High-precision correction method and system for wind speed forecasting in power systems

Abstract

The present invention discloses a high-precision correction method and system for wind power forecasting of electric power systems, which belongs to the field of new energy forecasting. The method is first based on the WRF model and ARW dynamic solver, and selects a three-layer nested structure and a physical process parameterization scheme with the wind farm location as the regional center, and combines static terrain and global weather forecast data to achieve high-resolution short-term wind speed forecasting; then, by modifying the boundary layer parameters and introducing data assimilation technology to form a variety of differentiated schemes, multiple downscaled forecast results within the same time period are obtained; then, based on the forecast results and the wind speed observation data of the station at the corresponding time, a variety of different types of improved artificial intelligence models are used to realize the correction of the forecast wind speed; finally, a variety of improved models are combined with adaptive weighting to obtain the final wind speed forecast result. By fully mining the information of terrain, observation data and global weather forecast data, the accuracy of wind speed forecast data is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of new energy power prediction for electric power systems, and specifically relates to a high-precision correction method and system for forecasted wind speed for wind power prediction in electric power systems. Background Art

[0002] Against the backdrop of the transition to a low-carbon energy structure, renewable energy generation, represented by wind power, has developed rapidly and expanded in scale, playing an increasingly significant role in ensuring energy supply. However, the output of renewable energy generation is significantly affected by weather factors and exhibits significant uncertainty, posing a challenge to the safe and stable operation of the power system. The atmospheric system is inherently chaotic, and to fully exploit the advantages of wind resources, wind turbines are often located in areas with relatively complex meteorological conditions, such as mountaintops. This increases the difficulty of forecasting key factors influencing wind power output, such as wind speed. Currently, most wind speed forecasts rely on physical modeling, statistical analysis, or neural network fitting. The open-source WRF (Weather Research and Forecasting Model) model has been widely used in practical wind speed forecasting research. However, changes in topography caused by human activities are not always reflected in static geographic data, and the operation of wind turbines significantly affects atmospheric turbulence. Consequently, WRF's forecasts still have room for improvement. With the continuous increase in wind power installed capacity, the current wind speed forecast data with low spatial resolution and large deviation can no longer meet the accuracy requirements of wind power prediction in new power systems. There is an urgent need for a high-resolution and accurate forecasting method that can combine the actual observation information of wind farms. Summary of the Invention

[0003] To address the current issues of low resolution and large deviations in wind speed forecast data used in wind power forecasting, this paper proposes a high-precision correction method and system for wind speed forecasting in power systems. Based on the WRF model and the ARW (Advanced Research WRF) dynamic solver, this method utilizes multiple differentiated parameterization schemes and an artificial intelligence model that focuses on different aspects of the data. Combined with an ensemble learning strategy that adaptively assigns weights, this method fully exploits terrain, observational data, and global weather forecast data, significantly improving the accuracy of wind speed forecast data.

[0004] The technical solutions adopted in the present invention are as follows:

[0005] A high-precision correction method for wind speed forecast for power system wind power forecasting includes the following steps:

[0006] Step 1: Based on the WRF model and within the ARW dynamic solver framework, a three-layer nested structure and the number of vertical layers are selected with the latitude and longitude of the wind farm to be predicted as the regional center. The physical process parameterization scheme is selected in combination with the terrain location of the wind farm to minimize the forecast deviation.

[0007] Step 2: Based on static terrain data and Global Forecast System (GFS) meteorological data, set operating parameters such as forecast time and step size, and form a WRF dynamic downscaling model to achieve 24-72 hour WRF dynamic downscaling forecasts. According to the application requirements of wind power prediction in power systems, key meteorological information such as wind speed is extracted.

[0008] Step 3: Based on the WRF dynamic downscaling model constructed in Step 2, the boundary layer parameterization scheme is modified (which has a significant impact on wind speed forecasts) and the surface layer scheme is modified to match it. Furthermore, four-dimensional data assimilation (FDDA) technology is introduced to generate multiple differentiated parameterization schemes. Forecasts are performed under each differentiated parameterization scheme, ultimately obtaining multiple different wind speed forecasts for the same time period.

[0009] Step 4: Use the multiple wind speed forecast results within the same time period obtained in step 3 as the input of the correction model, and take the wind speed observation data at the corresponding moment at the hub height of the wind turbine at the site as the target to train multiple artificial intelligence models of different types. Different artificial intelligence models are then subjected to different performance tests based on the characteristics of the wind speed data to improve the accuracy, and multiple wind speed forecast correction models are obtained.

[0010] Step 5: Calculate the mean absolute error and root mean square error of the wind speed forecast results of different wind speed forecast correction models on the validation set in step 4. Based on the error results, use an ensemble learning strategy to adaptively assign different weights to different correction models. By weightedly combining multiple correction models, the final wind speed forecast result is obtained, thereby improving the accuracy and stability of the wind speed forecast results.

[0011] In the above technical solution, further, in step 1, since wind power and wind speed show an approximate cubic relationship, the nested structure and the vertical layer setting method need to take into account the high-precision wind speed forecast requirements for wind power prediction. Since the WRF model cannot dynamically adjust the resolution of the grid, using all high-level resolution grids will significantly reduce the efficiency of the calculation. Therefore, the present invention selects a three-layer nested structure, and uses coarse resolution to provide boundary conditions in the parent area of ​​the three-layer nested structure, and uses high resolution in the sub-area containing the wind farm to balance the timeliness of the calculation and the simulation accuracy. Unlike conventional methods, the accuracy selection of the high resolution should not only ensure matching with static terrain data, but also pay special attention to the actual location distribution of the wind turbine group in the wind farm, so that the wind turbine is located in the center of the grid as much as possible to effectively reduce the prediction deviation caused by the grid edge effect, and significantly improve the accuracy of wind speed and wind power prediction; at the same time, in the WRF model, The vertical layers of the coordinate system should be densely distributed at the actual hub height of the wind turbines in the wind farm to ensure sufficient vertical resolution near the hub height.

[0012] Furthermore, the selection of a physical process parameterization scheme must consider the general characteristics of wind farm location selection (mountainous areas, coastal areas) and be compatible with data such as static terrain. The present invention requires: enabling ocean physical processes to account for land-sea interactions; enabling slope radiation modulation and terrain wind field modulation to account for complex mountainous terrain and enhance wind farm simulation; configuring high-frequency radiation calls to accommodate high-temporal-resolution wind speed forecasts; and disabling cumulus convection parameterization in favor of explicit convection simulation. This approach avoids potential bias introduced by parameterization schemes while maintaining high horizontal resolution and capturing short-term wind speed fluctuations.

[0013] Furthermore, in order to meet the meteorological data input requirements for wind power forecasting, the WRF dynamic downscaling model uses a job scheduling system to efficiently schedule the parallel computing tasks of the WRF model and automatically submit the geogrid, ungrib, metgrid, real and wrf modules in the WRF model in sequence. After generating the forecast result file, the post-processing part extracts the wind speed component at the hub height of the wind turbine at the longitude and latitude of the wind farm. , And vector addition integration into horizontal wind speed, if the vertical layer If it is not completely consistent with the hub height of the wind turbine, the average of the results of two vertical layers near the hub height is used as the horizontal wind speed forecast result at the hub height.

[0014] Furthermore, in step 3, due to the chaotic nature of the atmospheric system, the boundary layer scheme is modified to introduce a variety of dynamic downscaling forecasts, thereby enhancing the robustness of the present invention's results. The boundary layer is closely related to surface friction, heat transport, and transpiration, and through internal turbulent motion, it enables the exchange of momentum, heat, water vapor, and other substances and energies between the surface and the atmosphere, influencing the evolution of the weather system. Therefore, the selection of boundary layer parameters significantly affects wind speed forecast results. However, different boundary layer schemes have different underlying theories and approaches to turbulence considerations, and they target significantly different forecast scenarios. Considering the wind speed characteristics that are most critical for wind power prediction, there are differences in scenarios such as daytime and nighttime, strong and weak winds, and sunny and cloudy days, with differences in whether the convective boundary layer is active or stable. If only one scheme is used for year-round wind speed and wind power forecasts, significant prediction errors will occur when the meteorological scenarios vary significantly. Therefore, when selecting multiple boundary layer schemes, the present invention needs to fully consider their adaptability to different weather conditions, seasons, and time periods. For example, the YSU scheme can be selected to improve the accuracy of severe convective weather simulations, while the QNSE scheme can be selected to improve the forecast accuracy for stable boundary layer scenarios, such as those at night. Furthermore, a more accurate boundary layer scheme can be selected for time periods with large wind power forecast deviations, providing sufficient differentiated information for subsequent correction and weighted averaging steps. Furthermore, to correct for the cumulative errors that may arise during long-term integration of large-scale fields such as wind fields, the present invention also introduces four-dimensional data assimilation (FDDA) technology (i.e., grid nudging). By adding a relaxation forcing term at each grid point, it gradually approximates the analysis field. To ensure the accuracy of wind speed forecasts at turbine hub height, the hourly shift of wind speed, temperature, and water vapor to the GFS reanalysis field is restricted to the outer boundary layer, preserving the turbulent characteristics within. By modifying the boundary layer and applying grid assimilation techniques, wind speed forecasts for multiple different WRF models after dynamic downscaling are obtained for the same time period.

[0015] Furthermore, in step 4, the artificial intelligence model uses the different wind speed forecast results in the same time period in step 3 as model input, and is trained with the wind speed observation data at the corresponding moment at the hub height of the wind turbine at the station as the target, to obtain the wind speed correction forecast results of each model. The wind speed correction for the wind farm to be predicted not only needs to consider the complex nonlinear relationship between the forecast results of different WRF models and the actual observation data, but also needs to explore the periodicity and dynamic evolution laws reflected by the historical observation wind speed data itself. In addition, traditional neural network models usually rely on a large amount of local data for training, while the observation data of a single wind farm is often limited, making it difficult to effectively capture the chaotic characteristics presented by the atmospheric system. The introduction of the pre-trained large model effectively solves the above problems and provides implicit information of various time series in the real world. Only a small amount of data is needed to achieve fine-tuning and quickly adapt to the wind speed correction task for the wind farm. The present invention preferentially employs three different AI models for training: a deep neural network (DBN), a long short-term memory network (LSTM), and a decoder-only time series prediction basic pre-training model (TimesFM). These models focus on extracting high-dimensional features, mining time series features, and providing general time series characteristic information, respectively. Based on these three trained AI correction models, three different wind speed data correction results can be obtained for the same time period.

[0016] Furthermore, the present invention improves the model based on the distribution characteristics of wind speed to improve the accuracy of the correction results. Unlike the normal distribution, wind speed usually presents a Weibull distribution with a long-tail characteristic, which is specifically manifested in that low wind speeds occur more frequently, while high wind speed extreme scenes occur less frequently but have a significant impact on wind power prediction. Therefore, in the present invention, for the two types of small models trained on local data, DBN and LSTM, an improved Huber loss is used to replace the mean square error loss function (MSE) and the sample weights are improved to better adapt to the wind speed correction function. Specifically, the Huber loss processes the absolute error between the predicted value and the true value in segments. When the error is less than or equal to the critical value, a square error form is used to provide a smooth gradient, while when the absolute error is greater than the critical value, a linear error form is used, reducing sensitivity to outliers. Integrated revised forecasts typically exhibit the characteristic of smoothed peak values. Based on the approximate cubic relationship between wind power and wind speed, wind power forecast deviations are particularly significant under high wind speed conditions. This invention assigns greater weight to sample data with wind speeds exceeding 10 meters per second, allowing the revised model optimization process to focus more on high wind speed areas and effectively enhancing the ability to fit high wind speed scenarios. For the pre-trained large model, TimesFM, covariates are integrated to aid prediction and a linear layer is added to the end of the model. Local wind speed data is used to post-process and correct the results, allowing for better adaptation to the wind speed forecast task for the wind farm being predicted, driven by data and model fusion.

[0017] Furthermore, the ensemble learning strategy adaptively assigns different weights to different correction models. Specifically, based on the mean absolute error (MAE) and root mean square error (RMSE) evaluation results of different correction models on the validation set, different normalized weights are assigned to the correction models, so that the correction models with relatively smaller error and variance are given higher weights. Finally, a weighted combination based on the weight calculation results is used to obtain the final wind speed forecast result. By dynamically assigning weights, the predictive advantages of each correction model are fully utilized, significantly improving the accuracy and stability of wind speed forecasts, and providing more reliable meteorological data input support for wind power forecasting in power systems.

[0018] The present invention also provides a high-precision wind speed forecast correction system for wind power prediction in electric power systems, which is used to implement any of the methods described above.

[0019] The present invention further provides an electronic device, comprising:

[0020] one or more processors;

[0021] a memory for storing one or more programs;

[0022] When the one or more programs are executed by the one or more processors, the one or more processors implement any of the above methods.

[0023] The present invention also provides a computer-readable storage medium storing computer-executable instructions, which are used to implement any of the above methods when executed.

[0024] The beneficial effects of the present invention are:

[0025] (1) The high-precision correction method and system for wind speed forecast for power system wind power prediction proposed in the present invention provides downscaled high-resolution wind speed forecast results based on the latitude and longitude of the selected wind farm to be predicted as the regional center, and at the same time, the correction is based on the observation data at the height of the wind turbine hub, which effectively improves the accuracy of the wind speed forecast.

[0026] (2) The present invention adopts the method of changing the boundary layer parameters and introducing data assimilation technology to obtain multiple forecast results after dynamic downscaling, and trains the artificial intelligence model and improves the model based on the wind speed and the characteristics of each model. The wind speed is corrected according to the observation information. By combining the rules of historical data and the advantages of multiple parameterization schemes, the wind speed prediction accuracy is significantly improved, which can provide more reliable meteorological forecast input support for the wind power prediction of the power system.

[0027] (3) The present invention adopts an ensemble learning method of adaptive weighted combination, which assigns weights to different correction models according to the size of the error, and can improve the stability and generalization ability of predictions in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The present invention is a flow chart of a method for high-precision correction of wind speed forecast for power system wind power prediction.

[0029] Figure 2 Flowchart of WRF dynamic downscaling for wind power forecasting in power systems.

[0030] Figure 3 This is the WRF dynamic downscaling wind speed forecast result under different boundary layer parameter schemes in one embodiment of the present invention.

[0031] Figure 4 This is a diagram of wind speed forecast results after correction of the improved DBN model in one embodiment of the present invention.

[0032] Figure 5 This is a graph of wind speed forecast results after correction of the improved LSTM model in one embodiment of the present invention.

[0033] Figure 6 This is a graph of wind speed forecast results after correction using the improved TimesFM model in one embodiment of the present invention. DETAILED DESCRIPTION

[0034] To make the present invention more clear and understandable, the following is a clear and complete description of an embodiment of the present invention in conjunction with the accompanying drawings. Obviously, the embodiment described below is only one embodiment of the present invention, not all embodiments. Based on the embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0035] The present invention proposes a high-precision correction method and system for wind speed forecast for power system wind power forecasting. According to a specific embodiment of the present invention, Figure 1 As shown, the method includes the following steps:

[0036] Step 1: Based on the WRF model and within the ARW dynamic solver framework, a three-layer nested structure and the number of vertical layers are selected with the latitude and longitude of the wind farm to be predicted as the regional center. The physical process parameterization scheme is selected in combination with the terrain location of the wind farm to minimize the forecast deviation.

[0037] Step 2: Based on static terrain data and Global Forecast System (GFS) meteorological data, set operating parameters such as forecast time and step size, and form a WRF dynamic downscaling model to achieve 24-72 hour WRF dynamic downscaling forecasts. According to the application requirements of wind power prediction in power systems, key meteorological information such as wind speed is extracted.

[0038] Step 3: Based on the WRF dynamic downscaling model constructed in Step 2, the boundary layer parameterization scheme is modified (which has a significant impact on wind speed forecasts) and the surface layer scheme is modified to match it. Furthermore, four-dimensional data assimilation (FDDA) technology is introduced to generate multiple differentiated parameterization schemes. Forecasts are performed under each differentiated parameterization scheme, ultimately obtaining multiple different wind speed forecasts for the same time period.

[0039] Step 4: Use the multiple wind speed forecast results within the same time period obtained in step 3 as the input of the correction model, and take the wind speed observation data at the corresponding moment at the hub height of the wind turbine at the site as the target to train multiple artificial intelligence models of different types. Different artificial intelligence models are then subjected to different performance tests based on the characteristics of the wind speed data to improve the accuracy, and multiple wind speed forecast correction models are obtained.

[0040] Step 5: Calculate the mean absolute error and root mean square error of the wind speed forecast results of different wind speed forecast correction models on the validation set in step 4. Based on the error results, use an ensemble learning strategy to adaptively assign different weights to different correction models. By weightedly combining multiple correction models, the final wind speed forecast result is obtained, thereby improving the accuracy and stability of the wind speed forecast results.

[0041] Specifically:

[0042] In step 1, the WRF dynamic downscaling method uses the ARW dynamic solver as the core to solve a set of fully compressible Eulerian non-static equations describing atmospheric motion. The time integration uses the Runge-Kutta algorithm. The horizontal direction uses the staggered Arakawa C grid points, and the vertical direction uses the terrain following sigma-pressure hybrid vertical method based on dry static balance. The coordinate system is as shown in formula (1).

[0043]

[0044] in is the static pressure of the fluid, represents the static pressure value of the fluid at the top boundary of the computational domain, Represents the static pressure value of the surface. Compared with the traditional numerical prediction method, the present invention The values ​​should be densely distributed at the actual hub height of the wind turbines in the wind farm to ensure sufficient vertical resolution near the hub height.

[0045] In addition, the appropriate projection method (Lambert projection is usually used in mid-latitude areas) and nested structure should be selected with the longitude and latitude of the wind farm as the center. Since the WRF model cannot dynamically adjust the resolution of the grid, using high-resolution grids for all grids will significantly reduce the efficiency of the calculation. The nested structure balances the timeliness of the calculation and the simulation accuracy by using coarse resolution to provide boundary conditions in the parent area and high resolution in the key sub-areas containing the wind farm. WRF also needs to select a parameterization scheme based on the location environment of the wind farm and the forecast requirements, which usually includes microphysical processes, long-wave radiation, short-wave radiation, etc.

[0046] Furthermore, in step 2, the step size and other operational parameters should be set based on the determined resolution and parameterization scheme, efficiently utilizing computing resources to achieve regionalized wind speed forecasts for wind farms. To meet the needs of short-term wind power forecasting for power systems, the Global Forecast System meteorological data uses UTC-12:00 as the benchmark, extracting forecast data for the next 24-72 hours as global field input data to provide initial and boundary conditions.

[0047] like Figure 2 As shown, the WRF model typically implements wind speed data forecasting through a WPS pre-processing module, a WRF main module, and post-processing modules such as NCL. The WPS module contains three submodules: the geogrid module takes static terrain data as input and inserts it into the grid points of the model's forecast area where the wind farm is located; the ungrib module takes the GFS global weather forecast as input, decompresses it, and extracts the required meteorological variables; and the metgrid module horizontally inserts the meteorological data output by ungrib into the model area grid established by geogrid, providing the output file to the main module. The main module is divided into the real module and the wrf module. The former uses the file generated by metgrid to insert it into the vertical grid to form the initial field and boundary conditions, while the latter performs numerical integration and generates forecast results. In one embodiment of the present invention, the Slurm job scheduling system is used to efficiently schedule WRF's parallel computing tasks and automatically submit the geogrid, ungrib, metgrid, real, and wrf modules in sequence. After the forecast result file is generated, the post-processing part extracts the two horizontal direction components of the wind speed at the hub height from the forecast result nc file generated by the wrf module, u and v variables, and adds them together to form the horizontal wind speed. If the height of the layer is not completely consistent with the hub height of the wind turbine, the average value of the two layers near the hub height is used to obtain the corresponding horizontal wind speed forecast result, which is used as the meteorological forecast input data for short-term wind power.

[0048] Furthermore, in step 3, due to the chaotic characteristics of the atmospheric system, the purpose of modifying the boundary layer scheme is to introduce a variety of dynamic downscaling forecasts to improve the robustness of the results of the present invention. The boundary layer is closely related to the friction resistance, heat transfer, transpiration and other effects of the surface, and realizes the exchange of material energy such as momentum, heat and water vapor between the surface and the atmosphere through internal turbulent motion, affecting the evolution of the weather system. Therefore, the selection of boundary layer parameters has a significant impact on the wind speed forecast results. However, the basic theories and ways of considering turbulence of different boundary layer schemes are different, and the forecast scenarios they target are also quite different.

[0049] Considering wind speed characteristics, the most critical for wind power prediction, there are differences in scenarios such as daytime and nighttime, strong winds and weak winds, and sunny and cloudy days, with active convective boundary layers and predominant stable boundary layers. If only one scheme is used for year-round wind speed and wind power forecasting, significant prediction errors will occur when the meteorological scenarios vary significantly. Therefore, when selecting multiple boundary layer schemes, the present invention needs to fully consider their adaptability to different weather conditions, seasons, and time periods. For example, the YSU scheme can be selected to improve the simulation accuracy of severe convective weather, while the QNSE scheme can be selected to improve the prediction accuracy of stable boundary layer scenarios such as nighttime. At the same time, a separate, more accurate boundary layer scheme can be selected for time periods with large wind power prediction deviations, thereby providing sufficient differentiated information for subsequent correction and weighted averaging steps. Taking the QNSE scheme as an example, it utilizes the quasi-normal scaling elimination theory, which can account for small-scale turbulence characteristics under stable atmospheric conditions. Its spectral space closure theory considers the combined effects of turbulence and waves. The corresponding QNSE surface layer scheme considers the higher-order effects of turbulence closure.

[0050] Furthermore, to correct for the potential cumulative errors that can arise from long-term integration of large-scale fields like wind fields, this paper incorporates grid nudging technology. By adding a relaxed forcing term at each grid point, the field gradually approximates the analysis field. To ensure the accuracy of wind speed forecasts at turbine hub height, this paper restricts the hourly shift of wind speed, temperature, and water vapor to the GFS reanalysis field to the outer boundary layer, preserving the turbulent characteristics within. By modifying the boundary layer and applying grid nudging, wind speed forecasts for multiple different WRF models, dynamically downscaled, are obtained for the same time period.

[0051] Furthermore, in step 4, different artificial intelligence correction models are trained based on the different wind speed forecast results within the same time period output in the previous step as model input, with the wind speed observation data at the corresponding moment at the hub height of the wind turbine at the station as the target, to obtain the wind speed correction forecast results of each model. The wind speed correction for the wind farm to be predicted not only needs to consider the complex nonlinear relationship between the forecast results of different WRF models and the actual observation data, but also needs to explore the periodicity and dynamic evolution laws reflected by the historical observation wind speed data itself. In addition, traditional neural network models usually rely on a large amount of local data for training, while the observation data of a single wind farm is often limited, making it difficult to effectively capture the chaotic characteristics presented by the atmospheric system. The introduction of the pre-trained large model effectively solves the above problems and provides implicit information of various time series in the real world. Only a small amount of data is needed to achieve fine-tuning and quickly adapt to the wind speed correction task for the wind farm.

[0052] Based on the above analysis, the correction method described in the present invention adopts three different artificial intelligence models for training, namely deep neural network (DBN), long short-term memory network (LSTM), and decoder-only time series prediction basic pre-training model (TimesFM), which respectively focus on high-dimensional feature extraction, time series feature mining, and providing time series general characteristic information.

[0053] Among them, DBN is composed of multiple RBM layers and has very powerful feature mining capabilities. The energy function expression between the input and output of each RBM is:

[0054]

[0055] in represents the hidden layer neurons, represents the visual layer neurons, 、 represents the bias coefficient, Represents the weight. The joint probability between the hidden layer and the visual layer neurons is expressed as:

[0056]

[0057] The calculation results of the RBM of the previous layer are used as the input vector of the RBM of the next layer to complete the unsupervised greedy layer-by-layer pre-training, and finally the back-propagation algorithm is used for supervised fine-tuning training. The LSTM network can fully mine the time series information of the wind speed data. It adds gate structures and cell memory units on the basis of RNN, solves the gradient vanishing problem in its long time series, and enables LSTM to apply information from a long time ago to the data processing at the current moment. TimesFM is a large pre-trained time series prediction model based on the Transformer architecture. It has 200 million parameters and has learned data from more than 300 billion time points. It has fully learned the implicit information of various time series in the real world. The performance of zero-sample evaluation on most data sets is close to or exceeds the most advanced supervised model. At the same time, it can handle different time series prediction lengths and time resolutions, and supports covariate input. The input of its Transformer layer It can be expressed as

[0058]

[0059] in represents the time series input, represents the mask vector, represents the Hadamard product, The input residual block is essentially a multi-layer perceptron with one hidden layer and skip connections. The model parameters are concentrated in the Transformer stack layer, each layer contains a multi-head self-attention mechanism and a feedforward network, and the output Expressed as

[0060]

[0061] The final output layer is also a multi-layer perceptron, with a time step arrive The prediction results are:

[0062]

[0063] In addition, the present invention improves the model according to the characteristics of wind speed data and adopts different improvement methods for different models to improve the accuracy of the correction results.

[0064] For the small models of DBN and LSTM trained on local data, the improved Huber loss is first used to replace the mean square error loss function (MSE) to better adapt to the wind speed correction function. The Huber loss is shown in the following formula, where represents the true value, represents the predicted value, is the hyperparameter of HuberLoss

[0065]

[0066] Specifically, the Huber loss performs segmented processing on the absolute error between the predicted value and the true value. When the error is less than or equal to a critical value, a squared error form is used to provide a smooth gradient, while when the absolute error is larger, a linear error form is used, reducing sensitivity to outliers. On the other hand, integrated revised forecasts typically exhibit the characteristic of smoothing peak values. Based on the approximate cubic relationship between wind power and wind speed, the predicted deviation of wind power is particularly significant under high wind speed conditions. In this invention, by assigning greater weight to sample data with wind speeds above 10 meters per second, the revised model optimization process increases its focus on high wind speed areas, effectively enhancing the fitting capability of high wind speed scenarios.

[0067] For the pre-trained large model TimesFM, covariates are integrated to assist prediction and a linear layer is added at the end of the model. Local wind speed data is used to post-process and correct the results, so as to better adapt to the wind speed forecast task of the wind farm to be predicted in a data-model fusion-driven manner.

[0068] Furthermore, in step 5, the adaptive weighted combination method belongs to an integrated learning strategy. Based on the mean absolute error (MAE) and root mean square error (RMSE) evaluation results of the three correction models on the validation set, different normalized weights are assigned to the three correction models, so that the prediction model with relatively smaller error and variance has a higher weight. Taking the MAE weight calculation as an example, the represents the number of samples, Indicates the number of models, represents the true value, represents the predicted value, represents the model weight, Represents the output of the revised model, and the MAE evaluation method is expressed as:

[0069]

[0070] The weights of each model are expressed as

[0071]

[0072] Final wind speed forecast results Expressed as:

[0073]

[0074] Similarly, the weight calculation using RMSE as the metric is performed, and the average of the two weights is used to obtain the final weighted result. By dynamically assigning weights, the forecasting advantages of each revised model are fully utilized, significantly improving the accuracy and stability of wind speed forecasts and providing more reliable meteorological data input support for wind power forecasting in power systems.

[0075] Example

[0076] A wind farm in South China is used as an example to describe the method in detail. The method for high-precision correction of wind speed forecast for wind power prediction in this wind farm includes the following steps:

[0077] Step 1: Build the WRF dynamic downscaling model. Using the Lambert projection, the wind farm location is used as the reference latitude and longitude. 437 grid points are set in the east-west direction and 480 in the north-south direction, with a grid resolution of 4 kilometers and 55 vertical layers. NASA MODIS data with a resolution of 30 arc seconds are used as static terrain data, while GFS0P25 data with a 3-hour input interval are used as global forecast data. The WSM6 scheme is used for microphysical processes, and the RRTMG scheme is used for longwave and shortwave radiation. The Noah land surface model is used. The integration step is set to 20 seconds, and the forecast time is set to 72 hours. This achieves high-resolution short-term wind speed forecasts using dynamic downscaling.

[0078] Step 2: Modify the boundary layer and near-surface layer parameters. Based on the WRF model described in step 1, select the MYJ, QNSE, and BouLac boundary layer schemes, and modify the corresponding Eta similarity theory, QNSE surface layer, and MM5 similarity theory near-surface layer schemes. Based on the BouLac scheme, grid assimilation technology is introduced, and the pushing intensity of the wind field is defined as 0.0003 per second. Four different WRF dynamic downscaling wind speed forecast results for the same time period are obtained as follows: Figure 3 shown.

[0079] Step 3: Correction using an improved AI model. Based on the wind speed forecast results obtained in Step 2 for the wind farm for a particular month using four WRF dynamic downscaling methods, these were combined with the corresponding wind speed data at the wind farm's hub height to form a complete dataset with a temporal resolution of 15 minutes. The training, validation, and test sets were divided in a ratio of 7:1.5:1.5. Three different AI models were trained: an improved deep neural network (DBN), an improved long short-term memory network (LSTM), and an improved decoder-only time series prediction basic pre-trained model (TimesFM). The corrected wind speed forecast results obtained for each of the three improved models are shown in Figures 4-6.

[0080] Step 4: Correct the results using an ensemble learning method that combines adaptive weighting methods. The average MAE and RMSE of the three models in the validation set were calculated using the weighting method described above. The weights of the three models were obtained as shown in Table 1. The final wind speed forecast results were obtained through weighted combination, and the error statistics are shown in Table 2.

[0081] Table 1 Validation set errors and weight distribution of three AI correction models

[0082]

[0083] Table 2 Statistics of final wind speed forecast error

[0084]

[0085] Results from specific examples demonstrate that the high-precision wind speed correction method and system designed by the present invention for power system wind power forecasting can provide high-resolution wind speed forecast information for selected wind farms, effectively improving the accuracy of forecast results by combining observed data. Furthermore, it can combine the advantages of different parameter schemes and correction models to improve the stability and generalization of forecasts.

[0086] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0087] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0088] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0089] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0090] The embodiments described above are merely some preferred embodiments of the present invention and are not intended to limit the present invention. Persons skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent substitution or equivalent transformation falls within the scope of protection of the present invention.

Claims

1. A high-precision correction method for wind speed forecast for power system wind power prediction, characterized by: The following steps are involved: Step 1: Based on the WRF model and within the ARW dynamic solver framework, a three-layer nested structure and the number of vertical layers are selected with the latitude and longitude of the wind farm to be predicted as the regional center. The physical process parameterization scheme is selected based on the topographic location of the wind farm to minimize the forecast bias. Step 2: Based on static terrain data and Global Forecast System (GFS) meteorological data, set the forecast time and step size operating parameters, and form a WRF dynamic downscaling model to achieve a 24-72 hour WRF dynamic downscaling forecast. The wind speed is extracted according to the application requirements of wind power prediction in the power system. Step 3: Based on the WRF dynamic downscaling model built in Step 2, modify the boundary layer parameterization scheme and the surface layer scheme to match it. In addition, introduce the four-dimensional data assimilation (FDDA) technology to form multiple differentiated parameterization schemes. Forecast under each differentiated parameterization scheme, and finally obtain multiple different wind speed forecast results for the same time period. Step 4: Use the multiple wind speed forecast results for the same time period obtained in step 3 as input to the correction model. Using the wind speed observation data at the corresponding moment at the hub height of the wind turbine at the site as the target, train multiple different types of artificial intelligence models. Based on the characteristics of the wind speed data, make different improvements to the different artificial intelligence models to obtain multiple wind speed forecast correction models. Step 5: Calculate the mean absolute error and root mean square error of the wind speed forecast results of different wind speed forecast correction models on the validation set in step 4. Based on the error results, use an ensemble learning strategy to adaptively assign different weights to different correction models. By weightedly combining multiple correction models, the final wind speed forecast result is obtained.

2. The high-precision correction method for wind speed forecast for power system wind power prediction according to claim 1 is characterized in that: In step 1, a coarse resolution is used to provide boundary conditions in the mother area of ​​the three-layer nested structure, and a high resolution is used in the sub-area containing the wind farm to balance the timeliness of the calculation and the simulation accuracy. In addition to ensuring the matching with the static terrain data, the accuracy selection of the high resolution should also consider the actual location distribution of the wind turbine group in the wind farm, so that the wind turbine is located in the center of the grid as much as possible. At the same time, in the WRF model, The vertical layers of the coordinate system should be densely distributed at the actual hub height of the wind turbines in the wind farm to ensure sufficient vertical resolution near the hub height.

3. The high-precision correction method for wind speed forecast for power system wind power prediction according to claim 1 is characterized in that: In step 1, the selection of the physical process parameterization scheme must meet the following requirements: enabling ocean physical processes to consider land-sea interaction, enabling slope radiation modulation and terrain wind field modulation, setting high-frequency radiation calls to adapt to high-time-resolution wind speed forecast requirements, and disabling cumulus convection parameterization and adopting explicit convection simulation.

4. The high-precision correction method for wind speed forecast for power system wind power prediction according to claim 1 is characterized in that: In step 2, the WRF dynamic downscaling model uses a job scheduling system to efficiently schedule the parallel computing tasks of the WRF model and automatically submits the geogrid, ungrib, metgrid, real and wrf modules in the WRF model in sequence. After generating the forecast result file, the post-processing part extracts the wind speed component at the hub height of the wind turbine at the longitude and latitude of the wind farm. , And vector addition integration into horizontal wind speed, if the vertical layer If it is not completely consistent with the hub height of the wind turbine, the average of the results of two vertical layers near the hub height is used as the horizontal wind speed forecast result at the hub height.

5. The high-precision correction method for wind speed forecast for power system wind power prediction according to claim 1 is characterized in that: In step 3, the modified boundary layer parameterization scheme is used to introduce diversified dynamic downscaling forecasts. When selecting multiple boundary layer parameterization schemes, it is necessary to fully consider the adaptability under different weather conditions, different seasons, and different time periods to provide sufficient differentiated information. In addition, the four-dimensional data assimilation technology is introduced to gradually approach the analysis field by adding a relaxation forcing term to each grid point. In the process of shifting the wind speed, temperature, and water vapor to the GFS reanalysis field every hour, the shift is limited only to the outside of the boundary layer, retaining the internal turbulence characteristics.

6. The high-precision correction method for wind speed forecast for power system wind power prediction according to claim 1, characterized in that: In step 4, the artificial intelligence model adopts three different types of artificial intelligence models: deep neural network DBN, long short-term memory network LSTM, and decoder-only time series prediction basic pre-training model TimesFM.

7. The high-precision correction method for wind speed forecast for power system wind power prediction according to claim 6, characterized in that: For the two types of artificial intelligence models, DBN and LSTM, their loss functions and sample weights are improved. Specifically, the improved Huber loss is used to replace the mean square error loss function MSE. The absolute error between the predicted value and the true value is segmented. When the error is less than or equal to the critical value, the square error form is used to provide a smooth gradient, and when the absolute error is greater than the critical value, the linear error form is used to reduce the sensitivity to outliers; by giving greater weight to sample data with wind speeds higher than 10 meters per second, the correction model optimization process pays more attention to high wind speed areas and enhances the fitting ability of high wind speed scenarios; for the pre-trained large model TimesFM, covariates are fused to assist in prediction and a linear layer is added at the end of the model. The local wind speed data is used to post-process and correct the results, so as to better adapt to the wind speed forecast task of the wind farm to be predicted in a data model fusion-driven manner.

8. The high-precision correction method for wind speed forecast for power system wind power prediction according to claim 1 is characterized in that: In step 5, the ensemble learning strategy is used to adaptively assign different weights to different correction models. Specifically, based on the mean absolute error (MAE) and root mean square error (RMSE) evaluation results of different correction models on the validation set, different correction models are assigned different normalized weights, so that the correction model with relatively smaller error and variance has a higher weight. Finally, the final wind speed forecast result is obtained by weighted combination based on the weight calculation results.

9. A high-precision wind speed correction system for wind power forecasting in power systems, characterized by: Used to implement the method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing computer-executable instructions, wherein the instructions are used to implement the method according to any one of claims 1 to 8 when executed.

Images