A multi-model fusion regional-level wind farm power prediction method and device

Abstract

This invention relates to the field of power prediction technology for wind, solar, and hydropower, and discloses a regional-level wind farm power prediction method and device using multi-model fusion. The prediction method effectively integrates meteorological information, wind turbine maintenance, and wind turbine failure factors to improve prediction accuracy. A regional-level wind farm power prediction model is constructed using a CNN-TCN-Transformer model. Through the collaborative work of multiple sub-models, the spatial dependence and feature interaction between wind farms are effectively captured, fully considering the spatial correlation between wind farms and improving the accuracy of wind power prediction. The introduction of beta distribution quantifies the uncertainties under extreme weather, wind turbine maintenance, and wind turbine failure scenarios, characterizing the probability distribution of prediction errors and improving the uncertainty handling capability of wind power prediction.

Description

Technical Field

[0001] This invention relates to the field of power prediction technology for water, wind, and solar power, specifically to a regional wind farm power prediction method and apparatus using multi-model fusion. Background Technology

[0002] As wind power accounts for an increasing proportion of the energy structure, regional-level wind power forecasting is of great significance for grid dispatch, energy planning, and power supply and demand balance. Currently, existing regional-level wind power forecasting methods mainly fall into two categories: one is to directly add up the output forecasts of each wind farm in the region to obtain the overall wind power forecast data; the other is to conduct correlation analysis or cluster analysis on the wind farms in the region to select representative typical wind farms, and then use the output forecast data of these typical wind farms as a basis to calculate the overall wind power forecast result.

[0003] However, the actual output of wind farms is affected by multiple factors, including meteorological information in the area where the wind farm is located, turbine maintenance, and turbine failures. Existing forecasting methods do not incorporate these factors into the forecasting system, and for wind power forecasting of a single wind farm, they cannot consider the spatiotemporal dependencies between wind farms in the region, lacking full utilization of existing information. This results in a large deviation between the forecast results and the actual output, and the forecast accuracy is insufficient to meet the actual needs of grid dispatching and energy management. Summary of the Invention

[0004] In view of this, the present invention provides a regional wind farm power prediction method and device with multi-model fusion to solve the problem that existing prediction methods do not incorporate comprehensive factors such as meteorological information, wind turbine maintenance and wind turbine failure in the wind farm area into the prediction system, resulting in a large deviation between the prediction results and the actual output.

[0005] In a first aspect, the present invention provides a regional-level wind farm power prediction method using multi-model fusion, the method comprising: Historical data of all wind farms within the regional scope were collected to construct a training dataset. The historical data included meteorological data, wind curtailment rate data, and power generation data. A regional wind farm power prediction model was constructed using the CNN-TCN-Transformer model, and the model was trained using the training dataset. For extreme weather, wind turbine maintenance and wind turbine failure factors, the beta distribution is used to describe the probability characteristics of each factor. Monte Carlo simulation is used to sample each uncertainty factor and calculate the disturbance factor to generate multiple disturbance sample paths. Multiple disturbance sample paths are integrated into the regional wind farm power prediction model to generate probabilistic prediction results. The model parameters are updated using a composite loss function and the AdamW optimizer until the model converges.

[0006] This invention provides a regional wind farm power prediction method using multi-model fusion, which effectively integrates meteorological information, wind turbine maintenance, and wind turbine failure factors to improve prediction accuracy. A regional wind farm power prediction model is constructed using a CNN-TCN-Transformer model. Through the collaborative work of multiple sub-models, the method effectively captures the spatial dependence and feature interactions between wind farms, fully considering the spatial correlation between wind farms and improving the accuracy of wind power prediction. The introduction of beta distribution quantifies the uncertainties under extreme weather, wind turbine maintenance, and wind turbine failure scenarios, characterizing the probability distribution of prediction errors and improving the uncertainty handling capability of wind power prediction.

[0007] In one alternative implementation, historical data from all wind farms within a regional scope are collected to construct a training dataset, including: Based on the meteorological data, the mean, standard deviation, maximum value, minimum value, and average wind direction vector of each meteorological variable for each wind farm within each week are calculated to obtain meteorological characteristics; The actual power generation and installed capacity of each wind farm are obtained weekly. Based on the actual power generation and installed capacity, the wind curtailment rate data is calculated, and a time series feature reflecting the weekly power curtailment intensity is constructed to obtain the power curtailment feature. Obtain weekly power generation data for each wind farm to determine wind power characteristics; The meteorological features, the power rationing features, and the wind power features are combined to construct a training dataset.

[0008] In one optional implementation, a regional wind farm power prediction model is constructed using a CNN-TCN-Transformer model, and the regional wind farm power prediction model is trained using a training dataset, including: The prediction model is constructed using the CNN-TCN-Transformer model with an encoder-decoder structure; The training dataset is subjected to dimensionality reduction and local feature extraction using a 1D CNN in the encoder; The temporal dependencies in the training dataset are captured by the TCN in the encoder; The Transformer encoder in the encoder captures the global spatial dependencies in the training dataset and outputs a context vector. The regional wind farm power prediction model is trained using the context vector and the actual total power generation of each wind farm in the region in the previous week as inputs through the Transformer decoder. This simulates the weekly predicted total power generation of each wind farm in the region and the weekly predicted total power generation of the region obtained by summing the results.

[0009] In one optional implementation, for factors such as extreme weather, wind turbine maintenance, and wind turbine failure, a beta distribution is used to describe the probabilistic characteristics of each factor. Monte Carlo simulation is used to sample and calculate the disturbance factor for each uncertainty factor, generating multiple disturbance sample paths, including: Based on beta distribution logic, the indexes for extreme weather, wind turbine maintenance and wind turbine failure factors are defined and the distribution parameters are calibrated to obtain the factor probability characteristics of each wind farm single-site dimension. Based on the probabilistic characteristics of factors at each station, the Monte Carlo simulation method was used to generate multiple sets of the same number of sampled values ​​for each station. Based on the disturbance factor formula and multiple sets of sampled values ​​for each individual wind farm, the comprehensive disturbance factor corresponding to each sample path for each wind farm is calculated. By combining the comprehensive perturbation factor with the features of the training dataset, multiple perturbation sample paths are generated.

[0010] In one optional implementation, multiple disturbance sample paths are integrated into the regional wind farm power prediction model to generate probabilistic prediction results. A composite loss function and the AdamW optimizer are used to update the model parameters until the model converges, including: Multiple disturbance sample paths are sequentially input into the regional wind farm power prediction model to obtain the prediction value for each single path. Statistical analysis is performed on the predicted total power generation values ​​of multiple single-path regions to calculate statistical characteristics and probability prediction results, including confidence intervals and probability distribution tables. Calculate the quantile loss between the predicted and actual total power generation for each path region, and then sum and average the quantile losses of multiple paths. Calculate the KL divergence between the fitted predicted distribution of total regional power generation and the preset beta prior distribution of total regional power generation, and multiply the KL divergence by the regularization coefficient to obtain the distribution regularization term. The quantile prediction loss is added to the distribution regularization term to obtain the composite loss function value for the current batch. The composite loss function value is fed into the AdamW optimizer, and all parameters of the regional wind farm power prediction model are updated through the backpropagation algorithm until the model converges.

[0011] In one optional implementation, the composite loss function is as follows:

[0012] in, N Indicates the number of sample paths. Representing a path i Forecast of total regional power generation Y This represents the actual total power generation capacity of the region. Represents the regularization coefficient. P The model predicts the distribution of the total power generation in the region. Q The beta prior distribution of the total power generation in the region is represented. This represents the KL divergence.

[0013] In one alternative implementation, for the first region k The first wind farm i The formula for the perturbation factor of the path is as follows:

[0014] in, Indicates the first in the region k Extreme weather power attenuation coefficient of a wind farm Indicates the first k The first wind farm i Sampled values ​​under wind turbine failure scenarios along a single path. Indicates the first k The first wind farm i Sampling values ​​in a wind turbine maintenance scenario with a single path. Indicates the first k The first wind farm i Sampled values ​​under extreme weather scenarios along the path.

[0015] Secondly, the present invention provides a regional wind farm power prediction device with multi-model fusion, the device comprising: The data acquisition module is used to collect historical data from all wind farms within the regional scope and construct a training dataset. The historical data includes meteorological data, wind curtailment rate data, and power generation data. The training module is used to construct a regional wind farm power prediction model using the CNN-TCN-Transformer model, and to train the regional wind farm power prediction model using the training dataset. The calculation module is used to describe the probability characteristics of each factor, such as extreme weather, wind turbine maintenance and wind turbine failure, using beta distribution. It samples each uncertainty factor and calculates the disturbance factor through Monte Carlo simulation, generating multiple disturbance sample paths. The prediction module is used to integrate multiple disturbance sample paths into the regional wind farm power prediction model to generate probabilistic prediction results. It uses a composite loss function and the AdamW optimizer to update the model parameters until the model converges.

[0016] This invention provides a regional wind farm power prediction device that integrates multiple models, effectively combining meteorological information, wind turbine maintenance, and wind turbine failure factors to improve prediction accuracy. It employs a CNN-TCN-Transformer model to construct a regional wind farm power prediction model. Through the collaborative work of multiple sub-models, it effectively captures the spatial dependencies and feature interactions between wind farms, fully considering the spatial correlation between wind farms and improving the accuracy of wind power prediction. Furthermore, it introduces beta distribution to quantify uncertainties under extreme weather, wind turbine maintenance, and wind turbine failure scenarios, characterizing the probability distribution of prediction errors and improving the uncertainty handling capability of wind power prediction.

[0017] Thirdly, the present invention provides a computer device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the regional wind farm power prediction method of multi-model fusion described in the first aspect or any corresponding embodiment.

[0018] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the multi-model fusion regional wind farm power prediction method described in the first aspect or any corresponding embodiment thereof. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a regional wind farm power prediction method using multi-model fusion according to an embodiment of the present invention. Figure 2 This is a structural block diagram of a regional wind farm power prediction device with multi-model fusion according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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] According to an embodiment of the present invention, a method for predicting regional wind farm power through multi-model fusion is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0023] This embodiment provides a regional wind farm power prediction method using multi-model fusion. Figure 1 This is a flowchart of a regional wind farm power prediction method using multi-model fusion according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps: Step S1: Collect historical data of all wind farms within the regional scope and construct a training dataset. The historical data includes meteorological data, wind curtailment rate data, and power generation data.

[0024] Specifically, historical data from all wind farms within a regional scope will be collected, with a data span of at least one year. This data will primarily cover meteorological data, wind curtailment rate data, and power generation data. "Region" can refer to a province, a city, or other areas. Feature extraction and organization will be performed on the meteorological, wind curtailment rate, and power generation data, using weekly time units, to form a training dataset. Meteorological data includes raw meteorological records of wind speed, temperature, and wind direction for each wind farm's location. Wind curtailment rate data includes the installed capacity and actual power generation of all wind farms within the region. The wind curtailment rate is calculated as follows: Wind curtailment rate (%) = (Theoretical power generation - Actual power generation) / Theoretical power generation * 100%. The theoretical power generation is calculated as follows: Theoretical power generation = Installed capacity * 24 hours * 7 days.

[0025] Furthermore, step S1 above includes: Step S11: Calculate the mean, standard deviation, maximum value, minimum value, and average wind direction vector of each meteorological variable for each wind farm within each week based on meteorological data to obtain meteorological characteristics.

[0026] Step S12: Obtain the actual power generation and installed capacity of each wind farm each week. Based on the actual power generation and installed capacity, calculate the wind curtailment rate data, construct a time series feature reflecting the weekly power curtailment intensity, and obtain the power curtailment feature.

[0027] Step S13: Obtain weekly power generation data for each wind farm to obtain wind power characteristics.

[0028] Step S14: Combine meteorological features, power rationing features, and wind power features to construct a training dataset.

[0029] In this embodiment of the invention, the training data preparation and feature extraction are completed according to the above steps S11-S14, and a training dataset containing at least 52 sets of samples is finally formed.

[0030] Step S2: Construct a regional wind farm power prediction model using the CNN-TCN-Transformer model, and train the regional wind farm power prediction model using the training dataset.

[0031] Specifically, step S2 above includes: Step S21: Use the CNN-TCN-Transformer model to build a prediction model with an encoder-decoder structure.

[0032] Step S22: Perform dimensionality reduction and local feature extraction on the training dataset using the 1D CNN in the encoder.

[0033] Step S23: Capture the temporal dependencies in the training dataset through the TCN in the encoder.

[0034] Step S24: Capture the global spatial dependencies in the training dataset and output the context vector through the Transformer encoder in the encoder.

[0035] Step S25: Using the context vector and the actual total power generation of each wind farm in the region in the previous week as input, the Transformer decoder is used to train the regional wind farm power prediction model, and the weekly predicted total power generation of each wind farm in the region is simulated and obtained by summing the weekly predicted total power generation of the region.

[0036] In this embodiment of the invention, the first 1D CNN uses 3 convolutional layers with a kernel size of 3 and an activation function of ReLU. The number of output channels is 64, 128, and 256 respectively. The input is the training dataset constructed in step S1 above, and the output is the local feature vector after dimensionality reduction.

[0037] The second TCN layer contains four residual blocks, employs dilated convolution, and uses GELU as the activation function to maintain the sequence length. The input is the local feature vector output by the 1D CNN, and the output is a feature sequence that captures long-term dependent features such as seasonal wind patterns and monthly power rationing trends.

[0038] The third-layer Transformer encoder is designed with 3 layers, 8 multi-head attention heads, and 256 hidden layer dimensions. The input is the feature sequence output by the TCN. The self-attention mechanism captures the global spatial relationships such as meteorological linkage and power correlation between different wind farms in the region. The final output is a context vector (or feature sequence) that can represent the comprehensive information of spatiotemporal-operation-policy.

[0039] The decoder employs a Transformer decoder, whose parameters are matched with those of the Transformer encoder. During the training phase, the Transformer decoder takes the context vector output by the encoder plus the actual total power generation of the previous week as input, and outputs the predicted value for the corresponding week in the training / validation set. During the prediction phase, the Transformer decoder takes the context vector output by the encoder plus the actual / predicted total power generation of the previous week as input, and iteratively calculates and outputs the predicted total power generation within the regional scope for the current week.

[0040] The model parameters are configured as follows: the Adam optimizer is used in the early stage of training, and the learning rate adopts a warmup + decay strategy; the batch size is set according to the GPU memory; the number of training epochs must meet the model convergence, and an early stopping strategy is used to prevent overfitting; overfitting is further suppressed by dropout layers and weight decay.

[0041] Step S3: For extreme weather, wind turbine maintenance and wind turbine failure factors, beta distribution is used to describe the probability characteristics of each factor. Monte Carlo simulation is used to sample each uncertainty factor and calculate the disturbance factor to generate multiple disturbance sample paths.

[0042] Specifically, step S3 above includes: Step S31: Based on beta distribution logic, define the indicators and calibrate the distribution parameters for extreme weather, wind turbine maintenance and wind turbine failure factors respectively, and obtain the factor probability characteristics of each wind farm single station dimension.

[0043] Step S32: Based on the factor probability characteristics of each station, the Monte Carlo simulation method is used to generate multiple sets of the same number of sampled values ​​for each station.

[0044] Step S33: Based on the disturbance factor formula and multiple sets of sampled values ​​for each wind farm, calculate the comprehensive disturbance factor corresponding to each sample path for each wind farm.

[0045] Step S34: Combine the integrated perturbation factor with the features of the training dataset to generate multiple perturbation sample paths.

[0046] In this embodiment of the invention, the first [item] within the region k One wind farm ( , K (The total number of wind farms in the region), and based on their historical operating data, define quantitative indicators for extreme weather, wind turbine maintenance, and wind turbine failure respectively: Extreme weather index: Percentage of weeks with extreme weather at a single weather station within a year , =Number of extreme weather weeks per station / 52, with a value range of [0,1]; Fan failure index: Weekly failure rate of fans at a single site , The total number of wind turbine failure hours in a week at the station / (number of wind turbines in the station × 7 days × 24 hours), with a value range of [0,1]. Wind turbine maintenance index: Percentage of weekly downtime caused by wind turbine maintenance at a single wind turbine site , =Total weekly downtime due to turbine maintenance at the wind turbine site / (Number of turbines at the wind turbine site) 7 days (24 hours), with a value range of [0,1].

[0047] Based on the historical statistical mean and variance of the above indicators for each station, the shape parameter of the beta distribution was calibrated. and This yields the factor probability characteristics for each individual station dimension, namely: , , .

[0048] At the same time, based on the terrain characteristics, wind turbine characteristics, and meteorological sensitivity of each individual wind farm, a differentiated extreme weather power attenuation coefficient is set for each wind farm. This is used for subsequent calculations of disturbance factors for individual stations.

[0049] In this embodiment, for a certain station k Assume the percentage of weeks with extreme weather p k Obey Beta ( =12.6, =50.4) Distribution (the percentage of weeks with extreme weather is 20% with a variance of 0.0025); assume the weekly failure rate of the wind turbine. f k Obey Beta ( =3, =96.7) distribution (corresponding to a weekly failure rate of mean 3% and variance of 0.000289); assume the percentage of weekly downtime due to fan maintenance. m k Obey Beta ( =2, =48.2) distribution (corresponding to a weekly downtime rate of approximately 4%, with a variance of 0.00075).

[0050] Specifically, the beta distribution is typically used to describe random variables in the interval [0,1], and its probability density function is:

[0051] in, and For shape parameters, This is the beta function. In this embodiment, the beta distribution is used to describe the probability characteristics of each factor.

[0052] For a beta distribution, the mean and variance are related to the parameters. , The relationship is:

[0053]

[0054] Inversely derived from mean and variance and :

[0055]

[0056] The probability beta distribution parameters for extreme weather events are calculated as follows: According to this embodiment, for a certain station k， The average percentage of extreme weather events is 20%, with a variance of 0.0025, meaning: mean

[0057] variance

[0058] The calculation yielded:

[0059]

[0060] The calculation of beta distribution parameters for wind turbine maintenance and wind turbine failure factors is the same as above.

[0061] The criteria for determining an extreme weather week are as follows: Extreme values ​​and fluctuations of meteorological elements such as wind speed, temperature, precipitation, and wind direction are statistically analyzed on a weekly basis. A week is defined as an extreme weather week if any of the following occurs: wind speed exceeding the cutoff threshold, calm winds, extreme temperature changes, heavy precipitation, or a drastic change in wind direction. Specific indicators are as follows: Extreme wind speeds: Maximum wind speed ≥25 m / s per hour within a week, or minimum wind speed ≤2 m / s; Extreme temperatures: A maximum temperature of ≥38℃ or a minimum temperature of ≤-10℃ within a week; Heavy precipitation / blizzard: Maximum hourly rainfall ≥ 30 mm, or blizzard weather may occur; Dramatic changes in wind direction: The circumferential variance of wind direction within one week is >0.8, indicating continuous and drastic fluctuations in wind direction.

[0062] If any of the above conditions are met, the week is determined to be an extreme weather week.

[0063] The circumferential variance of wind direction is calculated as follows: a. Take the hourly wind direction angle over a week. , , ..., (Unit: degree) b. Convert to unit circle coordinates

[0064]

[0065] c. Find the average vector

[0066]

[0067] d. Find the average vector length

[0068] e. Wind direction circumferential variance Wind direction circumferential variance = 1 - R Balancing computational efficiency and scene coverage, the total number N of disturbance sample paths to be generated is determined. For each wind farm within the region, based on its beta distribution probability characteristics at the single-site dimension, N independent Monte Carlo sampling operations are performed to obtain N sets of sample values ​​for each wind farm: a. No. k Extreme weather sampling values ​​for each wind farm: , ,..., ( i =1,2,...,N, i (For the perturbation sample path number), each The first in the corresponding area k The first wind farm i The probability of extreme weather occurring along each path.

[0069] b. The k Wind turbine fault sampling values ​​for each wind farm: , ,..., Each The first in the corresponding area k The first wind farm i The probability of wind turbine failure scenarios occurring along a specific path.

[0070] c. k Wind turbine maintenance sampling values ​​for a wind farm: , ,..., Each The first in the corresponding area k The first wind farm i The probability of a wind turbine along a certain path being in a maintenance scenario.

[0071] d. The number of sampled values ​​for all wind farms is consistent with the total number of regional disturbance sample paths N, and the Monte Carlo sampling process for each individual wind farm is independent of each other.

[0072] For the region of the first k The first wind farm, its first i Group sample values Substituting into the disturbance factor formula, calculate the first disturbance factor of the wind farm. i Single-site disturbance factor corresponding to each sample path The perturbation factor formula is:

[0073] in, Indicates the first in the region k Extreme weather power attenuation coefficient of a wind farm Indicates the first k The first wind farm i Sampled values ​​under wind turbine failure scenarios along a single path. Indicates the first k The first wind farm i Sampling values ​​in a wind turbine maintenance scenario with a single path. Indicates the first k The first wind farm i Sample values ​​for extreme weather scenarios along the path.

[0074] Using the method described above, the single-site disturbance factor for each wind farm under N paths within the region is calculated, resulting in N disturbance factors for each wind farm. ,..., For each wind farm, the disturbance factor is used. The meteorological characteristics and power rationing characteristics were corrected respectively.

[0075] It is a "degradation factor" that integrates the effects of faults, maintenance, and extreme weather. Therefore, for positive factors affecting power generation (the greater the influencing factor, the greater the power generation, such as wind speed), the correction logic is usually "original feature ×". (This represents the degree of attenuation after the characteristic is affected). For negative factors affecting power generation, the correction logic is usually "original characteristic / (This represents the degree of attenuation of the characteristic after it is affected). For some specific factors, such as the wind direction average vector, no correction is made.

[0076] Step S4: Integrate multiple disturbance sample paths into the regional wind farm power prediction model to generate probabilistic prediction results. Use a composite loss function and AdamW optimizer to update the model parameters until the model converges.

[0077] Specifically, step S4 above includes: Step S41: Input multiple disturbance sample paths into the regional wind farm power prediction model in sequence to obtain the predicted value of each single path.

[0078] In this embodiment of the invention, a regional wind farm power prediction model is invoked, and the input features of all disturbance sample paths are sequentially input into the model to predict all paths and store the predicted values ​​of all paths.

[0079] Step S42: Perform statistical analysis on the predicted total power generation of multiple single-path regions, calculate statistical characteristics and probability prediction results, including confidence intervals and probability distribution tables.

[0080] In this embodiment of the invention, statistical analysis is performed on the predicted total power generation of N single-path regions. The predicted values ​​corresponding to commonly used quantiles are calculated, and based on a beta distribution, the predicted total power generation of the N single-path regions is fitted to a beta distribution to obtain the probability density curve of the predicted total power generation of the regions. Then, based on the quantile predicted values, 90% confidence intervals are output; and the probability of occurrence of each power interval is calculated according to the probability density curve of the predicted total power generation of the regions to obtain a probability distribution table.

[0081] Step S43: Calculate the quantile loss between the predicted and actual total power generation of each path region, and then sum and average the quantile losses of multiple paths. This average value reflects the deviation between the predicted and actual total power generation of the region.

[0082] In this embodiment of the invention, the predicted total regional power generation value for each path is... and actual power value Y The error is measured using Pinball loss (quantile loss), as shown in the following formula:

[0083] It is the target quantile, and its value range is... , The larger the value, the more severely the loss function penalizes cases where the actual value is greater than the predicted value. The smaller the quantile, the heavier the penalty for the actual value being less than the predicted value. The value is entirely determined by the wind power probability prediction target. The specific determination method can be divided into three scenarios, which are selected based on business needs and statistical logic: Scenario 1: Output a specified confidence interval, for example, the goal is to output a 90% confidence interval. The value of directly corresponds to the lower and upper bounds of the confidence interval. , At this point, it is necessary to calculate the lower and upper bound quantile losses simultaneously and take the average, so that the model can fit the predicted total power generation values ​​of the region at both ends.

[0084] Scenario 2: If more attention is paid to a certain type of extreme risk (such as insufficient power supply due to low power output, or wind curtailment due to excessive power output), a single-sided quantile can be selected: Pay attention to the risk of low power output (such as a sharp drop in output due to extreme weather). choose , For these small quantiles, the model will focus on fitting the predicted value at the "low power end" to reduce the risk of underestimating the power.

[0085] Pay attention to high-power risks (such as wind turbines operating at full capacity leading to wind curtailment). choose , With these large quantiles, the model will focus on fitting the predicted value at the "high power end" to reduce the risk of overestimating power.

[0086] Scenario 3: Quantile distribution based on historical data Select the quantile interval with the largest fluctuations from historical data as... For example, if the total power generation in a historical region fluctuates most dramatically at the 0.05 and 0.95 quantiles, then these two quantiles should be selected so that the model can focus on learning the pattern in this interval. The losses of the lower and upper quantiles should be calculated simultaneously and averaged so that the model can fit the power prediction values ​​at both ends.

[0087] Step S44: Calculate the KL divergence between the fitted predicted distribution of total regional power generation and the preset beta prior distribution of total regional power generation, and multiply the KL divergence by the regularization coefficient to obtain the distribution regularization term.

[0088] In this embodiment of the invention, the prior distribution of the beta of the total regional power generation is obtained by fitting historical data on the total regional power generation. Its probability density function is:

[0089] , For shape parameters. It is a Beta function. It is a gamma function.

[0090] Predicted distribution of total power generation in the fitted region: Model output N Each path prediction value is typically fitted to another beta distribution. Its probability density function is:

[0091] It is a Beta function. It is a gamma function.

[0092] Gamma function It is a generalization of factorial to the real and complex number domains, and its calculation method varies depending on the type of the independent variable (i.e., the value inside the parentheses). Common calculation scenarios are as follows: Scenario 1: Positive Integers n Use the factorial formula directly:

[0093] Scenario 2: Half-integer

[0094]

[0095] Scenario 3: General Real Numbers

[0096] For non-integer and non-half-integer cases, numerical methods or table lookups are usually used, such as using Stirling's formula or Lanczos approximation. In practical applications, calculations are almost always performed directly using specialized software, such as MATLAB, Python, or R.

[0097] KL divergence measures the difference between these two "power distributions," thus constraining the predicted distribution to a reasonable range to closely approximate the actual power. The smaller the KL divergence, the closer the "predicted power generation distribution" is to the "historical prior power distribution," and the more the model output conforms to the real-world scenario.

[0098] The KL divergence formula for the fitted predicted distribution of total regional power generation and the preset beta prior distribution of total regional power generation is as follows:

[0099] in, It is a Beta function. It is the digamma function (the logarithmic derivative of the gamma function, i.e.) ).

[0100] Step S45: Add the quantile prediction loss to the distribution regularization term to obtain the composite loss function value for the current batch.

[0101] In this embodiment of the invention, the composite loss function is as follows:

[0102] in, N Indicates the number of sample paths; Representing a path i Forecast of total regional power generation; Y This represents the actual total power generation capacity of the region. Represents the regularization coefficient; P The model predicts the distribution of the total power generation in the region; Q The beta prior distribution of the total power generation in the region is represented. This represents the KL divergence; Pinball is the "Pinball loss function" (also called the quantile loss function), a loss function specifically used for probability prediction / quantile regression.

[0103] Regularization coefficient Its function is to balance the weights of "quantile loss (the deviation between predicted and actual values)" and "KL divergence (the deviation between predicted and prior distributions)". Its value needs to be determined in conjunction with business requirements and model validation results. There are three common methods: Method 1: Empirical values ​​based on business requirements Based on the priorities of "prediction accuracy" and "distribution rationality", empirical values ​​are set directly: If more emphasis is placed on the closeness of the predicted value to the actual value (quantile loss has higher priority): take the smaller value. ,for example ; If greater emphasis is placed on the consistency between the predicted distribution and the prior distribution (to avoid unreasonable distributions in the model output): take the larger value. ,for example .

[0104] Method 2: Cross-validation Optimal results were selected using validation set performance metrics (MAE, RMSE, 90% confidence interval coverage). The following is a specific example: a. Prepare a set of candidates Value, for example ; b. For each candidate The model was trained and metrics (MAE, RMSE, 90% confidence interval coverage) were computed on the validation set. c. Select the smallest value that simultaneously satisfies "MAE ≤ 6%, RMSE ≤ 8%, and 90% confidence interval coverage ≥ 88%". : Method 3: Grid Search / Random Search (Automatic Parameter Tuning) Automatically find the optimal solution using optimization algorithms : Grid search: in Within the candidate interval (e.g., 0.001~2), iterate with a fixed step size (e.g., 0.01) to select the candidate with the best validation set index. ; Random search: Randomly sample several items within the interval. The value is selected based on the optimal validation set metric. .

[0105] Step S46: Input the composite loss function value into the AdamW optimizer, and update all parameters of the regional wind farm power prediction model through the backpropagation algorithm until the model converges.

[0106] In this embodiment of the invention, the composite loss function value C is passed to the AdamW optimizer, and all parameters of the CNN-TCN-Transformer model are updated through the backpropagation algorithm to minimize the composite loss function, so that the model prediction results are closer to the true values.

[0107] Every 5 epochs of training, the validation set from the perturbation sample path is input into the model, and the composite loss function value C of the validation set is calculated. 验证 If the validation set loss C over 20 consecutive epochs... 验证 There was no decrease, and the training set loss C 训练If the model tends to stabilize and simultaneously meets the following validation criteria: mean absolute error (MAE) ≤ 6%, root mean square error (RMSE) ≤ 8%, and 90% confidence interval coverage ≥ 88%, then the model is considered to have converged and parameter updates are stopped.

[0108] This invention provides a regional wind farm power prediction method using multi-model fusion, which effectively integrates meteorological information, wind turbine maintenance, and wind turbine failure factors to improve prediction accuracy. A regional wind farm power prediction model is constructed using a CNN-TCN-Transformer model. Through the collaborative work of multiple sub-models, the method effectively captures the spatial dependence and feature interactions between wind farms, fully considering the spatial correlation between wind farms and improving the accuracy of wind power prediction. The introduction of beta distribution quantifies the uncertainties under extreme weather, wind turbine maintenance, and wind turbine failure scenarios, characterizing the probability distribution of prediction errors and improving the uncertainty handling capability of wind power prediction.

[0109] This embodiment also provides a regional-level wind farm power prediction device with multi-model fusion, which is used to implement the above embodiments and preferred embodiments, and will not be repeated as already described. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0110] This embodiment provides a regional-level wind farm power prediction device with multi-model fusion, such as... Figure 2 As shown, it includes: The data acquisition module 21 is used to collect historical data of all wind farms within the regional scope and construct a training dataset. The historical data includes meteorological data, wind curtailment rate data, and power generation data.

[0111] Training module 22 is used to construct a regional wind farm power prediction model using the CNN-TCN-Transformer model and to train the regional wind farm power prediction model using the training dataset.

[0112] The calculation module 23 is used to describe the probability characteristics of each factor, such as extreme weather, wind turbine maintenance and wind turbine failure, using beta distribution. It samples each uncertainty factor and calculates the disturbance factor through Monte Carlo simulation, generating multiple disturbance sample paths.

[0113] Prediction module 24 is used to integrate multiple disturbance sample paths into the regional wind farm power prediction model to generate probabilistic prediction results. It uses a composite loss function and AdamW optimizer to update model parameters until the model converges.

[0114] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0115] In this embodiment, the regional wind farm power prediction device with multi-model fusion is presented in the form of functional units. Here, a unit refers to an ASIC (Application Specific Integrated Circuit), a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0116] This invention provides a regional wind farm power prediction device with multi-model fusion, which effectively integrates meteorological information, wind turbine maintenance, and wind turbine failure factors to improve prediction accuracy. It employs a CNN-TCN-Transformer model to construct a regional wind farm power prediction model. Through the collaborative work of multiple sub-models, it effectively captures the spatial dependence and feature interaction between wind farms, fully considering the spatial correlation between wind farms and improving the accuracy of wind power prediction. Furthermore, it introduces beta distribution to quantify the uncertainties under extreme weather, wind turbine maintenance, and wind turbine failure scenarios, characterizing the probability distribution of prediction errors and improving the uncertainty handling capability of wind power prediction.

[0117] This invention also provides a computer device having the above-described features. Figure 2 The regional wind farm power prediction module shown is a multi-model fusion module.

[0118] Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 3 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 3 Take a processor 10 as an example.

[0119] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0120] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0121] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0122] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0123] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0124] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0125] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A regional wind farm power prediction method using multi-model fusion, characterized in that, The method includes: Historical data of all wind farms within the regional scope were collected to construct a training dataset. The historical data included meteorological data, wind curtailment rate data, and power generation data. A regional wind farm power prediction model was constructed using the CNN-TCN-Transformer model, and the model was trained using the training dataset. For extreme weather, wind turbine maintenance and wind turbine failure factors, the beta distribution is used to describe the probability characteristics of each factor. Monte Carlo simulation is used to sample each uncertainty factor and calculate the disturbance factor to generate multiple disturbance sample paths. Multiple disturbance sample paths are integrated into the regional wind farm power prediction model to generate probabilistic prediction results. The model parameters are updated using a composite loss function and the AdamW optimizer until the model converges.

2. The regional wind farm power prediction method based on multi-model fusion according to claim 1, characterized in that, Historical data from all wind farms within the regional scope were collected to construct a training dataset, including: Based on the meteorological data, the mean, standard deviation, maximum value, minimum value, and average wind direction vector of each meteorological variable for each wind farm within each week are calculated to obtain meteorological characteristics; The actual power generation and installed capacity of each wind farm are obtained weekly. Based on the actual power generation and installed capacity, the wind curtailment rate data is calculated, and a time series feature reflecting the weekly power curtailment intensity is constructed to obtain the power curtailment feature. Obtain weekly power generation data for each wind farm to determine wind power characteristics; The meteorological features, the power rationing features, and the wind power features are combined to construct a training dataset.

3. The regional wind farm power prediction method based on multi-model fusion according to claim 1, characterized in that, A regional wind farm power prediction model is constructed using the CNN-TCN-Transformer model. The model is trained using a training dataset, including: The prediction model is constructed using the CNN-TCN-Transformer model with an encoder-decoder structure; The training dataset is subjected to dimensionality reduction and local feature extraction using a 1D CNN in the encoder; The temporal dependencies in the training dataset are captured by the TCN in the encoder; The Transformer encoder in the encoder captures the global spatial dependencies in the training dataset and outputs a context vector. The regional wind farm power prediction model is trained using the context vector and the actual total power generation of each wind farm in the region in the previous week as inputs through the Transformer decoder. This simulates the weekly predicted total power generation of each wind farm in the region and the weekly predicted total power generation of the region obtained by summing the results.

4. The regional wind farm power prediction method based on multi-model fusion according to claim 2, characterized in that, For factors such as extreme weather, wind turbine maintenance, and wind turbine failure, a beta distribution is used to describe the probabilistic characteristics of each factor. Monte Carlo simulation is used to sample and calculate the disturbance factor for each uncertainty factor, generating multiple disturbance sample paths, including: Based on beta distribution logic, the indexes for extreme weather, wind turbine maintenance and wind turbine failure factors are defined and the distribution parameters are calibrated to obtain the factor probability characteristics of each wind farm single-site dimension. Based on the probabilistic characteristics of factors at each station, the Monte Carlo simulation method was used to generate multiple sets of the same number of sampled values ​​for each station. Based on the disturbance factor formula and multiple sets of sampled values ​​for each individual wind farm, the comprehensive disturbance factor corresponding to each sample path for each wind farm is calculated. By combining the comprehensive perturbation factor with the features of the training dataset, multiple perturbation sample paths are generated.

5. The regional wind farm power prediction method based on multi-model fusion according to claim 1, characterized in that, Multiple disturbance sample paths are integrated into the regional wind farm power prediction model to generate probabilistic prediction results. A composite loss function and the AdamW optimizer are used to update the model parameters until the model converges, including: Multiple disturbance sample paths are sequentially input into the regional wind farm power prediction model to obtain the prediction value for each single path. Statistical analysis is performed on the predicted total power generation values ​​of multiple single-path regions to calculate statistical characteristics and probability prediction results, including confidence intervals and probability distribution tables. Calculate the quantile loss between the predicted and actual total power generation for each path region, and then sum and average the quantile losses of multiple paths. Calculate the KL divergence between the fitted predicted distribution of total regional power generation and the preset beta prior distribution of total regional power generation, and multiply the KL divergence by the regularization coefficient to obtain the distribution regularization term. The quantile prediction loss is added to the distribution regularization term to obtain the composite loss function value for the current batch. The composite loss function value is fed into the AdamW optimizer, and all parameters of the regional wind farm power prediction model are updated through the backpropagation algorithm until the model converges.

6. The regional wind farm power prediction method based on multi-model fusion according to claim 5, characterized in that, The composite loss function is as follows: in, N Indicates the number of sample paths. Representing a path i Forecast of total regional power generation Y This represents the actual total power generation capacity of the region. Represents the regularization coefficient. P The model predicts the distribution of the total power generation in the region. Q The beta prior distribution of the total power generation in the region is represented. This represents the KL divergence.

7. The regional wind farm power prediction method based on multi-model fusion according to claim 4, characterized in that, For the region of the first k The first wind farm i The formula for the perturbation factor of the path is as follows: in, Indicates the first in the region k Extreme weather power attenuation coefficient of a wind farm Indicates the first k The first wind farm i Sampled values ​​under wind turbine failure scenarios along a single path. Indicates the first k The first wind farm i Sampling values ​​in a wind turbine maintenance scenario with a single path. Indicates the first k The first wind farm i Sampled values ​​under extreme weather scenarios along the path.

8. A regional-level wind farm power prediction device with multi-model fusion, characterized in that, The device includes: The data acquisition module is used to collect historical data from all wind farms within the regional scope and construct a training dataset. The historical data includes meteorological data, wind curtailment rate data, and power generation data. The training module is used to construct a regional wind farm power prediction model using the CNN-TCN-Transformer model, and to train the regional wind farm power prediction model using the training dataset. The calculation module is used to describe the probability characteristics of each factor, such as extreme weather, wind turbine maintenance and wind turbine failure, using beta distribution. It samples each uncertainty factor and calculates the disturbance factor through Monte Carlo simulation, generating multiple disturbance sample paths. The prediction module is used to integrate multiple disturbance sample paths into the regional wind farm power prediction model to generate probabilistic prediction results. It uses a composite loss function and the AdamW optimizer to update the model parameters until the model converges.

9. A computer device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the regional wind farm power prediction method based on multi-model fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the regional wind farm power prediction method of multi-model fusion as described in any one of claims 1 to 7.

Images