Wind power probability load prediction method, system and equipment based on meta-learning and medium
By calculating the gradients of quantiles and monotonic constraint losses in wind farm data, eliminating conflicting components, and iteratively optimizing, the problems of insufficient generalization ability and inconsistent probability distribution in wind power probabilistic prediction are solved, enabling rapid adaptation and accurate prediction in newly built wind farms.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ECONOMIC TECH RES INST OF STATE GRID ANHUI ELECTRIC POWER
- Filing Date
- 2026-04-17
- Publication Date
- 2026-05-15
AI Technical Summary
Existing wind power probabilistic load forecasting methods lack generalization ability in scenarios with few samples, such as newly built wind farms, and lack unified modeling of probability distribution structure and uncertainty characteristics, leading to quantile crossover and probability calibration distortion problems.
By acquiring source wind farm data, a meta-training task is constructed. The gradients of quantile loss and monotonic constraint loss are calculated, components with opposite gradient directions are removed, and gradients are superimposed to update model parameters. The meta-loss is calculated by combining task adaptation parameters and target set, and iterative optimization is performed until convergence, outputting probability prediction results.
It achieves rapid adaptation and accurate prediction in target wind farms, avoids quantile crossover and probability calibration distortion, and improves the model's generalization ability in scenarios with few samples.
Smart Images
Figure CN122051950A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind power load forecasting technology, and more specifically, to a wind power probabilistic load forecasting method, system, device, and medium based on meta-learning. Background Technology
[0002] As wind power penetration in power systems continues to increase, the intermittency, randomness, and volatility of wind power output place extremely high demands on the safe operation of the power grid. Wind power probabilistic load forecasting, because it can quantify forecast uncertainty and output forecast intervals or probability distributions, provides a more comprehensive decision-making basis for grid risk dispatching and has become a current research hotspot.
[0003] To address the issue of insufficient generalization ability of traditional methods in scenarios with limited samples, such as newly built wind farms and extreme weather, existing technologies introduce meta-learning methods. These methods train on multiple source tasks to obtain models with good initialization characteristics, enabling them to quickly adapt to the target task using a small number of samples, thereby improving cross-scenario prediction capabilities. Existing methods typically employ a two-layer optimization framework of "meta-training + fast adaptation." In the meta-training phase, point prediction error or quantile loss is used as the optimization objective, while in the adaptation phase, model transfer is achieved through a small number of gradient updates.
[0004] Existing methods, in the process of combining meta-learning and probabilistic prediction, lack a unified modeling of the overall structure and uncertainty characteristics of the probability distribution. On the one hand, under conditions of few samples, the model is insufficient in characterizing cognitive uncertainty, which can easily lead to excessive shrinkage of the prediction interval and probability calibration distortion. On the other hand, optimization methods based on discrete quantiles lack constraints on the monotonicity of the distribution, and are prone to quantile crossover when the distribution shifts, thus disrupting the consistency of the probability distribution.
[0005] Furthermore, the aforementioned problems are intertwined during training and adaptation. Insufficient characterization of uncertainty exacerbates tail-partition bias, while the lack of distribution structure constraints further worsens calibration results, leading to simultaneous failure of probability prediction in both accuracy and consistency. Therefore, how to achieve collaborative modeling of probability distribution structure and uncertainty within a meta-learning framework has become a critical technical problem that urgently needs to be solved.
[0006] To address the above problems, this invention proposes a solution. Summary of the Invention
[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a wind power probabilistic load prediction method, system, device, and medium based on meta-learning. The aim is to solve the technical problems in existing wind power probabilistic prediction methods, such as quantile crossover, probability calibration distortion, and insufficient generalization ability in scenarios with few samples, caused by the lack of unified modeling of the overall structure and uncertainty characteristics of the probability distribution.
[0008] To achieve the above objectives, the present invention provides the following technical solution: The wind power probabilistic load prediction method based on meta-learning includes the following steps: acquiring source wind farm data to construct meta-training tasks, each task containing a support set and a target set; calculating the first gradient of quantile loss and the second gradient of monotonic constraint loss based on the support set, removing components in the second gradient that are opposite in direction to the first gradient, and then superimposing them with the first gradient to obtain the superimposed gradient; updating model parameters based on the superimposed gradient to obtain task-fitting parameters; calculating meta-loss and updating meta-parameters using the task-fitting parameters and the target set, iterating until the meta-loss converges to obtain the final meta-parameters; and adapting the final meta-parameters to the target wind farm to output the probabilistic prediction results.
[0009] In a preferred embodiment, the step of acquiring source wind farm data to construct a meta-training task includes: dividing multiple meta-training tasks based on multiple source wind farm data; performing time-series partitioning on the data samples corresponding to each meta-training task to form a support set and a target set that satisfy time sequence constraints, wherein the time corresponding to the support set is earlier than that of the target set.
[0010] In a preferred embodiment, the calculation of the first gradient of the quantile loss and the second gradient of the monotonic constraint loss specifically involves: inputting support set samples into a prediction network with shared parameters to obtain predicted values for multiple quantiles; calculating the quantile loss based on the predicted values and the true values, and backpropagating to obtain the first gradient; constructing a ranking consistency constraint based on the predicted values of the quantiles, and obtaining a constraint error signal by quantizing the ranking consistency deviation; calculating the monotonic constraint loss based on the constraint error signal, and backpropagating to obtain the second gradient.
[0011] In a preferred embodiment, obtaining the superimposed gradient specifically involves: calculating the dot product of the first gradient and the second gradient; when the dot product is negative, decomposing the second gradient into a parallel component along the direction of the first gradient and an orthogonal component perpendicular to the direction of the first gradient; suppressing the parallel component and scaling the orthogonal component using a preset scaling factor; and superimposing the first gradient with the scaled orthogonal component to obtain the superimposed gradient.
[0012] In a preferred embodiment, the preset scaling factor is obtained by: determining the preset scaling factor based on the dot product of the first gradient and the second gradient, such that the preset scaling factor decreases as the dot product decreases.
[0013] In a preferred embodiment, the step of updating the model parameters based on superimposed gradients to obtain task adaptation parameters specifically involves: calculating a violation rate based on the monotonicity violation between quantiles in the support set prediction results; determining the gain coefficient of the orthogonal components according to the violation rate, wherein the gain coefficient is positively correlated with the violation rate; multiplying the preset scaling factor by the gain coefficient to obtain an adjusted scaling factor; scaling the orthogonal components based on the adjusted scaling factor, and updating the model parameters after superimposing them with the first gradient to obtain task adaptation parameters.
[0014] In a preferred embodiment, obtaining the final meta-parameters specifically involves: applying the task adaptation parameters to the target set, calculating the quantile regression loss and the calibration bias loss, wherein the calibration bias loss is constructed based on the deviation between the prediction interval coverage and the confidence level; weighting and summing the quantile regression loss and the calibration bias loss to obtain the meta-loss, and updating the initial parameters through backpropagation; iterating the above process until the meta-loss converges to obtain the final meta-parameters.
[0015] A system for wind power probabilistic load prediction based on meta-learning includes: a task construction module for acquiring source wind farm data to construct meta-training tasks, each task containing a support set and a target set; a gradient processing module for calculating the first gradient of quantile loss and the second gradient of monotonic constraint loss based on the support set, removing components in the second gradient that are opposite in direction to the first gradient, and then superimposing them with the first gradient to obtain a superimposed gradient; an adaptation and update module for updating model parameters based on the superimposed gradient to obtain task adaptation parameters; calculating meta-loss and updating meta-parameters using the task adaptation parameters and the target set, iteratively executing until the meta-loss converges to obtain final meta-parameters; and a prediction output module for adapting the target wind farm based on the final meta-parameters and outputting probabilistic prediction results.
[0016] An electronic device, comprising a memory and a processor: the memory for storing a program; the processor for executing the program to implement the various steps of the meta-learning-based wind power probabilistic load forecasting method.
[0017] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned meta-learning-based wind power probabilistic load forecasting method.
[0018] The technical effects and advantages of the wind power probabilistic load forecasting method, system, equipment, and medium based on meta-learning of this invention are as follows: 1. This invention calculates the first gradient of the quantile loss and the second gradient of the monotonic constraint loss based on the support set. The component with the opposite direction to the first gradient is removed and superimposed on the first gradient to obtain the superimposed gradient. The model parameters are then updated to obtain the task adaptation parameters. This effectively solves the problem of mutual interference in optimization caused by the conflict of gradient directions of multiple loss functions in the prior art, ensures the monotonic consistency of quantile prediction, and avoids the destruction of the probability distribution structure.
[0019] 2. This invention calculates the meta-loss and updates the meta-parameters by utilizing task adaptation parameters and target sets, iterating until the meta-loss converges to obtain the final meta-parameters, and adapting the model to the target wind farm based on the final meta-parameters and outputting probability prediction results. This enables the model to quickly adapt to the target wind farm with only a small number of samples, solving the problems of overfitting and insufficient generalization ability of existing methods in scenarios with scarce data such as newly built wind farms, and realizing rapid transfer and accurate prediction across wind farms. Attached Figure Description
[0020] Figure 1 This is a flowchart illustrating the wind power probabilistic load forecasting method based on meta-learning of the present invention. Figure 2 This is a schematic diagram of the system structure of the wind power probabilistic load prediction method based on meta-learning of the present invention. Figure 3 A structural block diagram of an exemplary electronic device provided for implementing embodiments of the present disclosure; Figure 4 This is a comparison chart of the average validation loss under different sample sizes in this invention; Figure 5 This is a comparison chart of the multi-cycle probabilistic wind power prediction evaluation indicators in test set A of this invention; Figure 6 This is a comparison chart of the multi-cycle probabilistic wind power prediction and evaluation indicators in test set B of this invention. Detailed Implementation
[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] Example 1, Figure 1 The present invention provides a wind power probabilistic load forecasting method based on meta-learning, comprising the following steps: S1, acquire source wind farm data to construct meta-training tasks, each task includes a support set and a target set; In this embodiment, the acquisition of source wind farm data to construct meta-training tasks, each task including a support set and a target set, specifically: First, power time-series data and synchronous meteorological data from multiple source wind farms are acquired, with each source wind farm corresponding to an independent data source.
[0023] S11, collect historical operating data from at least two source wind farms. The historical operating data includes, but is not limited to, wind speed, wind direction, temperature, humidity, air pressure, and historical wind power output. In this embodiment, preferably, three geographically dispersed source wind farms are selected, denoted as Wind Farm A, Wind Farm B, and Wind Farm C, respectively. The data for each wind farm spans one year, with a sampling interval of 15 minutes.
[0024] Multiple meta-training tasks are constructed using either a "one task per wind farm" or "random sampling" approach. Specifically, each meta-training task corresponds to a continuous time-series sample extracted from a source wind farm. For example, M continuous time windows are randomly selected from wind farm A, with each window serving as an independent meta-training task. In this way, a task pool containing K meta-training tasks is constructed. Where K is much larger than the number of source wind farms to ensure mission diversity.
[0025] S12, for each meta-training task constructed in step S11 The corresponding continuous temporal samples are then non-overlappingly segmented to form a support set that satisfies strict temporal order constraints. and target set Suppose a meta-training task contains a continuous time series of length L. ,in The input features at time t (such as wind speed, historical power). Let be the actual wind power output (load value) at time t. The samples at a given time point are assigned to the support set, and the remaining samples are assigned to the target set.
[0026] It should be noted that the timestamps of all samples in the support set are strictly earlier than the timestamps of all samples in the target set.
[0027] Compared to existing technologies that divide training data into single wind farms and single time periods, this method significantly increases the diversity of meta-training tasks by dividing multi-source wind farm data into tasks and using random sampling within time windows. This allows the model to learn richer wind power variation patterns. Simultaneously, strict temporal partitioning (support set time precedes target set time) ensures that no data leakage occurs during model adaptation, aligning with the physical scenarios of actual wind power prediction.
[0028] S2, calculate the first gradient of quantile loss and the second gradient of monotonic constraint loss based on the support set, remove the component of the second gradient that is opposite to the direction of the first gradient and then superimpose it with the first gradient to obtain the superimposed gradient; S21, in this embodiment, the calculation of the first gradient of the quantile loss and the second gradient of the monotonic constraint loss based on the support set specifically includes: S21-1 will support set samples Input to a prediction network with shared parameters In this embodiment, the prediction network uses a multi-layer residual LSTM as the basic prediction network, and its structural parameters are shown in Table 1.
[0029] Table 1
[0030] Let the set of quantile levels be... ,in This embodiment is preferred. , .
[0031] Network output is , indicating that the i-th sample is in the i-th order. Predicted wind power output at the quantile level. Calculate quantile losses. :
[0032] in, To support a concentrated sample size, For sample index, Quantile-level index, For the first Each quantile level represents a predicted quantile. For the first The actual wind power output (actual load) of each sample, quantile loss function , For prediction error, For the characteristic function, when The value is 1 if it is less than 0, and 0 otherwise. This represents the current quantile level.
[0033] Furthermore, regarding quantile loss Perform backpropagation to calculate model parameters First gradient :
[0034] S21-2, Constructing ranking consistency constraints based on quantile predicted values. Based on the definition of quantiles, for any two quantile levels... The corresponding predicted value should satisfy the monotonically increasing constraint: If this constraint is violated, a sorting consistency deviation occurs. First, sorting consistency constraints are constructed, and then the sorting consistency deviation is quantified to obtain the constraint error signal. :
[0035] in, To constrain the error signal, quantize the quantile of the i-th sample. and Discrepancy in sorting consistency between them For lower quantiles The corresponding predicted value, Higher quantiles The corresponding predicted value, To maximize the function, ensure that the deviation is non-negative.
[0036] Furthermore, the monotonic constraint loss is calculated based on the constraint error signal. :
[0037] in, For all quantile pairs The number of combinations.
[0038] Furthermore, regarding Perform backpropagation and calculate the second gradient. :
[0039] S22, in this embodiment, the step of removing the component in the second gradient that is opposite in direction to the first gradient and then superimposing it with the first gradient to obtain the superimposed gradient specifically involves: S22-1, Calculate the first gradient With the second gradient dot product :
[0040] like This indicates that the two gradient directions do not conflict, and their direct superposition yields the superimposed gradient. ; like ,show Existence and For components with opposite directions, a decomposition operation is performed, specifically: [decompose the components]. Decomposed into along Parallel components of direction and perpendicular to orthogonal components of direction :
[0041] Furthermore, for parallel components In this embodiment, the parallel component is completely eliminated (set to zero) and suppressed using a preset scaling factor. Scaling orthogonal components: .
[0042] Furthermore, the first gradient With scaled orthogonal components Superposition yields the superimposed gradient. .
[0043] S23, in this embodiment, the preset proportional coefficient Based on dot product Determine, make Follow The decrease is proportional to the decrease. This embodiment uses a linear mapping:
[0044] in, This is the L2 norm (modulus) of the first gradient. It is the L2 norm (modulus) of the second gradient.
[0045] It should be noted that, compared to existing techniques that ignore quantile monotonicity constraints or simply stack multiple loss gradients, this method has the following advantages: (1) By constraining the error signal, the order consistency constraint is quantized into a differentiable loss function, so that the model can learn the quantile prediction that satisfies monotonicity in backpropagation, thus avoiding the quantile crossover problem from the source. (2) By using gradient decomposition and conflict detection mechanism, the parallel components in the second gradient that are opposite to the direction of the first gradient are removed, thus eliminating the interference of monotonic constraint loss on the optimization direction of quantile loss. At the same time, the orthogonal components are retained so that the monotonic constraint can still play a role in the direction without conflict. (3) By setting a preset ratio coefficient The contribution strength of orthogonal components is dynamically adjusted according to the dot product, especially when gradient conflicts are severe. Approaching 0 when there is no conflict This achieves an automatic balance between the two optimization objectives. The above mechanism enables the model to effectively maintain the monotonically increasing relationship between quantiles while ensuring the accuracy of quantile prediction.
[0046] S3, update the model parameters based on superimposed gradients to obtain the task adaptation parameters; In this embodiment, the step of updating the model parameters based on superimposed gradients to obtain the task adaptation parameters specifically involves: S31, after obtaining the quantile prediction results for the support set, count the monotonicity violations between all quantile pairs. For each sample... and each pair of quantile levels If it appears Then record one violation. Calculate the violation rate. :
[0047] Among them, the total number of quantile pairs , It is an indicator function, taking the value 1 if the condition is true, and 0 otherwise.
[0048] S32, based on the violation rate Determine the gain coefficient of the orthogonal components The gain coefficient is positively correlated with the violation rate; that is, the more severe the violation, the larger the gain coefficient, thereby enhancing the contribution of the monotonic constraint gradient (orthogonal component). This embodiment uses a linear positive correlation:
[0049] in, This is a preset scaling factor.
[0050] Furthermore, the preset scaling factor is multiplied by the gain factor to obtain the adjusted scaling factor. Based on the adjusted scaling factor The orthogonal components are scaled and superimposed with the first gradient to obtain the final gradient used for parameter updates:
[0051] Furthermore, the superimposed gradients are used to update the model parameters to obtain the task adaptation parameters for the current task:
[0052] in, To adapt parameters for the updated task, These are the current model parameters (meta-parameters). This is the inner loop learning rate (adaptive learning rate).
[0053] It should be noted that this method will be based on gradient conflict adjustment ( ) and adjustments based on violation rate ( This combination forms a two-layer adjustment mechanism. The former responds to instantaneous conflicts at the gradient level, while the latter responds to cumulative biases at the result level. After being decoupled, the two are multiplied, enabling the model to quickly learn monotonicity constraints in the early stages of training (when the violation rate is high) and finely optimize quantile accuracy in the later stages of training (when the violation rate is low), achieving a smooth transition from rapid correction to fine optimization.
[0054] S4, calculate the meta-loss using the task adaptation parameters and target set and update the meta-parameters, iterate until the meta-loss converges to obtain the final meta-parameters; In this embodiment, the step of calculating the meta-loss and updating the meta-parameters using task adaptation parameters and target set, and iteratively executing this process until the meta-loss converges to obtain the final meta-parameters, specifically involves: S41, obtain the task adaptation parameters Acting on the target set ,in Let be the number of samples in the target set. For each sample in the target set, calculate the predicted quantile:
[0055] in, For the target set One sample in Predicted values under quantiles This is the sample index.
[0056] Furthermore, the quantile regression loss is calculated on the target set:
[0057] S42, the calibration bias loss is used to quantify the degree of calibration between the predicted probability distribution and the actual observed distribution, and is constructed based on the deviation between the predicted interval coverage and the confidence level. For any confidence level... The corresponding prediction interval is:
[0058] In this embodiment, a commonly used confidence level is used. Regarding confidence levels The formula for calculating the predicted interval coverage is:
[0059] Calculate calibration deviation:
[0060] Furthermore, the calibration bias loss can be constructed using the absolute value of mean square calibration error, quantile calibration error, or interval coverage deviation. This embodiment uses mean square calibration error, specifically:
[0061] in, The number of confidence levels selected. For the first One confidence level.
[0062] S43. The meta-loss is obtained by weighted summation of the quantile regression loss and the calibration bias loss. :
[0063] in, These are the weighting coefficients.
[0064] Furthermore, the original loss For initial parameters Find the derivative and calculate the elementary gradient. :
[0065] Furthermore, the meta-parameters (initial parameters) are updated using the meta-gradient:
[0066] in, These are the meta-parameters (initial parameters, objects of outer optimization). This is the meta-learning rate (outer learning rate).
[0067] S44, Repeat the above inner and outer loop training process until the meta-loss is reached. The convergence condition is met, and the final meta-parameters are obtained. The convergence condition includes: intrinsic loss for 20 consecutive epochs. The rate of change is less than 0.01%; the average monotonicity violation rate on the validation task set and ; Reach the maximum preset number of iterations (e.g., 500 iterations).
[0068] S5 adapts the target wind farm based on the final meta-parameters and outputs the probability prediction results.
[0069] In this embodiment, the process of adapting the target wind farm based on the final meta-parameters and outputting the probability prediction result specifically involves: For newly built or data-scarce target wind farms, a small number of continuous time-series samples are collected as an adaptation set. In this embodiment, the adaptation set contains 24 time points, corresponding to 6 hours of historical data (based on a 15-minute sampling rate). Each sample includes input features (such as wind speed, wind direction, temperature, humidity, historical power, etc.) and the corresponding actual wind power output value. The final meta-parameters obtained through training are then used... Load it into the prediction network and use it to initialize the model.
[0070] The model is quickly adapted using an adaptation set to obtain parameters specific to the target wind farm. The adapted parameters Applying to the forecast period, it performs probabilistic prediction of wind power for the next 24 hours (96 time points in total, 15-minute resolution), and outputs wind power prediction values at 9 quantile levels (10%, 20%, ..., 90%). Based on the quantile prediction results, a prediction interval with a 90% confidence level is constructed.
[0071] Example 2, Figure 2 The present invention provides a system for predicting industry electricity consumption based on industrial chain linkages, comprising: The task construction module is used to acquire source wind farm data to construct meta-training tasks. Each task contains a support set and a target set. The gradient processing module is used to calculate the first gradient of the quantile loss and the second gradient of the monotonic constraint loss based on the support set. The component in the second gradient that is opposite to the direction of the first gradient is removed and then superimposed with the first gradient to obtain the superimposed gradient. The adaptation and update module is used to update the model parameters based on the superimposed gradient to obtain the task adaptation parameters; calculate the meta-loss using the task adaptation parameters and the target set and update the meta-parameters, and iterate until the meta-loss converges to obtain the final meta-parameters; The prediction output module is used to adapt the target wind farm based on the final meta-parameters and output the probability prediction results.
[0072] Example 3, an electronic device, such as Figure 3 As shown, it includes a memory and a processor: the memory is used to store a program; the processor is used to execute the program to implement any of the embodiments in Example 1.
[0073] Since the electronic device described in this embodiment is the device used to implement the method in Embodiment 1 of the present invention, those skilled in the art can understand the specific implementation method and various variations of the electronic device in this embodiment based on the method described in Embodiment 1 of this application. Therefore, how the electronic device implements the method in the embodiments of this application will not be described in detail here. Any device used by those skilled in the art to implement the method in the embodiments of this application falls within the scope of protection of this application.
[0074] Example 4: A readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the embodiments in Example 1.
[0075] Example 5: This example uses two sets of real wind farm datasets for verification, and the specific information is shown in Table 2.
[0076] Table 2 Test Dataset Details
[0077] To verify the rapid adaptability of this method under small sample conditions, comparative experiments were conducted at different sample sizes, and the results are shown in Table 3.
[0078] Table 3 Comparison of average validation loss under different sample sizes
[0079] The results show that the proposed method achieves the lowest validation loss for all sample sizes, and performs particularly well when the training samples are limited. With only 10 samples, the validation loss of the proposed method (0.0452) is even lower than the performance of STL with 1000 samples (0.0485), which fully demonstrates the superiority of the proposed method in scenarios with few samples.
[0080] Table 4 shows a comparison between our method and various benchmark methods in multi-step prediction scenarios.
[0081] Table 4 Multi-period probabilistic wind power prediction evaluation indicators
[0082] As shown in Table 4, our method achieves dependencies of 1.15% and 1.08% on test sets A and B, respectively, both the lowest among all methods, indicating that the predicted probability distribution is most consistent with actual observations. Our method achieves an sharpness of 0.0677 on test set A, the lowest among all methods (narrowest prediction interval); and an sharpness of 0.326 on test set B, comparable to the optimal method. Our method achieves skill scores of -0.258 and -1.17 on test sets A and B, respectively, both the highest among all methods (closest to 0), indicating the best overall performance.
[0083] The above formulas are all dimensionless and numerical. The formulas are derived from software simulation based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.
[0084] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.
[0085] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0086] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.
[0087] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
[0088] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A wind power probabilistic load forecasting method based on meta-learning, characterized in that, Includes the following steps: Obtain source wind farm data to construct meta-training tasks, each task containing a support set and a target set; The first gradient of the quantile loss and the second gradient of the monotonic constraint loss are calculated based on the support set. The component in the second gradient that is opposite to the direction of the first gradient is removed and then superimposed with the first gradient to obtain the superimposed gradient. The model parameters are updated based on superimposed gradients to obtain the task adaptation parameters; The meta-loss is calculated and the meta-parameters are updated using the task adaptation parameters and the target set. The process is iterated until the meta-loss converges to obtain the final meta-parameters. The target wind farm is adapted based on the final meta-parameters, and the probability prediction results are output.
2. The wind power probabilistic load forecasting method based on meta-learning according to claim 1, characterized in that, The task of acquiring source wind farm data to construct the meta-training task includes: Multiple meta-training tasks are divided based on data from multiple wind farm sources; The data samples corresponding to each meta-training task are divided into time sequences to form a support set and a target set that satisfy the time order constraint, wherein the time corresponding to the support set is earlier than that of the target set.
3. The wind power probabilistic load forecasting method based on meta-learning according to claim 2, characterized in that, The calculation of the first gradient of the quantile loss and the second gradient of the monotonic constraint loss are specifically as follows: By supporting a prediction network with shared parameters for a set of sample inputs, predictions for multiple quantiles can be obtained. The quantile loss is calculated based on the predicted and true values, and the first gradient is obtained through backpropagation. Based on the predicted values of the quantiles, a ranking consistency constraint is constructed, and the constraint error signal is obtained by quantifying the ranking consistency deviation. The monotonic constraint loss is calculated based on the constraint error signal, and the second gradient is obtained by backpropagation.
4. The wind power probabilistic load forecasting method based on meta-learning according to claim 3, characterized in that, The process of obtaining the superimposed gradient is as follows: Calculate the dot product of the first gradient and the second gradient. When the dot product is negative, decompose the second gradient into a parallel component along the direction of the first gradient and an orthogonal component perpendicular to the direction of the first gradient. The parallel components are suppressed, and the orthogonal components are scaled using a preset scaling factor. The first gradient is superimposed with the scaled orthogonal component to obtain the superimposed gradient.
5. The wind power probabilistic load forecasting method based on meta-learning according to claim 4, characterized in that, The preset proportional coefficient is obtained as follows: The preset scaling factor is determined based on the dot product of the first gradient and the second gradient, such that the preset scaling factor decreases as the dot product decreases.
6. The wind power probabilistic load forecasting method based on meta-learning according to claim 5, characterized in that, The process of updating model parameters based on superimposed gradients to obtain task adaptation parameters is as follows: The violation rate is calculated based on the monotonicity violations among quantiles in the support set prediction results; The gain coefficient of the orthogonal component is determined based on the violation rate, wherein the gain coefficient is positively correlated with the violation rate. The preset scaling factor is multiplied by the gain coefficient to obtain the adjusted scaling factor. The orthogonal components are scaled based on the adjusted scaling factor, and then superimposed with the first gradient to update the model parameters, thus obtaining the task adaptation parameters.
7. The wind power probabilistic load forecasting method based on meta-learning according to claim 6, characterized in that, The final meta-parameters are obtained as follows: The task adaptation parameters are applied to the target set to calculate the quantile regression loss and calibration bias loss, wherein the calibration bias loss is constructed based on the deviation between the prediction interval coverage and the confidence level. The weighted sum of the quantile regression loss and the calibration bias loss is used to obtain the meta-loss, and the initial parameters are updated through backpropagation. Iterate through the above process until the meta-loss converges to obtain the final meta-parameters.
8. A system using the wind power probabilistic load forecasting method based on meta-learning as described in any one of claims 1-7, characterized in that, include: The task construction module is used to acquire source wind farm data to construct meta-training tasks. Each task contains a support set and a target set. The gradient processing module is used to calculate the first gradient of the quantile loss and the second gradient of the monotonic constraint loss based on the support set. The component in the second gradient that is opposite to the direction of the first gradient is removed and then superimposed with the first gradient to obtain the superimposed gradient. The adaptation update module is used to update the model parameters based on the superimposed gradient to obtain the task adaptation parameters. The meta-loss is calculated and the meta-parameters are updated using the task adaptation parameters and the target set. The process is iterated until the meta-loss converges to obtain the final meta-parameters. The prediction output module is used to adapt the target wind farm based on the final meta-parameters and output the probability prediction results.
9. An electronic device, characterized in that, The electronic device includes a memory and a processor: The memory is used to store programs; The processor is configured to execute the program to implement each step of the wind power probabilistic load forecasting method based on meta-learning as described in any one of claims 1-7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the wind power probabilistic load forecasting method based on meta-learning as described in any one of claims 1 to 7.