Wind power short-term prediction method, terminal device and computer readable storage medium
By introducing dynamic bias and confidence calculation into wind power forecasting, and combining multiple correction methods, a closed-loop feedback mechanism from ultra-short-term to short-term is formed, which solves the problem of insufficient forecast accuracy in existing technologies and achieves high-precision and robust forecasting under complex meteorological conditions.
Patent Information
- Application Number
- CN202610632035.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-09
- Publication Date
- 2026-07-31
AI Technical Summary
Existing wind power forecasting methods struggle to consistently meet forecasting accuracy requirements under complex weather conditions. There is no effective feedback mechanism between ultra-short-term and short-term forecasts, and they lack adaptive error correction capabilities, leading to a rapid decline in forecast accuracy and an increase in wind curtailment.
By determining the power error between the predicted and measured wind power values, calculating the dynamic bias and confidence level, and combining multiple correction methods (bias correction, online parameter fine-tuning, and similar sample adjustment) to correct the short-term prediction model, a closed-loop feedback mechanism from ultra-short-term to short-term is formed, and unified regulation is carried out using multi-temporal and spatial scale characteristics.
It significantly improves the consistency and accuracy of wind power forecasting, enhances forecast precision, and possesses high robustness. In particular, it can effectively correct errors under extreme weather conditions, thereby reducing the amount of wind power curtailed.
Smart Images
Figure CN122495332A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wind power prediction technology, and specifically discloses a short-term wind power prediction method, terminal equipment and computer-readable storage medium. Background Technology
[0002] Wind power, as a major form of clean and renewable energy, has been connected to the grid on a large scale globally and has become an important part of the power system. However, wind power output has significant randomness, volatility, and intermittency. Its forecasting errors directly affect the accuracy of grid dispatching plans, leading to increased wind curtailment and higher system operating costs.
[0003] Existing wind power prediction methods mainly include physical methods, statistical methods, and deep learning methods. Physical methods rely on numerical weather prediction and microscale models of wind farms, which are computationally complex and difficult to adapt to complex terrains; statistical methods are suitable for stationary sequences but struggle to capture the strong nonlinear characteristics of wind power; while deep learning methods possess strong nonlinear fitting capabilities, they generally suffer from the following problems:
[0004] 1. After the model is trained and the parameters are fixed, the prediction accuracy will rapidly decrease when faced with systematic biases in numerical weather forecasts, drift in wind measurement data, or sudden meteorological events.
[0005] 2. Ultra-short-term forecasts are independent of short-term forecasts, and no effective feedback mechanism has been formed, so ultra-short-term error information has not been fully utilized.
[0006] 3. Traditional correction methods often use single bias correction or simple weighted fusion, lacking adaptive confidence adjustment, and are prone to over-correction or under-correction under extreme weather conditions.
[0007] In summary, existing forecasting methods typically process ultra-short-term forecasts (e.g., the next 0-4 hours) and short-term forecasts (e.g., the next 1-10 days) separately, failing to fully utilize the correlation between information at different time scales. Under complex weather conditions, they are difficult to reliably meet the forecasting accuracy requirements. There is an urgent need for a new wind power forecasting method with strong robustness and adaptive error correction capabilities. Summary of the Invention
[0008] The purpose of this invention is to provide a short-term wind power forecasting method, terminal equipment, and computer-readable storage medium to solve the technical problem that existing methods are unable to stably meet the forecasting accuracy requirements under complex meteorological conditions.
[0009] A first aspect of the present invention provides a method for short-term wind power forecasting, comprising:
[0010] Step 1: Determine the power error between the predicted and measured values of wind power at each moment within the first time period;
[0011] Step 2: Determine the dynamic bias at the corresponding time based on the power error;
[0012] Step 3: Determine the confidence level at the corresponding time based on the dynamic bias;
[0013] Step 4: Compare the confidence level with a preset threshold, and determine the correction method for the output power of the short-term prediction model in the second time period based on the comparison result; the second time period is longer than the first time period.
[0014] Preferably, step 2 specifically includes:
[0015] The smoothing coefficient is determined based on the number of time nodes within the first time period;
[0016] The dynamic bias at the current moment is determined based on the smoothing coefficient, the dynamic bias at the previous moment, and the power error at the current moment within the first time period.
[0017] Preferably, step 3 specifically includes:
[0018] Obtain the installed capacity and sensitivity coefficient of the wind farm;
[0019] The confidence level at the corresponding time is determined based on the wind farm's installed capacity, sensitivity coefficient, and dynamic bias.
[0020] Preferably, step 4 specifically includes:
[0021] When the confidence level is less than or equal to the first threshold, the output power of the short-term prediction model in the second time period is corrected using the bias correction method.
[0022] When the first threshold < the confidence level ≤ the second threshold, the output power of the short-term prediction model in the second time period is corrected using the online parameter fine-tuning method and the bias correction method;
[0023] When the confidence level is greater than the second threshold, the output power of the short-term prediction model in the second time period is corrected using the similar sample adjustment method, the online parameter fine-tuning method, and the bias correction method.
[0024] Preferably, the output power of the short-term prediction model in the second time period is corrected using a bias correction method, specifically as follows:
[0025] Determine the decay time coefficient for each moment in the second time period predicted by the short-term prediction model;
[0026] The output power of the short-term prediction model in the second time period is corrected by the product of the decay time coefficient and the dynamic bias, and is denoted as the first output power.
[0027] Preferably, the output power of the short-term prediction model in the second time period is corrected using an online parameter fine-tuning method and the bias correction method, specifically as follows:
[0028] Freeze the previous short-term prediction model The parameters of the layer are determined, and the loss function is determined based on the difference between the output power and the actual power of the short-term prediction model after the parameters are frozen;
[0029] The parameters of the remaining layers of the short-term prediction model are updated using the loss function, and the output power of the short-term prediction model after updating the parameters is obtained, denoted as the second output power;
[0030] The final output power is determined based on the first output power and the second output power.
[0031] Preferably, the output power of the short-term prediction model in the second time period is corrected using the similar sample adjustment method, the online parameter fine-tuning method, and the bias correction method, specifically as follows:
[0032] The error between the predicted and measured values of wind power within a historical period is defined as the first error.
[0033] Determine the similarity between the first error and the power error;
[0034] The output power of the short-term prediction model is determined based on the similarity and the true power of the historical period, and is denoted as the third output power.
[0035] The final output power is determined based on the first output power, the second output power, and the third output power.
[0036] Preferably, the output power of the short-term prediction model is determined based on the similarity and the true power over historical periods, specifically as follows:
[0037] The weights in the short-term prediction model are determined based on the similarity.
[0038] The output power of the short-term prediction model is determined based on the actual power over historical periods and the weights.
[0039] A second aspect of the present invention provides a terminal device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described short-term wind power forecasting method.
[0040] A third aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described short-term wind power prediction method.
[0041] The wind power short-term prediction method, terminal equipment, and storage medium of the present invention have the following advantages compared with the prior art:
[0042] This invention significantly improves both prediction accuracy and robustness. Traditional wind power prediction methods often treat ultra-short-term and short-term predictions as separate components, neglecting the multi-temporal and spatiotemporal scale characteristics between them. This invention achieves closed-loop feedback from ultra-short-term to short-term predictions through a multi-real-time feedback mechanism with unified confidence level control. This effectively utilizes multi-temporal and spatiotemporal scale characteristics, thereby significantly improving the consistency and accuracy of wind power prediction. Furthermore, under the influence of dynamically similar samples, the short-term wind power prediction method of this invention exhibits high robustness to extreme weather conditions. Attached Figure Description
[0043] Figure 1 This is a flowchart of a short-term wind power prediction method according to an embodiment of the present invention.
[0044] Figure 2 The figure shows the results of a specific embodiment of the wind power short-term prediction method of the present invention. Detailed Implementation
[0045] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.
[0046] A first aspect of this invention provides a short-term wind power forecasting method, such as... Figure 1 As shown, it includes:
[0047] Step 1: Determine the power error between the predicted and measured values of wind power at each moment within the first time period.
[0048] This invention first acquires multi-source data, including numerical weather prediction data, real-time data collected from actual wind turbines at a resolution of at least 15 minutes, and meteorological tower data. Then, the acquired data undergoes preprocessing, including outlier removal, missing value imputation, and normalization, to construct a multimodal spatiotemporal sequence.
[0049] The preprocessing specifically involves: aligning multi-source data, including wind turbine acquisition data, numerical weather prediction data, meteorological tower data, and wind turbine location information, by time to construct a multimodal spatiotemporal sequence. Outliers are removed based on physical constraints, missing values are filled using KNN interpolation, and Min-Max normalization is applied to map the data to the [0,1] interval.
[0050] (1)
[0051] in It is the normalized value. These are the original data points. It is the minimum value of the entire sequence. It is the maximum value of the entire sequence.
[0052] Furthermore, in this embodiment of the invention, an ultra-short-term power prediction model is used to generate predicted wind power values for each moment within the first future time period. For example, the ultra-short-term power prediction model can be a neural network model, which is used to generate predicted wind power values for each moment within the next 0-4 hours, thus obtaining an ultra-short-term power prediction sequence.
[0053] The power error between the predicted and measured wind power values at each moment within the first time period is determined as follows:
[0054] Definition containing recent The ultra-short-term power prediction error sequence at each time point:
[0055] (2)
[0056] in, yes During the period Power error at time, yes Measured wind power output at any given time. yes The wind power prediction value output by the ultra-short-term power prediction model.
[0057] Step 2: Determine the dynamic offset at the corresponding time based on the power error, specifically as follows:
[0058] Step 2.1: Determine the smoothing coefficient based on the number of time nodes within the first time period. The smoothing coefficient is a core parameter of time series models (such as exponential smoothing), and its function is to assign weights to recent actual observations and historical forecasts.
[0059] Step 2.2: Determine the dynamic bias at the current moment based on the smoothing coefficient, the dynamic bias at the previous moment, and the power error at the current moment within the first time period.
[0060] The embodiments of the present invention will contain the most recent The exponentially weighted moving average of the ultra-short-term power prediction error sequence at each time point is used as the dynamic bias. The deadline corresponding to each time point is: Taking dynamic bias as an example:
[0061] (2)
[0062] in, It is the smoothing coefficient. yes Dynamic bias at any given moment; yes Dynamic bias at time.
[0063] Step 3: Determine the confidence level at the corresponding time point based on the dynamic bias, specifically as follows:
[0064] Step 3.1: Obtain the installed capacity of the wind farm. and sensitivity coefficient ;
[0065] Step 3.2: Based on the installed capacity of the wind farm Sensitivity coefficient And dynamic bias, to determine the confidence level at the corresponding time.
[0066] This invention, in its embodiments, calculates in real time based on at least one of the following: numerical weather prediction ensemble dispersion, mean absolute error (MAE) of ultra-short-term forecasts, and power ramp-up event detection results. Time confidence .
[0067] (3)
[0068] Step 4: Compare the confidence level with a preset threshold, and determine the correction method for the output power of the short-term prediction model in the second time period based on the comparison results; the second time period is greater than the first time period, specifically:
[0069] (1) When the confidence level is less than or equal to the first threshold, the output power of the short-term prediction model in the second time period is corrected using the bias correction method. For example, the first threshold can be 0.2.
[0070] (2) When the first threshold < confidence level ≤ second threshold, the output power of the short-term prediction model in the second time period is corrected using online parameter fine-tuning method and bias correction method. For example, the second threshold can be 0.6.
[0071] (3) When the confidence level is greater than the second threshold, the output power of the short-term prediction model in the second time period is corrected by using the similar sample adjustment method, the online parameter fine-tuning method and the bias correction method.
[0072] The embodiments of the present invention utilize a short-term prediction model to obtain prediction results for the next 1-10 days.
[0073] In this embodiment of the invention, the bias correction method is used to correct the output power of the short-term prediction model in the second time period, specifically as follows:
[0074] S1. Determine the decay time coefficient of the short-term prediction model for each moment in the second time period, as shown in formula (4).
[0075] (4)
[0076] in, For the second time period The decay time coefficient of the node, when the second period is 1 day. ; This is the decay time constant.
[0077] S2. The output power of the short-term prediction model in the second time period is corrected by the product of the decay time coefficient and the dynamic bias, and is denoted as the first output power, as shown in formula (5).
[0078] (5)
[0079] in, yes The first output power at any given moment; for The raw output power (i.e., the predicted power) of the short-term prediction model at time point.
[0080] In this embodiment of the invention, the product of the decay time coefficient and the dynamic bias is added to the original output power of the next day, thereby correcting the original output power of the short-term prediction model.
[0081] This invention utilizes online parameter fine-tuning and bias correction methods to correct the output power of the short-term prediction model in the second time period, specifically as follows:
[0082] S11, Freezing the front of the short-term forecast model The parameters of the layer are determined, and the auxiliary loss function is determined based on the difference between the output power and the actual power of the short-term prediction model after the parameters are frozen, as shown in formula (6).
[0083] (6)
[0084] in, It is an auxiliary loss function. Indicates recent Each time point yes The actual power at any given moment Before freezing The output power of the short-term prediction model after the layer.
[0085] by Taking a layer LSTM+fully connected layer as an example, before freezing The specific layer parameters are as follows:
[0086] (7)
[0087] in, It was before Layer freezing parameters, It refers to the number of frozen layers.
[0088] S12. Update the parameters of the remaining layers of the short-term prediction model using the loss function, and obtain the output power of the short-term prediction model after updating the parameters, denoted as the second output power.
[0089] This embodiment of the invention applies a learning rate of only the last 1-3 layers of parameters in the next-day short-term prediction model. Gradient updates of magnitude:
[0090] (8)
[0091] in, These are adjustable parameters in the back layer. The online learning rate employs a cosine annealing hot restart strategy:
[0092] (9)
[0093] The fine-tuned model was used to generate power prediction values for 96 points the following day:
[0094] (10)
[0095] in, It is the predicted value of the fine-tuned model, i.e., the second output power; It is a finely tuned short-term forecasting model; Input data for a short-term time-based prediction model.
[0096] S13. Determine the final output power based on the first output power and the second output power.
[0097] (11)
[0098] in, It is the final output power obtained by the two-fold fusion correction; It is the change in the first output power compared to the original output power; It is the change in the second output power compared to the original output power; ; , where are the weights of the two correction methods, respectively.
[0099] This invention utilizes a similar sample adjustment method, an online parameter fine-tuning method, and a bias correction method to correct the output power of the short-term prediction model in the second time period, specifically as follows:
[0100] S101. Determine the error between the predicted and measured values of wind power within a historical period, and denot it as the first error.
[0101] The embodiments of the present invention are in daily At any given moment, retrieve the most recent Hourly ultra-short-term error vector:
[0102] (12)
[0103] Error feature vectors pre-stored in the database of days with historical similarities:
[0104] (13)
[0105] S102. Determine the similarity between the first error and the power error.
[0106] (14)
[0107] in, , These are the weighting coefficients; The normalized scaling parameter is usually taken as... , The installed capacity is expressed in kilowatts (kW).
[0108] S103. Determine the output power of the short-term prediction model based on the similarity and the true power of the historical period, denoted as the third output power. Specifically, determine the weights in the short-term prediction model based on the similarity, and determine the output power of the short-term prediction model based on the true power and weights of the historical period.
[0109] This invention calculates the similarity of meteorological processes over the next 24 hours, adaptively fuses them according to the current error magnitude to obtain a comprehensive dynamic similarity, reorders the historical similar day database, and calculates dynamic weights using Softmax.
[0110] Meteorological similarity:
[0111] (15)
[0112] in For the first time in history The similarity of the weather between today and the present moment; It is the first Meteorological elements (wind speed, wind direction, air pressure, temperature, humidity, etc.); The weights are assigned to the corresponding elements (wind speed has the highest weight, and the sum is 1). This is a 24-hour weather forecast sequence. For the first time in history The corresponding time of day; This is the maximum value of the meteorological element in the entire dataset, used for normalization.
[0113] The overall dynamic similarity is calculated by considering the similarity between the first error and the power error. Similarity to weather It is obtained through confidence-weighted fusion, and its calculation formula is as follows:
[0114] (16)
[0115] in, For the first time in history Heaven is in response to the present Overall similarity at any given moment; for Confidence level at any given moment.
[0116] Dynamic weights:
[0117] (17)
[0118] in, The number of similar days selected.
[0119] Generate next day's power forecast:
[0120] (18)
[0121] in, This is the predicted power value after correction for similar dates. It is the first in history The actual power value at the corresponding moment.
[0122] S104. Determine the final output power based on the first output power, the second output power, and the third output power.
[0123] (19)
[0124] in, It is the final output power obtained by triple fusion correction; It is the change in the first output power compared to the original output power; It is the change in the second output power compared to the original output power; It is the change in the third output power compared to the original output power; ; , where are the weights of the three correction methods.
[0125] The short-term prediction model in this embodiment of the invention is updated at least once a day.
[0126] A second aspect of the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described short-term wind power prediction method.
[0127] A third aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described short-term wind power prediction method.
[0128] The effectiveness of the wind power short-term prediction method of the present invention will be verified below with more specific embodiments.
[0129] The method of this invention was verified using 15-minute resolution measured data from a 99MW wind farm in Northeast China from May 1, 2024 to May 20, 2025. The wind farm has an installed capacity of 99MW and a total of approximately 38,880 valid sampling points.
[0130] At a moderate level of 17.90% for the original short-term forecast normalized root mean square error (nRMSE), taking the March-April period of this wind field over 25 years as an example, the results are shown in Table 1 and... Figure 2 As shown.
[0131] Table 1 Comparison of results from March to April 2025
[0132] According to Table 1 and Figure 2 As can be seen, this invention achieves a stable 1.22% accuracy improvement through an ultra-short-term forecast feedback closed-loop mechanism, with the corrected nRMSE reaching 16.68%, fully verifying the error correction capability and engineering applicability of this invention under complex meteorological conditions.
[0133] This invention significantly improves both prediction accuracy and robustness. Traditional wind power prediction methods often treat ultra-short-term and short-term predictions as separate components, neglecting the multi-temporal and spatial scale characteristics between them. This invention achieves a closed-loop feedback method from ultra-short-term to short-term predictions through a multi-real-time feedback mechanism with unified confidence level control. The system can automatically select the most appropriate correction intensity and combination strategy based on the magnitude, direction, and trend of errors in the most recent 1-4 hours, effectively utilizing the characteristics of multiple temporal and spatial scales, improving the consistency and accuracy of predictions, and exhibiting high robustness to extreme weather conditions under the influence of dynamically similar samples.
[0134] The above descriptions are merely a few embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any modifications or alterations made by those skilled in the art without departing from the scope of the technical solution of the present invention using the disclosed technical content are equivalent to equivalent implementation cases and fall within the scope of the technical solution.
Claims
1. A method for short-term wind power forecasting, characterized in that, include: Step 1: Determine the power error between the predicted and measured values of wind power at each moment within the first time period; Step 2: Determine the dynamic bias at the corresponding time based on the power error; Step 3: Determine the confidence level at the corresponding time based on the dynamic bias; Step 4: Compare the confidence level with a preset threshold, and determine the correction method for the output power of the short-term prediction model in the second time period based on the comparison result; the second time period is longer than the first time period.
2. The short-term wind power forecasting method according to claim 1, characterized in that, Step 2 is as follows: The smoothing coefficient is determined based on the number of time nodes within the first time period; The dynamic bias at the current moment is determined based on the smoothing coefficient, the dynamic bias at the previous moment, and the power error at the current moment within the first time period.
3. The short-term wind power forecasting method according to claim 1, characterized in that, Step 3 specifically involves: Obtain the installed capacity and sensitivity coefficient of the wind farm; The confidence level at the corresponding time is determined based on the wind farm's installed capacity, sensitivity coefficient, and dynamic bias.
4. The short-term wind power forecasting method according to claim 1, characterized in that, Step 4 specifically involves: When the confidence level is less than or equal to the first threshold, the output power of the short-term prediction model in the second time period is corrected using the bias correction method. When the first threshold < the confidence level ≤ the second threshold, the output power of the short-term prediction model in the second time period is corrected using the online parameter fine-tuning method and the bias correction method; When the confidence level is greater than the second threshold, the output power of the short-term prediction model in the second time period is corrected using the similar sample adjustment method, the online parameter fine-tuning method, and the bias correction method.
5. The short-term wind power forecasting method according to claim 4, characterized in that, The output power of the short-term prediction model in the second time period is corrected using a bias correction method, specifically as follows: Determine the decay time coefficient for each moment in the second time period predicted by the short-term prediction model; The output power of the short-term prediction model in the second time period is corrected by the product of the decay time coefficient and the dynamic bias, and is denoted as the first output power.
6. The short-term wind power forecasting method according to claim 5, characterized in that, The output power of the short-term prediction model in the second time period is corrected using an online parameter fine-tuning method and the bias correction method, specifically as follows: Freeze the previous short-term prediction model The parameters of the layer are determined, and the loss function is determined based on the difference between the output power and the actual power of the short-term prediction model after the parameters are frozen; The parameters of the remaining layers of the short-term prediction model are updated using the loss function, and the output power of the short-term prediction model after updating the parameters is obtained, denoted as the second output power; The final output power is determined based on the first output power and the second output power.
7. The short-term wind power forecasting method according to claim 6, characterized in that, The output power of the short-term prediction model in the second time period is corrected using the similar sample adjustment method, the online parameter fine-tuning method, and the bias correction method, specifically as follows: The error between the predicted and measured values of wind power within a historical period is defined as the first error. Determine the similarity between the first error and the power error; The output power of the short-term prediction model is determined based on the similarity and the true power of the historical period, and is denoted as the third output power. The final output power is determined based on the first output power, the second output power, and the third output power.
8. The short-term wind power forecasting method according to claim 7, characterized in that, The output power of the short-term prediction model is determined based on the similarity and the true power over historical periods, specifically as follows: The weights in the short-term prediction model are determined based on the similarity. The output power of the short-term prediction model is determined based on the actual power over historical periods and the weights.
9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the short-term wind power prediction method according to any one of claims 1-8.
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 steps of the short-term wind power prediction method as described in any one of claims 1-8.