hidden layer nodes, initial learning rate, L2 regularization coefficient, and learning rate adjustment factor, are optimized with validation set MSE as the fitness function.
[0025] The beneficial effects of this invention are as follows: (1) The present invention introduces a spectrum isolation constraint term in the successive variational mode decomposition, which can reduce the mode aliasing of strong non-stationary power sequences under turning weather conditions, making low-frequency trends, local oscillations and high-frequency disturbances easier for the prediction model to learn respectively.
[0026] (2) This invention utilizes TCN, BiGRU and attention mechanism to form a hybrid neural network, which can simultaneously extract long-range temporal dependencies, local context changes and key time step contributions, and improve the model's response to sudden segments such as power ramp-up and cliff drop.
[0027] (3) The present invention utilizes the Newton-Raphson optimization algorithm to perform collaborative optimization of the decomposition penalty factor and the prediction model hyperparameters, which can reduce the mismatch problem caused by setting parameters separately for the decomposition module and the prediction module, and reduce the workload of manual parameter tuning.
[0028] (4) This invention is applicable to short-term wind power point prediction under transitional weather conditions, and can provide a basis for prediction of wind power grid connection scheduling, wind-storage coordinated control, reserve capacity arrangement and safe operation of power system.
[0029] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0030] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the TCN causal dilated convolution structure; Figure 2 A schematic diagram of the structure for modeling BiGRU bidirectional sequences; Figure 3 The figures show a comparison of ablation experiment curves; where (a) is a comparison of predictions from Model 2 and Model 4, and (b) is a comparison of predictions from Model 3 and Model 4. Figure 4 A comparison chart of the various models. Detailed Implementation
[0031] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0032] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.
[0033] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.
[0034] Example 1: This invention provides a short-term wind power forecasting method for weather transitions, comprising the following steps: S1: Collect wind power data and meteorological data of the target wind farm during and near-term weather transition periods. The meteorological data includes at least one of wind speed, wind direction, temperature, air pressure, and humidity, and construct a sample dataset according to a preset time resolution.
[0035] S2: Perform time alignment, outlier handling, normalization, and sliding window processing on the sample dataset to construct a historical input window; let the length of the historical input window be... L At any moment t The input feature sequence includes historical power sequences and corresponding meteorological variables, and the prediction target is future... h Step wind power output.
[0036] Let the length of the history input window be... L At any moment t The observable input feature sequence is represented as .in, Including historical wind power output as well as meteorological variables such as wind speed, wind direction, temperature, air pressure, and humidity, it can be expressed as: The forecasting task is to estimate the wind power output h steps in the future, given a historical window. The overall prediction mapping relationship is shown in equation (1).
[0037] (1) In equation (1), Indicates the model parameters The nonlinear mapping function is determined. This equation gives the input, output, and prediction step size of the model in this paper, and serves as the foundation for subsequent sequence decomposition, hybrid neural network modeling, and parameter optimization tasks.
[0038] S3: An improved successive variational mode decomposition is used to decompose the original wind power sequence, obtaining multiple mode components with different frequency characteristics. The improved successive variational mode decomposition introduces a spectral isolation constraint term on the basis of extracting modes step by step, so as to keep the currently extracted mode separate from the previously extracted modes in the frequency domain, thereby reducing mode aliasing.
[0039] SVMD avoids the problem of needing to preset the number of modes in traditional VMD by extracting modal components step by step. Compared with ordinary SVMD, the improved SVMD further introduces a spectral isolation constraint term into the objective function, so that the newly extracted modes and existing modes maintain better separation in the frequency domain. Its improved objective function is shown in Equation (2). Equation (2) describes the joint constraint of modal compactness and spectral isolation, which can reduce the modal aliasing phenomenon in strong fluctuation sequences of transition weather.
[0040] (2) During the solution process, the improved SVMD iteratively updates the current modal components according to equation (2). The corresponding modal update relationship is shown in equation (3), where the spectral isolation term is used to constrain the frequency domain overlap between the new mode and the extracted modes, so that each mode forms a clearer distribution boundary in the frequency domain.
[0041] (3) Equation (3) presents the improved SVMD update method in the frequency domain. Through this update, low-frequency trends, local oscillations, and high-frequency disturbances under transitional weather conditions can be assigned to different modal components, thereby reducing the difficulty of subsequent TCN-BiGRU-AM directly learning the strongly non-stationary original sequence. The decomposition process also needs to set stopping conditions to avoid over-decomposition producing spurious modes, or under-decomposition causing different frequency components to overlap.
[0042] The penalty factor directly affects the SVMD decomposition effect. A penalty factor that is too small may lead to insufficient mode constraints, while a penalty factor that is too large may make the decomposition results too smooth, weakening the ability to express mutation components. Therefore, this paper subsequently introduces NRBO to adaptively optimize the penalty factor in order to obtain decomposition results that are more suitable for transition weather power sequences.
[0043] S4: The power mode components obtained from the decomposition are recombined with meteorological variables to form the input feature matrix of the hybrid neural network, so that the power change information at different frequency scales and the corresponding meteorological driving factors are used as the input of the prediction model.
[0044] S5: Construct a TCN-BiGRU-AM hybrid prediction model. The temporal convolutional network extracts multi-scale long-range temporal dependencies through causal convolution, dilated convolution, and residual connections; bidirectional gated recurrent units perform forward and backward encoding within the historical input window to extract local contextual dynamic features; and an attention mechanism assigns weights to hidden states at different time steps, highlighting key observational information near weather turning points, power ramp-up points, or sudden wind direction changes.
[0045] After improved SVMD decomposition, the original power sequence is transformed into several modal components with different frequency characteristics. To fully explore the temporal relationship between these components and meteorological variables, this paper constructs a TCN-BiGRU-AM hybrid prediction model. This model consists of a temporal convolutional network, a bidirectional gated recurrent unit, and an attention mechanism, which respectively perform the functions of long-range dependency extraction, local context modeling, and key moment weighting.
[0046] 1) TCN temporal feature extraction TCN achieves efficient modeling of time series through causal convolution, dilated convolution, and residual connections. Causal convolution ensures that the model uses only current and historical information at the prediction time, avoiding the leakage of future information; dilated convolution expands the receptive field by introducing gaps between convolution kernel elements, enabling the model to capture long-scale dependencies with fewer network layers; residual connections help alleviate the gradient degradation problem during the training of deep networks. The calculation relationship of its dilated causal convolution is shown in Equation (4), and the structural connection method is as follows: Figure 1 As shown.
[0047] (4) like Figure 1 As shown, causal connections ensure that convolutional computations at the prediction time do not introduce future information, while dilated connections allow the model to cover a longer historical range without significantly increasing the number of layers. In turning point weather prediction, this structure can be used to capture long-range dependencies before and after weather changes, such as the slow rise in wind speed before the turning point, the lagged power response after the change in wind direction, and local perturbations before the power ramp-up. The multi-scale features obtained from TCN are then input into BiGRU to further characterize local contextual changes.
[0048] 2) BiGRU bidirectional sequence modeling Although TCN can extract long-range temporal features, its convolutional structure still has some shortcomings in fine-grained expression of local context dynamic changes. To further enhance the model's ability to characterize local fluctuations in transitional weather, this paper introduces a BiGRU module. GRU regulates the information transfer between historical states and current inputs through update and reset gates, and has fewer parameters and higher training efficiency compared to LSTM. BiGRU performs forward and backward encoding simultaneously within a given historical input window, and its bidirectional state concatenation relationship is shown in Equation (5). The network structure is as follows: Figure 2 As shown.
[0049] (5) like Figure 2As shown, BiGRU encodes the historical input window together with forward GRU and backward GRU, and splices the hidden states in both directions at the same time step. The update gate and reset gate inside the GRU unit determine the retention ratio of the historical state and the current input, and the specific calculation process is shown in Equation (6).
[0050] (6) In equation (6), the degree of retention of historical information in the updated gate control is used to reset the dependence of the candidate state on the past state. For power prediction of transitional weather, this gating mechanism enables the model to automatically adjust the weight of historical information before and after power abrupt changes, avoiding excessive interference from information during stable periods on the prediction of abrupt changes.
[0051] 3) Weighted key features of attention mechanism Under transitional weather conditions, the contribution of different historical time steps to future power is not the same. Observational information near the weather transition point, the starting point of power ramp-up, or the point of sudden change in wind direction is often more predictive than stable segments far away from the transition point. To this end, this paper introduces an attention mechanism after the BiGRU output to adaptively weight the hidden states at different time steps. The attention score, weight normalization, and weighted output process are shown in Equation (7).
[0052] (7) In Equation (7), the attention weights reflect the relative contribution of each time step to the prediction target. When the historical window contains sudden changes in wind speed, rapid changes in wind direction, or the starting point of power ramp-up, the attention mechanism can increase the weights of the hidden states at these critical moments, thereby enhancing the model's responsiveness to turning points and the interpretability of the results.
[0053] S6: Establish a unified optimization function for decomposition parameters and prediction model parameters, and use the Newton-Raphson optimization algorithm to optimize the unified optimization function to obtain the optimal decomposition parameters and the optimal hybrid prediction model hyperparameters; wherein, the penalty factor in the improved successive variational mode decomposition and the hyperparameters of the TCN-BiGRU-AM hybrid prediction model are used as joint optimization variables, and the hyperparameters include the number of bidirectional hidden layer nodes, the initial learning rate, the L2 regularization coefficient, and the learning rate adjustment factor.
[0054] The performance of decomposition-based hybrid neural networks is affected by both the quality of data decomposition and the network hyperparameters. Optimizing only the SVMD parameters without considering prediction performance may result in more regular decomposed sequences, but this may not be optimal for prediction. Conversely, optimizing network parameters on inputs with insufficient decomposition quality may lead to overfitting to noise or failure to capture effective modes. To mitigate this mismatch, this paper introduces NRBO as a collaborative optimization strategy.
[0055] NRBO consists of the Newton-Raphson search rule and a trap avoidance operator. The Newton-Raphson search rule enhances the algorithm's ability to explore near excellent candidate solutions; the trap avoidance operator expands the search range through random perturbations. The former is used to refine the search near better parameter combinations, while the latter is used to prevent premature convergence of parameter searches. Based on this mechanism, NRBO undertakes two types of optimization tasks: first, optimizing the SVMD penalty factor with permutation entropy as the objective; second, optimizing key hyperparameters of TCN-BiGRU-AM, including the number of bidirectional hidden layer nodes, initial learning rate, L2 regularization coefficient, and learning rate adjustment factor, with validation set MSE as the fitness function.
[0056] S7: Based on the optimized improved successive variational mode decomposition parameters and hybrid neural network hyperparameters, the prediction model is trained and outputs the wind power prediction value at the target prediction time.
[0057] Example 2: In this embodiment, wind power and meteorological data of a wind farm from March 1, 2022 to December 31, 2022 were collected, with a time resolution of 15 minutes. Data from April to December were used as the training set, and data from March were used as the test set. A turning point weather prediction experiment was conducted on the test set. All comparison models used the same data partitioning method.
[0058] Experiment 1 The first set of experiments was used to evaluate the contributions of the improved SVMD and NRBO optimization modules. Four model variants were set up: Model 1 was the basic VMD-TCN-BiGRU-AM model; Model 2 replaced VMD with improved SVMD but did not use NRBO hyperparameter optimization; Model 3 used NRBO optimization but did not introduce improved SVMD; Model 4 used both improved SVMD and NRBO, which was the complete co-optimization model in this set of ablation experiments.
[0059] The experimental results and experimental graphs are shown in Table 1 and 2. Figure 3 As shown in (a) and (b) in the figure, the experimental results show that both modules can independently improve prediction performance. Compared with Model 1, Model 2 reduces MSE by 12.97%, RMSE by 6.71%, MAE by 11.84%, MAPE by 5.93%, and R² by 0.0028. This indicates that the improved SVMD can provide more stable and predictive modal components for the prediction network. Model 3 also achieves performance improvement, with MSE reduced by 15.94%, RMSE by 8.31%, MAE by 10.87%, MAPE by 4.33%, and R² by 0.0045. This shows that NRBO can find more suitable model parameters compared to manual tuning or isolated parameter tuning.
[0060] Table 1
[0061] The complete Model 4 achieved optimal results across all five metrics. Compared to Model 1, Model 4 reduced MSE by 32.83%, RMSE by 18.04%, MAE by 25.98%, MAPE by 22.52%, and R² by 0.9719 to 0.9811. Model 4 outperformed both Model 2 and Model 3, indicating that the improved SVMD and NRBO are not simply independent additives but rather complementary. The improved SVMD provides clearer input components, while NRBO allows the prediction network parameters to better fit these components, thus achieving higher prediction accuracy.
[0062] Experiment 2 The second set of experiments compared the proposed method with typical prediction models, including LSTM, VMD-LSTM, VMD-TCN-BiGRU-AM, and VMD-SSA-LSTM. All models were evaluated on the same transitional weather test data. The comparison figures are shown below. Figure 4 As shown.
[0063] As shown in Table 2, the proposed method achieves the best results across all evaluation metrics. Compared to LSTM, the RMSE of the proposed method decreases from 19.7571 to 7.3828, and the R² increases from 0.8994 to 0.9851, indicating that a single recurrent neural network is insufficient to adequately adapt to strong non-stationary wind power sequences under transitional weather conditions. The VMD-LSTM method shows only a limited improvement over LSTM, indicating that simply using a decomposition method without enhancing the feature extraction capability of the prediction network is still insufficient to fully solve the problem of transitional weather prediction.
[0064] Table 2
[0065] Compared to VMD-TCN-BiGRU-AM, our proposed method reduces RMSE by 28.70% and MAE by 20.51%. This demonstrates that, under the same hybrid neural network architecture, replacing basic decomposition and isolated parameter tuning with improved SVMD and NRBO co-optimization significantly enhances model performance. Compared to the second-best performing VMD-SSA-LSTM, our proposed method reduces RMSE by 2.55%, MAE by 5.80%, MAPE by 2.00%, MSE by 5.03%, and R² by 0.9751 to 0.9851.
[0066] The comparison of the predicted curves further supports the above conclusions. Within the relatively stable range of approximately 610 to 670 points, the predicted curves of each model show little difference in their fit with the actual power curves. However, within the highly volatile range of approximately 670 to 700 points, the proposed method can more quickly and accurately follow the power abrupt changes, significantly outperforming the comparative models. This is particularly important for forecasting transitional weather, as actual operational risks are often concentrated in sudden upswings and precipitous changes, rather than in stable periods.
[0067] Example 3: An electronic device, comprising a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the method described in Embodiment 1 when executing the computer program.
[0068] Example 4: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in Embodiment 1.
[0069] Example 5: A computer program product includes a computer program that, when executed by a processor, implements the method described in Example 1.
[0070] In the above embodiments, the reference to "this embodiment" in the specification indicates that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple appearances of "this embodiment" do not necessarily refer to the same embodiment.
[0071] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.
[0072] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.
[0073] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.
[0074] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.
[0075] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be 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, or discrete hardware components.
[0076] This invention can be used in a wide range of general-purpose or special-purpose computing system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.
[0077] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.
[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for predicting wind power under transitional weather conditions based on hybrid neural networks, characterized in that: Includes the following steps: S1: Collect wind power data and meteorological data of the target wind farm during and near-term weather transition periods, and construct a sample dataset according to the preset time resolution; S2: Preprocess the sample dataset; S3: The original wind power sequence is decomposed by improved successive variational mode decomposition (SVMD) to obtain multiple mode components with different frequency characteristics. S4: Recombine the power mode components obtained from the decomposition with meteorological variables to form the input feature matrix; S5: Construct a TCN-BiGRU-AM hybrid prediction model, including a temporal convolutional network, a bidirectional gated recurrent unit, and an attention mechanism, which respectively undertake the functions of long-range dependency extraction, local context modeling, and key moment weighting. S6: Establish a unified optimization function for decomposition parameters and prediction model parameters, and use the Newton-Raphson optimization algorithm (NRBO) to optimize the unified optimization function to obtain the optimal decomposition parameters and the optimal hybrid prediction model hyperparameters. S7: Based on the input feature matrix, the optimized improved successive variational mode decomposition parameters, and the hybrid neural network hyperparameters, train the TCN-BiGRU-AM hybrid prediction model and output the wind power prediction value at the target prediction time.
2. The wind power prediction method based on hybrid neural networks under transitional weather conditions according to claim 1, characterized in that: The meteorological data mentioned in step S1 includes at least one of wind speed, wind direction, temperature, air pressure, and humidity, and a sample dataset is constructed according to a preset time resolution.
3. The wind power prediction method based on hybrid neural networks under transitional weather conditions according to claim 1, characterized in that: Step S2 specifically includes: The sample dataset is subjected to time alignment, outlier handling, normalization, and sliding window processing to construct a historical input window; let the length of the historical input window be... L At any moment t The input feature sequence includes historical power sequences and corresponding meteorological variables, and the prediction target is future... h Step wind power output; Let the length of the history input window be... L At any moment t The observable input feature sequence is represented as ;in, Includes historical wind power output as well as meteorological variables such as wind speed, wind direction, temperature, air pressure, and humidity, expressed as follows: ; The forecasting task is to estimate the wind power output h steps in the future, given a historical window. The overall prediction mapping relationship is shown below: In the formula, Indicates the model parameters The nonlinear mapping function that determines the outcome.
4. The wind power prediction method based on hybrid neural networks under transitional weather conditions according to claim 1, characterized in that: Step S3 describes an improved successive variational mode decomposition (SVMD) that introduces a spectral isolation constraint term based on the stepwise mode extraction. This ensures that the currently extracted mode remains separated from previously extracted modes in the frequency domain, thereby reducing mode aliasing. The improved objective function is shown below: In the formula, u L ( t ) indicates the number of digits to be extracted. L One modal component; f L ( t ) indicates extracting the first L The residual components obtained after each modal component; u i ( t ) indicates the previously extracted number i One modal component; f ( t () represents the original wind power sequence to be decomposed; ω L This represents the center frequency of the i-th modal component; J 1 indicates a modal bandwidth constraint term; J 2 represents the reconstruction error constraint term; J 3 indicates a spectrum isolation constraint term; α Represents the modal bandwidth constraint term J A penalty weight of 1; L This indicates the sequence number of the current mode to be extracted. i This indicates the sequence number of the previously extracted mode; During the solution process, the improved SVMD iteratively updates the current modal components according to the above formula; the corresponding modal update relationship is shown in the following formula: In the formula, û L n+1 ( ω ) indicates the first n The +1 iteration yields the first... k Frequency domain representation of each modal component; ( ω This represents the frequency domain representation of the original wind power sequence; û i ( ω ) indicates the first i Frequency domain representation of the extracted modal components; λ ( ω ) represents the frequency domain representation of the Lagrange multiplier; ω Represents frequency variables; ω L and ω i Representing the current modal component and the th modal component respectively i The center frequencies of the extracted modal components; α Indicates the bandwidth penalty factor; β Indicates the spectral isolation constraint coefficient; n Indicates the number of iterations; i and k Indicates the modal number; The spectral isolation term is used to constrain the frequency domain overlap between the new mode and the extracted modes, so that each mode forms a clearer distribution boundary in the frequency domain.
5. The wind power prediction method based on hybrid neural networks under transitional weather conditions according to claim 1, characterized in that: Step S5, which involves constructing the TCN-BiGRU-AM hybrid prediction model, includes: Temporal Convolutional Networks (TCNs) extract multi-scale long-range temporal dependencies through causal convolution, dilated convolution, and residual connections. Causal convolution ensures the model uses only current and historical information at the prediction time. Dilated convolution expands the receptive field by introducing gaps between kernel elements, enabling the model to capture long-scale dependencies with fewer network layers. Residual connections mitigate gradient degradation during deep network training. The computational relationship of dilated causal convolution is shown below: In the formula, F ( t () indicates time t The output of the dilated causal convolution; K Indicates the kernel length; x s-d·k Indicates that the index in the input sequence is sd·k Historical samples; d Indicates the expansion factor; i Indicates the index of the convolution kernel element; The Bi-Gated Recurrent Unit (BiGRU) performs forward and reverse encoding within the historical input window to extract local contextual dynamic features. Its bi-directional state concatenation relationship is shown in the following equation: In the formula, x t express t Input feature vector at any time; and These represent the forward GRU and the reverse GRU respectively. t The hidden state at any given moment; This indicates the hidden state from the previous moment; Indicates the hidden state in the next time step; h t This represents the result of concatenating the two-way hidden states; the symbol || indicates vector concatenation. BiGRU uses both forward and backward GRU to encode the historical input window, and concatenates the hidden states in both directions at the same time step. The update and reset gates inside the GRU unit determine the retention ratio of the historical state to the current input, and the specific calculation process is as follows: In the formula, z t Indicates an update to the door; r t Indicates that the door is being reset; t Indicates the candidate hidden state; h t express t Always hide your status; h t-1 This indicates the hidden state in the previous moment; x t This represents the current input feature vector; W z , W r and W These represent the weight matrices used in the calculation of the update gate, reset gate, and candidate hidden state, respectively. σ The symbol represents the Sigmoid activation function; tanh represents the hyperbolic tangent activation function; ⊙ represents element-wise multiplication; The formula updates the degree of retention of historical information of the gate control and resets the dependence of the candidate state of the gate control on the past state; The attention mechanism assigns weights to the hidden states at different time steps, highlighting key observational information near weather turning points, power ramp-up points, or wind direction abrupt changes. Its attention score, weight normalization, and weighted output process are shown in the following equation: In the formula, α i Indicates the first i Attention weights for each time step; h i Indicates the first i The hidden state at each time step; q Represents the attention query vector; s( h i ,q) indicates that by h i and q The calculated attention scoring function; exp(·) represents the exponential function; j The index represents the time step in the normalized summation; the attention weight reflects the relative contribution of each time step to the prediction target.
6. The wind power prediction method based on hybrid neural networks under transitional weather conditions according to claim 1, characterized in that: In step S6, the penalty factor in the improved successive variational mode decomposition and the hyperparameters of the TCN-BiGRU-AM hybrid prediction model are used as joint optimization variables. The hyperparameters include the number of bidirectional hidden layer nodes, the initial learning rate, the L2 regularization coefficient, and the learning rate adjustment factor. NRBO is introduced as a collaborative optimization strategy, consisting of the Newton-Raphson search rule and a trap avoidance operator. The Newton-Raphson search rule enhancement algorithm is used to refine the search near excellent candidate solutions. The trap avoidance operator expands the search range through random perturbation to prevent premature convergence of parameter search. NRBO optimizes the SVMD penalty factor with permutation entropy as the objective. The key hyperparameters of TCN-BiGRU-AM, including the number of bidirectional hidden layer nodes, initial learning rate, L2 regularization coefficient, and learning rate adjustment factor, are optimized with validation set MSE as the fitness function.
7. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to implement the method as described in any one of claims 1-6 when executing the computer program.
8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1-6.
9. A computer program product, characterized in that: Includes a computer program that, when executed by a processor, implements the method as described in any one of claims 1-6.
