A mountain wind power probability prediction method fusing fast-powerformer and bayes

Abstract

The application discloses a kind of mountain wind power probability prediction methods of fusing Fast-Powerformer and bayes, it is related to wind power generation power prediction technical field, it solves the technical problems that existing mountain wind farm microclimate gradient is strong, traditional model is insufficiently adaptable to complex terrain, probability prediction lacks the explanation of uncertainty source.One, collect mountain wind farm multi-source data;Second, obtain wind turbine operating characteristics according to mountain wind farm multi-source data;Third, generate mountain three-dimensional microclimate field characteristics according to terrain parameters;Fourth, obtain terrain adaptability input tensor by fusing terrain embedding features, historical power characteristics, time characteristics and optional weather forecast coding features;Fifth, output power prediction mean and conditional variance of future prediction period;Sixth, total prediction variance is obtained by sampling terrain embedding Fast-Powerformer model based on bayesian posterior approximation technique, and is decomposed into accidental uncertainty and cognitive uncertainty.The application is suitable for complex mountain wind farm.

Description

Technical Field

[0001] This invention relates to the field of wind power generation prediction technology, and more specifically, to a probabilistic prediction method for mountain wind power that integrates Fast-Powerformer and Bayesian methods. Background Technology

[0002] With the large-scale grid connection of new energy sources, wind power forecasting has become an important technical foundation for power system dispatch planning, reserve capacity arrangement, spot market application, medium- and long-term transactions, and new energy consumption assessment. Compared with plain wind farms, mountain wind farms are located in more complex terrains, with significant differences in elevation, slope aspect, ridge acceleration, valley channel effect, leeward eddies, enhanced local turbulence, and wake superposition within the farm area. These phenomena result in significant spatial differences in wind speed, wind direction, turbulence intensity, and air density among different wind turbines within the same wind farm.

[0003] Traditional wind power forecasting methods mainly include physical models, statistical models, machine learning models, and ordinary deep learning models. Physical models typically rely on numerical weather prediction and wind turbine power curves, offering some mechanistic explanation, but they are computationally expensive, have limited update frequency, and are insufficient for correcting complex local terrain. Statistical models, such as ARIMA and SARIMA, have some fitting ability for linear trends and periodic terms, but they struggle to characterize the strong nonlinearity, non-stationarity, and multivariate coupling of wind power output. Traditional machine learning models, such as support vector regression, random forests, and gradient boosting trees, can learn nonlinear mappings, but they typically treat wind power forecasting as a static regression problem, failing to adequately characterize long-term dependencies, wind speed ramp-up, sudden wind direction changes, and multivariate interactions.

[0004] In recent years, deep learning models such as LSTM, GRU, CNN-LSTM, Transformer, Informer, and Reformer have been increasingly used for wind power prediction. Transformer-type models can model long-range dependencies through attention mechanisms, but the attention complexity of a standard Transformer is [missing information - likely OCR errors]. O(L 2 ) When the input history window is long and the prediction step size is wide, the computational cost and memory usage increase significantly. Efficient Transformers such as Informer and Reformer reduce the cost of long sequence modeling, but when directly applied to wind power prediction, they still have problems such as insufficient extraction of local short-term features, inadequate expression of cross-variable dependencies, and insufficient utilization of the periodic frequency domain features of wind power.

[0005] For mountain wind farms, time series models alone are insufficient to reflect the impact of topographic factors on wind speed shear, wind direction deflection, turbulence enhancement, and air density changes. Slope, aspect, curvature, surface roughness, valley passages, ridgelines, and shading factors not only affect the local wind field but also alter the upstream-downstream relationship between adjacent turbines, thereby influencing power aggregation at the farm level. Existing models, lacking topographic embedding and micrometeorological field reconstruction, are prone to significant prediction errors in the windward side of ridges, leeward slopes, valley mouths, and complex saddle areas.

[0006] On the other hand, wind power forecasting results require not only accurate point predictions but also reliable probability intervals and risk interpretations. Existing interval or quantile forecasts typically provide a total uncertainty but fail to further distinguish between accidental uncertainties caused by the inherent randomness of wind resources, sensor noise, and power curve dispersion, and cognitive uncertainties caused by insufficient samples, uncertain model parameters, complex terrain extrapolation, and distribution offsets. For scheduling and trading, accidental uncertainty usually represents irreducible risk, while cognitive uncertainty can be reduced by increasing samples, supplementing measurement points, updating models, or conducting active learning; the two have different engineering implications. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a probabilistic prediction method for mountain wind power that integrates Fast-Powerformer and Bayesian methods to address the shortcomings of existing technologies. This method solves the technical problems of strong micro-meteorological gradients in existing mountain wind farms, insufficient adaptability of traditional models to complex terrain, large computational burden for long-sequence prediction, and lack of explanation of uncertainty sources in probabilistic prediction.

[0008] The present invention discloses a probabilistic prediction method for mountain wind power that integrates Fast-Powerformer and Bayesian methods. This method includes: Step 1: Collect multi-source data from mountain wind farms; Step 2: Process the multi-source data of the mountain wind farm to obtain the wind turbine operating characteristics; Step 3: Obtain terrain parameters and generate three-dimensional micro-meteorological field characteristics of the mountain area based on the terrain parameters; Step 4: Obtain historical power features, time features, and optional weather forecast coding features. Convert the wind turbine operation features, mountain three-dimensional micro-meteorological field features, and terrain parameters into terrain embedding features. Fuse the terrain embedding features with historical power features, time features, and optional weather forecast coding features to obtain the terrain adaptability input tensor. Step 5: Input the terrain adaptability input tensor into the terrain embedding Fast-Powerformer model, and the terrain embedding Fast-Powerformer model outputs the power prediction mean and conditional variance for the future prediction period; Step 6: Based on the Bayesian posterior approximation technique, the terrain embedding Fast-Powerformer model is sampled multiple times to obtain the total prediction variance. According to the law of total variance, the total prediction variance is decomposed into random uncertainty and cognitive uncertainty. Step 7: Generate power prediction intervals and power generation prediction results based on the power prediction mean, total prediction variance, random uncertainty, and cognitive uncertainty; Step 8: Combine equipment availability, power curtailment status, predicted deviation power, and electricity trading price parameters to obtain conditional risk value and trading risk index. Based on the conditional risk value and trading risk index, output the multi-timescale power, power generation, prediction range, and risk assessment results of the mountain wind farm.

[0009] To further improve the process, in step two, the multi-source data is sequentially subjected to quality control, anomaly identification, missing data repair, time synchronization, spatial registration, normalization processing, and reliability assessment to form wind turbine operating characteristics.

[0010] Furthermore, in step three, the method for generating the three-dimensional micro-meteorological field characteristics of the mountain based on the terrain parameters is as follows: Meteorological data sources, physical correction terms, and neural residual correction terms are obtained. Based on the meteorological data sources, physical correction terms, and neural residual correction terms, the terrain parameters are reconstructed into three-dimensional micro-meteorological field features of the mountain by multi-source meteorological interpolation, physical correction, and neural residual correction.

[0011] Furthermore, in step four, the terrain embedding features include static terrain features, incoming flow direction features, and micro-meteorological gradient features.

[0012] Furthermore, in step five, the terrain embedding Fast-Powerformer model includes a lightweight LSTM embedding module, an input transpose mechanism module, a frequency-enhanced channel attention module, and a Reformer basic backbone network module.

[0013] Furthermore, in step five, the expressions for the power prediction mean and conditional variance are: ; ; Where μ is the power prediction mean, σ 2 Here, h is the conditional variance, h is the prediction step size, and θ is the model parameter. P_{i,t+h} For the i-th wind turbine in the future t+h The power random variable at time t.

[0014] Furthermore, in step six, the method for decomposing the total prediction variance into random uncertainty according to the law of total variance is to obtain S posterior samples. Θ s For the S posterior samples Θ s The total prediction variance TUvar is decomposed into random uncertainty AUvar and random uncertainty EUvar using Bayesian posterior prediction variance decomposition. The expression for decomposing the random uncertainty AUvar is as follows: ; in, Θ s Let S be the posterior sample. Θ s ; The expression for the cognitive uncertainty EUvar, decomposed into, is: ; The expression for calculating the total prediction variance TUvar is as follows: ; Where TUvar is the total prediction variance.

[0015] Furthermore, in step seven, the expression for calculating the power prediction interval is: ; Among them, PI γ (x) represents the prediction interval at confidence level γ, z γ These are the standard normal quantiles or the calibrated empirical quantiles.

[0016] Furthermore, in step three, the terrain parameters include slope, aspect, curvature, relative elevation difference, surface roughness, ridgeline, valley passage, windward and leeward relationship, and terrain shading factor.

[0017] Furthermore, in step one, the multi-source data of the mountain wind farm includes micro-meteorological observation data of the station, wind turbine SCADA operation data, wind turbine static parameters, digital elevation model data, wind measurement tower data, remote sensing wind data, meteorological reanalysis data, optional regional meteorological large model forecast data, optional numerical weather prediction data, historical climate data, and atmospheric circulation index.

[0018] Beneficial effects The advantages of this invention are: 1. This invention explicitly embeds mountain terrain factors into the power prediction model, enabling the model to perceive the impact of slope, aspect, curvature, shading, valley passages, and ridge acceleration on wind power output. Compared with models that rely solely on historical power or single weather forecasts, this is more suitable for complex mountain wind farms.

[0019] 2. This invention employs a Fast-Powerformer architecture consisting of lightweight LSTM embedding, input transpose, LSH attention, reversible residuals, and frequency-enhanced channel attention. This architecture can reduce computational complexity and memory consumption while maintaining long sequence prediction capabilities, making it suitable for field-side deployment and rolling prediction.

[0020] 3. This invention decomposes the total prediction uncertainty into random uncertainty and cognitive uncertainty through Bayesian posterior prediction variance decomposition, so that the prediction interval not only has probabilistic meaning, but also has the ability to explain the source of risk, making it easier to judge whether the error mainly comes from natural randomness or insufficient sample coverage, model extrapolation or distribution shift.

[0021] 4. This invention establishes a consistency correction mechanism among short-term power forecasting, medium-term power forecasting, and medium-to-long-term power generation forecasting, enabling the forecast results to simultaneously serve real-time dispatching, day-ahead planning, power generation assessment, and electricity market transactions.

[0022] 5. This invention offers flexibility in terms of meteorological forecast input. For stations equipped with high-quality large-scale meteorological models or numerical weather prediction capabilities, these can be used as optional external forecast features; for scenarios with insufficient meteorological forecast resources or high real-time requirements, forecasts can also be primarily based on station micrometeorological observations, historical power data, terrain embedding, and short-term rolling windows. Attached Figure Description

[0023] Figure 1 This is a flowchart of the overall process of the mountain wind power probability prediction method that integrates Fast-Powerformer and Bayesian methods of the present invention.

[0024] Figure 2 This is a schematic diagram of the Fast-Powerformer model structure for terrain embedding in this invention.

[0025] Figure 3 This is a flowchart of the Bayesian posterior prediction variance decomposition and probability interval generation process of this invention.

[0026] Figure 4 This is a flowchart of the multi-timescale power generation prediction and consistency correction process of the present invention. Detailed Implementation

[0027] The present invention will be further described below with reference to embodiments, but this does not constitute any limitation on the present invention. Any limited modifications made by any person within the scope of the claims of the present invention are still within the scope of the claims of the present invention.

[0028] See Figures 1-4 This invention provides a probabilistic prediction method for mountain wind power that integrates Fast-Powerformer and Bayesian methods, such as... Figure 1As shown, the method includes: Step 1: Collect multi-source data from the mountain wind farm. This multi-source data includes station micro-meteorological observation data, wind turbine SCADA operation data, wind turbine static parameters, digital elevation model data, anemometer data, remote sensing anemometer data, meteorological reanalysis data, optional regional meteorological large-scale model forecast data, optional numerical weather prediction data, historical climate data, and atmospheric circulation index. In Step 1, the station micro-meteorological observation data includes at least one or more of the following: wind speed, wind direction, temperature, humidity, air pressure, air density, turbulence intensity, wind shear index, precipitation, boundary layer height, or near-surface stability. The wind turbine SCADA operation data includes at least one or more of the following: active power, reactive power, rotor speed, pitch angle, yaw angle, nacelle wind speed, turbine status, fault status, power curtailment status, availability, and historical alarm records. This wind farm comprises multiple wind turbine generators distributed along ridges, slopes, saddles, and near valley passages, with significant elevation differences within the area. Some turbines are located on the windward slope or ridge acceleration zone, while others are located on the leeward slope, in a shaded area, or in a wake superposition zone. The wind farm is equipped with environmental monitoring instruments, meteorological towers, wind turbine SCADA systems, and power metering devices, and is capable of acquiring digital elevation models, meteorological reanalysis data, optional weather forecast data, historical power generation data, and atmospheric circulation indices.

[0029] A multi-source data sample library was established for model training, calibration, validation, and online prediction. The library is organized into three levels: wind turbine level, wind farm level, and regional climate level. Wind turbine-level data is used to characterize the power response of individual wind turbines; wind farm-level data is used for power aggregation and grid connection point calibration; and regional climate-level data is used for medium- and long-term power generation prediction. Wind turbine-level data includes active power, reactive power, nacelle wind speed, rotor speed, pitch angle, yaw angle, unit status, fault status, power curtailment status, availability, and historical alarm information for each wind turbine. For normally operating samples, power data serves as the model training label; for samples experiencing fault shutdowns, maintenance, communication anomalies, and severe power curtailment, power data is not directly used as the label for normal power prediction, but its status information is retained for availability correction and risk assessment.

[0030] The micrometeorological data from the stations include wind speed, wind direction, temperature, humidity, air pressure, air density, turbulence intensity, wind shear index, precipitation, boundary layer height, and near-surface stability. For anemometer data, it is preferable to retain wind speed and direction at multiple height levels; for environmental monitoring instrument data, it is preferable to retain continuous observation results at the minute or 5-minute level; for remote sensing anemometer data, it is preferable to retain vertical wind profile characteristics.

[0031] Step Two: Perform quality control, anomaly identification, missing data repair, time synchronization, spatial registration, normalization, and reliability assessment on the multi-source data to form a wind turbine-level multi-source standardized feature set. In Step Two, the reliability score of the multi-source data... q s,i,t It is determined by the missing forecast rate, historical rolling forecast error, release time delay, and spatial representativeness error. ; ; Where s is the s-th type of data source, i is the i-th wind turbine or grid point, and t is the time. For the missing test rate, Historical rolling error, Δt s,i,t The time difference between the release time and the target time. Δd s,i For spatial representativeness error, α 1 to α 4 It is a non-negative adjustment coefficient. ω s,i,t To correspond to the fusion weights, q s,i,t Assess credibility.

[0032] After data collection is completed, the multi-source data undergoes unified processing. First, data from different sampling frequencies are standardized to the required time scale for prediction. For second- or minute-level data, moving averages, maximum / minimum value statistics, and volatility statistics can be used to generate 5-minute or 15-minute features. For hourly forecast data, linear interpolation, spline interpolation, or keeping the most recent forecast value can be used to map it to the power prediction timeline. For daily, weekly, and monthly scale data, aggregation is performed according to the statistical period.

[0033] Next, anomaly identification is performed. Anomalies are identified if wind speed is less than zero, wind direction exceeds 0° to 360°, power is less than zero, power exceeds the unit's rated power, or temperature or air pressure significantly exceeds the reasonable range for the local climate. For power data, a joint judgment is made based on unit status, power curtailment status, and grid connection point power constraints to avoid misinterpreting power reductions caused by unnatural wind conditions as a meteorological-power mapping relationship.

[0034] For missing data, processing is based on the duration of the missing data and the importance of the data source. Short-term missing data can be filled by interpolation between adjacent time points, interpolation of similar measurement points, interpolation of similar states of adjacent wind turbines, or model estimation; long-term missing data are not directly used as normal samples for filling, but the credibility weight of the data source in the corresponding time period is reduced.

[0035] Step 3: Based on the digital elevation model and the spatial distribution of wind turbines, extract slope, aspect, curvature, relative elevation difference, surface roughness, ridgeline, valley passage, windward / leeward relationship, and topographic shading factor to construct a three-dimensional micro-meteorological field reconstruction model for the mountainous area. In Step 3, the described three-dimensional micro-meteorological field for the mountainous area... M(r,z,t) Reconstruction is achieved through multi-source meteorological interpolation, physical correction, and neural residual correction. ; Where r is the spatial position, z is the height, and X is the height. s For the s-th type of meteorological data source, I S (·) For spatial interpolation or downscaling operators, Φ phy (·) Let T be the set of terrain parameters, which includes boundary layer wind profiles, terrain acceleration, terrain shading, valley passages, and stability corrections. ε θ (·) This is a neural residual correction term. M(r,z,t) This describes the three-dimensional micro-meteorological field characteristics of mountainous areas.

[0036] In step three, the hub height wind speed is obtained by adjusting the stability-corrected wind profile, the terrain residual term, and the terrain gating coefficient: ; ; in, For the first Wind speed at typhoon hub height Let z be the wind speed at height z. H i For wheel hub height, d 0 Zero planar displacement height z 0 Let L be the roughness length, L be the Moning-Obukhov length, and ψ be the roughness length. m For stability correction function, si, ai,ci,bi,θi These are slope, aspect, curvature, shading factor, relative elevation difference, and incoming wind direction. f ter This is the terrain residual correction function. This is the terrain gating coefficient.

[0037] Digital elevation model (DEM) data is used to extract static terrain features. For each wind turbine, slope, aspect, curvature, relative elevation difference, roughness, valley passage, ridgeline, terrain shading, windward / leeward relationship, and upstream terrain obstacle features are extracted within a neighborhood of at least 300 to 2000 meters. The actual range can be adjusted according to turbine spacing, prevailing wind direction, and site complexity.

[0038] Meteorological forecast data is an optional input. When the station has reliable regional meteorological large model output, numerical weather prediction, or refined meteorological forecast services, it can be used as an external forecast feature. When meteorological forecasts are missing or have large errors, the model can make rolling predictions based on historical micro-meteorological windows, station observations, topographic embedding, and short-term autoregressive features.

[0039] Medium- and long-term power generation forecast data include daily, weekly, monthly, or quarterly power generation statistics, effective wind hours, average wind speed, equipment availability, maintenance hours, power curtailment, average air density, major weather events, and atmospheric circulation indices. These atmospheric circulation indices include the El Niño-Southern Oscillation Index, the Arctic Oscillation Index, the Western Pacific Subtropical High Index, the monsoon index, or regional wind resource anomaly indicators.

[0040] After the above-mentioned treatment, a standardized sample set is formed, including micro-meteorological characteristics, operational status characteristics, terrain embedding characteristics, historical power characteristics, optional weather forecast characteristics, and time characteristics.

[0041] Based on a digital elevation model, topographic analysis of the wind farm area is performed. First, the area is divided into several grids according to the site boundaries and wind turbine layout, and a multi-scale neighborhood is established with each wind turbine as the center. The small-scale neighborhood is used to describe the slopes around the wind turbine and local shading, the medium-scale neighborhood is used to describe the valley channels and ridge orientation, and the large-scale neighborhood is used to describe the topographic undulations in the direction of the dominant inflow.

[0042] For each wind turbine, extract parameters such as slope, aspect, curvature, relative elevation difference, roughness, ridgeline, valley passage, windward slope, leeward slope, and terrain shielding. For a given incoming flow direction, search the terrain profile along the upstream direction to determine whether the wind turbine is located in a mountain shielding area, valley passage area, ridge acceleration area, or leeward disturbance area.

[0043] When a wind turbine is located on the windward side of a ridge and the angle between the incoming flow direction and the slope aspect is small, it is determined that it is strongly affected by the acceleration of the ridge. When a wind turbine is located in a narrow valley and the wind direction is basically consistent with the valley direction, it is determined that it is strongly affected by the acceleration of the valley passage. When there is a terrain obstacle upstream that is higher than or close to the hub height, it is determined that it is strongly affected by the obstruction. When a wind turbine is located on the leeward slope or in a slope turning area, it is determined that it is strongly affected by the enhancement of turbulence and the deflection of wind direction.

[0044] In the horizontal direction, local corrections are made to wind speed, wind direction, turbulence intensity, and air density by taking into account factors such as slope, aspect, curvature, terrain shading, and valley passage.

[0045] Step 4: The three-dimensional micro-meteorological field features, wind turbine operation features, and terrain parameters of the mountainous area are converted into terrain embedding features, and then fused with historical power features, time features, and optional weather forecast coding features to obtain the terrain-adaptive input tensor. In Step 4, the terrain embedding features are composed of static terrain features, incoming flow direction features, and micro-meteorological gradient features. ; ; in, For the first i Typhoon machine t Terrain embedding vector at time step, Microclimate characteristics, For operational status characteristics, As a time feature, These are optional weather forecast coding features; when weather forecast data is not available, It can be set to an empty vector or replaced by historical micro-meteorological window encoding features.

[0046] After obtaining the three-dimensional micro-meteorological field, terrain embedding features are further constructed. These features include both static terrain attributes at the wind turbine location and dynamic terrain responses related to the current incoming flow direction. For the same wind turbine, the upstream terrain differs under different wind directions; therefore, the weight of terrain influence also varies over time.

[0047] For site-level forecasting, the input tensors of all wind turbines are organized into three-dimensional tensors according to turbine number, spatial location, or upstream-downstream topology. For single-turbine forecasting, the input is the feature set of the wind turbine and its neighboring measurement points, adjacent wind turbines, and meteorological grid points.

[0048] The terrain-embedded Fast-Powerformer is the core power prediction model of this invention. This model is used to output the mean, conditional variance, and site-level aggregated power prediction results for wind turbine-level power over the next 5 minutes to 72 hours, and can also be extended to medium-term power predictions of 1 to 3 days.

[0049] like Figure 2As shown, step five involves embedding the terrain-adaptive input tensor into a Fast-Powerformer, using a lightweight LSTM embedding to extract short-term local dynamics, modeling cross-variable dependencies using an input transpose mechanism, reducing the complexity of long-sequence modeling using locality-sensitive hashing attention and reversible residual structures in the Reformer backbone network, extracting periodic wind resource fluctuation features using frequency-enhanced channel attention, and outputting the power prediction mean and conditional variance for the future prediction period. In step five, the lightweight LSTM embedding layer of the Fast-Powerformer terrain embedding is used to extract short-term local dynamic features of the input sequence. ; ; Where B is the batch size, T is the historical window length, d_h is the LSTM hidden state dimension, and d model For the model dimension, w_e and b_e are linear projection parameters. In step five, the input transpose mechanism converts time step tokens into variable tokens to enhance the correlation modeling among multivariate meteorological factors, topographic factors, and operational status factors. ; ; Where V represents the number of variables and T represents the time length; since V is usually less than T, the input transpose mechanism reduces the number of effective tokens for attention computation and improves the ability to model cross-variable dependencies. In step five, the Reformer backbone network uses locality-sensitive hash attention, mapping queries and keys to hash buckets, and calculating attention only within the same or adjacent buckets: ; ; Where P_i is the first... The effective attention set corresponding to each query, h (r) (·) represents the local sensitive hash function for the r-th round, where R is the number of hash rounds; the local sensitive hash attention reduces the complexity of standard self-attention from O(L...) to O(L...) 2 The value is reduced to approximately O(LlogL).

[0050] In step five, the Reformer backbone network further employs a reversible residual structure to reduce the intermediate activation storage during long sequence training: ; in, X 1 and X 2 For the two components of the hidden state, Y1 and Y 2 The output components are reconstructed through reversible computation during backpropagation, thus reducing memory usage. In step five, the frequency enhancement channel attention module uses discrete cosine transform to extract periodic components from wind power, wind speed, and terrain correction features, and generates channel weights. ; in, V i For the time series features of the i-th channel, B i These are the basis functions of the discrete cosine transform. W 1 and W 2 Here are the parameters of the fully connected layer, δ(·) is the non-linear activation function, σ(·) is the sigmoid function, and Z is the Reformer output hidden state. final This is the feature representation after frequency domain enhancement.

[0051] In step five, the terrain embedding Fast-Powerformer outputs the power prediction mean μ and conditional variance σ. 2 This forms a heteroscedastic two-end predictor: ; ; Where μ is the power prediction mean, σ 2 Here, h is the conditional variance, h is the prediction step size, and θ is the model parameter. P_{i,t+h} For the i-th wind turbine in the future t+h The power random variable at time t.

[0052] The model first extracts short-term local dynamics through a lightweight LSTM embedding layer.

[0053] After the input is transposed, the trajectory of each variable within the historical window is treated as a variable token. Attention is primarily calculated among wind speed, wind direction, temperature, humidity, air pressure, turbulence intensity, terrain factors, and operational status factors. Since the number of variables V is usually significantly smaller than the time length T, this mechanism can both improve the ability to model cross-variable coupling and reduce the computational overhead of attention.

[0054] In the long sequence dependency modeling part, the model adopts a Reformer backbone network. This structure reduces the standard self-attention complexity from O(L...) 2 The cost is reduced to approximately O(LlogL), making it suitable for longer historical windows and multi-step prediction scenarios.

[0055] Through the aforementioned reversible relationship, the input can be reconstructed from the output during backpropagation, thereby reducing intermediate activation storage. For site-side servers or edge computing devices, this architecture helps reduce deployment costs.

[0056] A frequency-enhanced channel attention module is introduced before the model output layer. This module enhances the model's ability to identify diurnal cycles, seasonal cycles, meteorological process cycles, and topographically induced periodic disturbances.

[0057] The dual outputs enable the model to not only provide point prediction results, but also provide a foundation for subsequent uncertainty decomposition and probability interval generation.

[0058] like Figure 3 As shown, step six: Based on the Bayesian posterior approximation, the terrain embedding Fast-Powerformer is sampled multiple times to obtain multiple power prediction means and conditional variances, and the total prediction variance is decomposed into random uncertainty and cognitive uncertainty according to the law of total variance.

[0059] In step six, the method for decomposing the total prediction variance into random uncertainty according to the law of total variance is to obtain S posterior samples. Θ s For the S posterior samples Θ s The total prediction variance TUvar is decomposed into random uncertainty AUvar and random uncertainty EUvar using Bayesian posterior prediction variance decomposition. The expression for decomposing the random uncertainty AUvar is as follows: ; in, Θ s Let S be the posterior sample. Θ s ; The expression for the cognitive uncertainty EUvar, decomposed into, is: ; The expression for calculating the total prediction variance TUvar is as follows: ; Where TUvar is the total prediction variance.

[0060] Among them, AUvar characterizes the irreducible uncertainty caused by random fluctuations in wind resources, measurement noise, turbulence disturbances, and power curve discreteness, while EUvar characterizes the reducible uncertainty caused by insufficient training samples, complex terrain extrapolation, model parameter uncertainty, and distribution offset.

[0061] The posterior samples θ_s are obtained through one or more of deep ensemble, Monte Carlo Dropout, Monte Carlo DropConnect, variational inference, or Bayesian backpropagation; the training loss employs a β-NLL loss to balance the tradeoff between mean learning and variance learning. ; ; Here, stop(·) indicates stopping the gradient operation. β This is the loss adjustment coefficient.

[0062] After completing the training of the terrain embedding Fast-Powerformer, this implementation further performs Bayesian posterior approximation and variance decomposition. Specifically, S posterior model samples θ1, θ2, …, θs can be obtained using methods such as deep ensemble, Monte Carlo Dropout, Monte Carlo DropConnect, variational inference, or Bayesian backpropagation.

[0063] When a wind turbine is located in an area with dense historical samples, stable operation, and common terrain conditions, EUvar is typically low; conversely, when a wind turbine is under extreme wind direction, rare weather events, special power curtailment conditions, or is newly installed equipment, EUvar typically increases. The system can use this information to identify areas where additional measurement points, supplementary samples, or model retraining are needed. When β is close to 0, the model is closer to standard negative log-likelihood training; when β is close to 1, the model prioritizes mean-fit accuracy. In practical applications, β can be selected by comprehensively considering prediction error on the validation set, interval coverage, and interval width.

[0064] like Figure 4 As shown, step seven: Based on the power prediction mean, random uncertainty and cognitive uncertainty, generate wind turbine-level power prediction intervals and station-level power prediction intervals at different confidence levels, and generate daily, weekly, monthly or quarterly power generation prediction results based on power integration, consistency correction and medium- and long-term climate response models.

[0065] In step seven, the power prediction interval is generated based on the power prediction mean and the total prediction variance: ; Among them, PI γ (x) represents the prediction interval at confidence level γ, z γ The standard normal quantile or the calibrated empirical quantile is used; when there is significant non-Gaussian error, quantile calibration or mixed density calibration is further used to correct the interval boundaries. In step seven, the medium- and long-term power generation forecasting model decomposes the power generation series into trend terms, seasonal terms, climate response terms, equipment availability terms, and random disturbance terms: ; ; ; in, E τ Let T be the power generation in the τth statistical period. τ S is the trend term. τ For seasonal terms, C τ For the climate response item, A τ For equipment availability, ε τ For the random disturbance term, c m, τ-lm For having a lag period l m The m-th type of climate factor, β m, τ For climate mode attention weights.

[0066] In step seven, the predicted power generation of the wind farm during the target period is obtained by combining the short-term power prediction integral and the medium- and long-term power generation direct prediction: ; in, This is the predicted power value of the power station. Δt For time intervals, λ is the power generation forecast directly output by the medium- to long-term climate response model. E For consistency correction weights, R pow To improve the reliability of short-term power predictions, TUvar pow The total uncertainty in power prediction, R run The reliability of the running status is represented by k1 to k3, which are weight parameters.

[0067] For power forecasting at the wind farm level, the predicted power of each wind turbine is aggregated according to its availability, power curtailment status, and line constraints: ; in, N w For the number of wind turbines, These are the unit's available status indicators. This is a power curtailment correction index. When the unit is in a fault, shutdown, or maintenance state, its predicted power is corrected to zero or reduced according to availability; when there are power curtailment requirements at the grid connection point of the wind farm, boundary constraint corrections are made according to the power curtailment strategy. The uncertainty at the wind farm level can be obtained from the uncertainty at the turbine level and its correlations. If the correlation between turbines is ignored, the variances can be approximately summed; if the upstream-downstream relationship and common meteorological disturbances are considered, a covariance term is introduced: ; in, ρ ij Let be the correlation coefficient of the prediction error between the i-th and j-th wind turbines, which can be estimated from historical residuals, spatial distance, incoming flow direction, and relative terrain relationship.

[0068] Step 8: Combining equipment availability, power curtailment status, predicted deviation in power volume, and electricity trading price parameters, calculate the conditional risk value and trading risk indicators, and output the multi-timescale power, power generation, prediction range, and risk assessment results for the mountain wind farm.

[0069] In step eight, the conditional value at risk is calculated using the following formula: ; Where L is the loss variable caused by predicted power deviation, equipment failure, power curtailment, or power trading assessment, and VaR α (L) represents the confidence level. Value at Risk (VaR) α (L) represents the conditional expected loss exceeding the value at risk, π dev π fault and π cur These are the loss coefficients related to deviation, fault, and power outage, respectively.

[0070] Based on short-term and medium-term power forecasts, a medium- to long-term power generation forecasting model is further established to predict future daily, weekly, monthly, or quarterly power generation. First, historical power generation is aggregated according to statistical periods, and average wind speed, effective wind hours, air density, equipment availability, maintenance time, power curtailment volume, and major weather events are simultaneously compiled.

[0071] The climate response term is determined by the atmospheric circulation index and the climate modal attention mechanism: For newly built mountain wind farms or wind farms with insufficient historical samples, reference sites with similar terrain, climate zones, and wind resource distribution can be selected, and transfer learning or parameter fine-tuning methods can be used to enhance the stability of medium- and long-term power generation prediction.

[0072] To avoid discrepancies between short-term power integration results and direct medium- to long-term power generation forecasts, this implementation introduces a consistency correction mechanism. First, the short-term or medium-term power forecast results are integrated over time intervals to obtain the target cycle power generation: ; Then, the power integral results are weighted and fused with the output of the medium- and long-term climate response model. When short-term meteorological conditions are stable, the power prediction range is narrow, and the reliability of the operating status is high, λ E Increasing λ increases the system's trust in the power integral results; when future weather fluctuations are strong, the prediction interval is wide, or there are long-term climate anomaly signals, λ... EThe system reduces the weights of medium- and long-term climate response models.

[0073] This invention also discloses a probabilistic prediction system for mountain wind farms, which includes a multi-source data acquisition module, a data governance module, a terrain factor extraction module, a three-dimensional micro-meteorological field reconstruction module, a terrain embedding module, a Fast-Powerformer power prediction module, a Bayesian posterior sampling module, a variance decomposition module, a medium- and long-term power generation prediction module, a consistency correction module, a conditional risk value assessment module, and a result output module.

[0074] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method.

[0075] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0076] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention, and these will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A probabilistic prediction method for mountain wind power integrating Fast-Powerformer and Bayesian methods, characterized in that, The method includes, Step 1: Collect multi-source data from mountain wind farms; Step 2: Process the multi-source data of the mountain wind farm to obtain the wind turbine operating characteristics; Step 3: Obtain terrain parameters and generate three-dimensional micro-meteorological field characteristics of the mountain area based on the terrain parameters; Step 4: Obtain historical power features, time features, and optional weather forecast coding features. Convert the wind turbine operation features, mountain three-dimensional micro-meteorological field features, and terrain parameters into terrain embedding features. Fuse the terrain embedding features with historical power features, time features, and optional weather forecast coding features to obtain the terrain adaptability input tensor. Step 5: Input the terrain adaptability input tensor into the terrain embedding Fast-Powerformer model, and the terrain embedding Fast-Powerformer model outputs the power prediction mean and conditional variance for the future prediction period; Step 6: Based on the Bayesian posterior approximation technique, the terrain embedding Fast-Powerformer model is sampled multiple times to obtain the total prediction variance. According to the law of total variance, the total prediction variance is decomposed into random uncertainty and cognitive uncertainty. Step 7: Generate power prediction intervals and power generation prediction results based on the power prediction mean, total prediction variance, random uncertainty, and cognitive uncertainty; Step 8: Combine equipment availability, power curtailment status, predicted deviation power, and electricity trading price parameters to obtain conditional risk value and trading risk index. Based on the conditional risk value and trading risk index, output the multi-timescale power, power generation, prediction range, and risk assessment results of the mountain wind farm.

2. The method for probabilistic prediction of mountain wind power integrating Fast-Powerformer and Bayesian methods according to claim 1, characterized in that, In step two, the multi-source data are sequentially subjected to quality control, anomaly identification, missing data repair, time synchronization, spatial registration, normalization processing, and reliability assessment to form wind turbine operating characteristics.

3. The method for probabilistic prediction of mountain wind power integrating Fast-Powerformer and Bayesian methods according to claim 1, characterized in that, In step three, the method for generating the three-dimensional micro-meteorological field characteristics of the mountain based on the terrain parameters is as follows: Meteorological data sources, physical correction terms, and neural residual correction terms are obtained. Based on the meteorological data sources, physical correction terms, and neural residual correction terms, the terrain parameters are reconstructed into three-dimensional micro-meteorological field features of the mountain by multi-source meteorological interpolation, physical correction, and neural residual correction.

4. The method for probabilistic prediction of mountain wind power integrating Fast-Powerformer and Bayesian methods according to claim 1, characterized in that, In step four, the terrain embedding features include static terrain features, incoming flow direction features, and micro-meteorological gradient features.

5. The method for probabilistic prediction of mountain wind power integrating Fast-Powerformer and Bayesian methods according to claim 1, characterized in that, In step five, the terrain embedding Fast-Powerformer model includes a lightweight LSTM embedding module, an input transpose mechanism module, a frequency-enhanced channel attention module, and a Reformer basic backbone network module.

6. The method for probabilistic prediction of mountain wind power integrating Fast-Powerformer and Bayesian methods according to claim 1, characterized in that, In step five, the expressions for the power prediction mean and conditional variance are as follows: ； ； Where μ is the power prediction mean, σ 2 Here, h is the conditional variance, h is the prediction step size, and θ is the model parameter. P_{i,t+h} For the i-th wind turbine in the future t+h The power random variable at time Z final This is the feature representation after frequency domain enhancement.

7. The method for probabilistic prediction of mountain wind power integrating Fast-Powerformer and Bayesian methods according to claim 6, characterized in that, In step six, the method for decomposing the total prediction variance into random uncertainty according to the law of total variance is to obtain S posterior samples. Θ s For the S posterior samples Θ s The total prediction variance TUvar is decomposed into random uncertainty AUvar and random uncertainty EUvar using Bayesian posterior prediction variance.

8. The method for probabilistic prediction of mountain wind power integrating Fast-Powerformer and Bayesian methods according to claim 7, characterized in that, In step seven, the expression for calculating the power prediction interval is: ； Among them, PI γ (x) represents the prediction interval at confidence level γ, z γ These are the standard normal quantiles or the calibrated empirical quantiles.

9. A method for probabilistic prediction of mountain wind power integrating Fast-Powerformer and Bayesian methods according to claim 1, characterized in that, In step three, the terrain parameters include slope, aspect, curvature, relative elevation difference, surface roughness, ridgeline, valley passage, windward and leeward relationship, and terrain shading factor.

10. A method for probabilistic prediction of mountain wind power integrating Fast-Powerformer and Bayesian methods according to claim 1, characterized in that, In step one, the multi-source data of the mountain wind farm includes micro-meteorological observation data of the station, wind turbine SCADA operation data, wind turbine static parameters, digital elevation model data, wind measurement tower data, remote sensing wind data, meteorological reanalysis data, optional regional meteorological large model forecast data, optional numerical weather prediction data, historical climate data, and atmospheric circulation index.

Images