A wind power prediction method and system fusing regional weather prediction and weather condition state space power prediction

Abstract

The application discloses a wind power prediction method and system fusing regional weather map weather prediction and weather condition state space power prediction, and belongs to the field of new energy power generation prediction. The method comprises the following steps: constructing a regional weather map comprising weather grid nodes, terrain nodes and wind turbine target nodes, predicting future weather variables, weather hidden states and weather credibility of the wind turbine point by a graph neural network; encoding a historical power sequence into a power hidden state, fusing the weather prediction result into a weather-power condition alignment representation, and generating weather condition gating parameters step by step; taking the power hidden state as an initial state, modulating the state transition, input and output matrix of a selective state space model by the gating parameters, and outputting an initial power prediction; after wind turbine space coupling correction and dynamic physical consistency constraint, outputting a final power point prediction, a prediction interval and a risk level. The application improves the accuracy, reliability and dispatching availability of wind power prediction under complex wind conditions.

Description

Technical Field

[0001] This invention relates to the fields of new energy power generation forecasting, artificial intelligence, meteorological field forecasting, graph neural networks, state-space sequence modeling, and power system dispatching technology, specifically to a wind power forecasting method and system that integrates regional graph meteorological forecasting and meteorological condition state-space power forecasting. Background Technology

[0002] Wind power output is affected by factors such as weather, topography, and generator status, exhibiting strong nonlinearity, strong volatility, and significant spatiotemporal coupling characteristics.

[0003] First, existing methods mostly use NWP, ERA5, or weather station data as external inputs, without incorporating weather forecasts into power prediction models, making it difficult to coordinate and control weather uncertainties.

[0004] Second, existing methods mostly employ meteorological variable splicing or static attention, which means that future meteorological conditions cannot participate in the selective updating of the hidden power state, resulting in insufficient response during complex weather transition periods.

[0005] Third, existing time series models struggle to balance long-term power inertia with short-term wind condition abrupt changes; Transformer calculations are complex, and long-range memory in recurrent neural networks is prone to decay.

[0006] Fourth, existing physical constraints mostly employ rated power reduction or simple power curve penalties, which cannot be dynamically adjusted according to air density, hub wind speed, yaw, wake, and operating conditions. Summary of the Invention

[0007] This invention provides a wind power prediction method that integrates regional map meteorological forecasting and meteorological condition state space power prediction to solve problems in existing methods such as the separation of meteorological input and power prediction, the absence of meteorological variables in power state evolution, the difficulty in balancing long-sequence prediction and abrupt response, and the static nature of physical constraints.

[0008] To address the aforementioned technical problems, this invention provides a wind power prediction method that integrates regional map meteorological forecasting and meteorological condition-state spatial power prediction, comprising the following steps: Acquiring historical meteorological field data, topographic data, wind turbine spatial location data, wind turbine operating status data, and wind turbine power time series data of the target wind farm; Constructing a regional meteorological map based on the historical meteorological field data, topographic data, and wind turbine spatial location data, wherein the regional meteorological map includes a node set and an edge set, the node set including meteorological grid nodes, topographic nodes, and wind turbine target nodes, and the edge set including meteorological adjacency edges, topographic influence edges, grid-to-wind turbine mapping edges, and wind direction propagation edges; Inputting the regional meteorological map into a regional map meteorological prediction model, and predicting the meteorological variables, meteorological latent states, and meteorological confidence levels of wind turbine locations at multiple future time steps through multi-layer message passing in a graph neural network; Encoding the historical power time series data as power latent states, and fusing the power latent states with the meteorological variables, meteorological latent states, and meteorological confidence levels to generate a meteorological-power condition-aligned representation; Based on the meteorological-power condition alignment representation, meteorological characteristics and reliability of each prediction step, and location encoding, meteorological condition gating parameters are generated step by step. Using the power latent state as the initial state of the state-space model, the state transition matrix, input matrix, and output matrix of the state-space model are modulated by the meteorological condition gating parameters, performing multi-step state evolution and outputting the initial power prediction result. The initial power prediction result is then subjected to wind turbine spatial coupling correction to obtain the spatially corrected power prediction. A dynamic power feasible region is generated based on future meteorological variables and wind turbine operating states. This dynamic power feasible region is used to apply dynamic physical consistency constraints to the spatially corrected power prediction, outputting the final power point prediction, prediction interval, and risk level.

[0009] A wind power prediction system implementing the prediction method of the present invention includes: a data acquisition module for acquiring historical meteorological field data, topographic data, wind turbine spatial location data, wind turbine operating status data, and wind turbine power time series data of a target wind farm; a regional map meteorological prediction module for constructing a regional meteorological map based on the historical meteorological field data, topographic data, and wind turbine spatial location data, and predicting meteorological variables, meteorological latent states, and meteorological confidence levels for wind turbine locations at multiple future time steps using a graph neural network; a meteorological condition state-space power prediction module for encoding historical power sequences into power latent states, fusing them with the meteorological prediction results to generate meteorological condition gating parameters, modulating the state-space model for power evolution, and outputting an initial power prediction; a spatial coupling correction module for performing wind turbine spatial coupling correction on the initial power prediction; and a dynamic physical consistency constraint module for generating a dynamic power feasible region based on future meteorological variables and wind turbine operating status, constraining and correcting the power prediction, and outputting the final power point prediction, prediction interval, and risk level.

[0010] By adopting the above technical solution, the present invention has the following beneficial effects:

[0011] 1. Improve the adaptability of weather forecasting under complex terrain by generating future meteorological variables that are spatially aligned with wind turbine nodes through regional meteorological maps and graph neural networks.

[0012] 2. By modulating state-space power evolution through meteorological condition parameters, meteorological forecasts can directly participate in power state updates, thereby improving the ability to respond to meteorological abrupt changes.

[0013] 3. Through selective state-space structure to The complexity balances long-term power inertia and short-term weather change response.

[0014] 4. Improve the spatial consistency of turbine power prediction by correcting and supplementing the effects of wake and wind direction propagation through spatial coupling correction of wind turbines.

[0015] 5. Implement soft constraints for training and hard corrections for inference through dynamic physical feasible regions, and output power range and risk information. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the overall process of two-stage wind power prediction.

[0017] Figure 2 A schematic diagram for constructing a regional meteorological map;

[0018] Figure 3 This is a structural diagram of a regional meteorological forecasting model.

[0019] Figure 4 This is a structural diagram of the state-space power prediction model under meteorological conditions.

[0020] Figure 5 A schematic diagram of the weather condition gating status update mechanism;

[0021] Figure 6 This is a schematic diagram of dynamic physical consistency constraints and risk output. Detailed Implementation

[0022] This invention sequentially constructs regional meteorological maps and outputs them. , , ; Encode historical power as and will and merge into ;according to , , and Step-by-step generation ;by As the initial state, by Modulation state space evolution is obtained After spatial correction and physical constraints, the output power prediction, range, and risk level are determined.

[0023] The core innovations of this invention include:

[0024] 1. Regional map meteorological forecasting mechanism

[0025] This invention constructs regional meteorological maps Predicted by graph neural network , and .

[0026] Node set Including meteorological grid nodes (wind speed) ), terrain nodes Wind turbine node edge set Includes meteorological adjacency, topographic influence, grid mapping, and wind direction propagation edges.

[0027] After message passing and physical correction in the residual graph at layers 6-8, the output is... , and .

[0028] 2. Meteorological Condition-Based Spatial Power Prediction Mechanism

[0029] The second phase will Transform into ,in For alignment purposes, , , The first Meteorological characteristics, reliability, and location coding; , , Modulated, The feedforward matrix is ​​fixed. The state equation is:

[0030] ;

[0031] ;

[0032] The initial state is ;No. Step input by and Integration and filter:

[0033] ;

[0034] Represents element-wise product. For input gating.

[0035] The attenuation is caused by Modulation: Update faster when wind speed increases, and faster when weather is stable. Approaching unit array.

[0036] Reasoning complexity Parallel training complexity Lower than Transformer .

[0037] 3. Meteorological condition gating status update mechanism

[0038] This invention generates five types of gating parameters step by step based on future wind speed, wind direction, air density, meteorological latent state, and meteorological reliability: input selection gating. State retention gating State update gating Output response gating and interval width gating For the first In each forecast step, the gating parameters are represented by meteorological conditions. Meteorological characteristics of this step The reliability of this meteorological step and position encoding Co-generation:

[0039]

[0040] Four types of modulation gates are mapped to sigmoid. : To ensure constant positive values ​​and bidirectional scaling, the interval width gating adopts an exponential form. .

[0041] The gating function is as follows: modulation ; Control state maintained. Speed ​​up refresh when the value is reduced; modulation ; Control status update; modulation ; Acting on ; Adjustment range ; The basic quantile. When Below hour, Increase.

[0042] Five types of gating: Acting on ; Acting on ; Acting on ; Acting on ; Adjust the quantile interval.

[0043] remove In addition, the input includes , No. step , No. step and and join and ; Entered .thus Responding to changes in wind speed, and Responding to changes in wind direction, The response is of low reliability.

[0044] 4. Wind turbine spatial coupling correction mechanism

[0045] This invention utilizes the correlation between future wind direction, wind speed, relative wind turbine position, and historical power to generate spatial correction parameters in a lightweight manner. Initial prediction of the state space Perform upstream and downstream spatial coupling correction:

[0046] ;

[0047] Average wind direction For wind turbine relative wind turbine The displacement; ( (where the diameter is), only when And the fan Located in the wind turbine Effective when downwind. Correction amount. ,in For wind turbine Historical power The learnable coefficient; Obtained through constraints .

[0048] Space correction only increases Sparse matrix operations and a coefficient .

[0049] 5. Meteorological-driven dynamic physical consistency constraints and risk output mechanisms

[0050] This invention is based on future wind speed air density Yaw angle Pitch angle Tip speed ratio Wake reduction factor Rated power and operational efficiency The upper bound for dynamically generating power for each wind turbine at each forecast time step is:

[0051] ;

[0052] ;

[0053] in The power coefficient is determined by... and Sure; , , These are yaw, wake turbulence, and operational efficiency, respectively. Depend on Crop to .

[0054] Training uses soft constraints The inference process employs hard pruning and outputs the range and risk.

[0055] The following describes the specific implementation using SDWPF and ERA5 / WeatherBench2 data.

[0056] Data Dimension Definition

[0057] The variables are defined as follows:

[0058] Batch size.

[0059] Number of wind turbines. Taking the SDWPF dataset as an example, The code example can be set to... .

[0060] Historical window, 48 typical 10-minute steps.

[0061] Prediction window, typically 12 10-minute steps.

[0062] Power operating characteristics, including Reactive power, speed, yaw code, pitch angle, temperature, status code and ,common Time coding is not included. .

[0063] Weather forecast features, including hub height and wind speed. Wind direction cosine coding, wind direction sine coding, temperature air pressure , specific moisture air density ,common One channel. Weather reliability. Output as an independent tensor, not included in the calculation. Number of channels.

[0064] : Raw meteorological field features (first stage input), including 10-meter horizontal wind speed Components, near-surface temperature, surface air pressure, specific humidity, geopotential height, total .

[0065] Hidden state dimension, typical value .

[0066] Quantity output count, typical value (Corresponding to 0.1, 0.5, and 0.9 quantiles).

[0067] : Number of meteorological grid nodes, typical value ( (Grid).

[0068] Core tensors of each module:

[0069] Historical weather field input (ERA5 / WeatherBench2) Historical power and operating status sequence Phase 1 Output: Meteorological Variables for Future Wind Turbine Locations Phase 1 Output: Regional Weather Latent Status First-stage output: Meteorological confidence level (independent tensor, value 0~1) Historical power hidden state Meteorological-Power Condition Alignment Meteorological condition gating parameters (generated step by step in forecasting) No. Step state space power hidden state State-space initial power prediction (uncorrected) Power prediction after spatial coupling correction Final power point prediction after physical clipping Quantile prediction

[0070] like Figure 1 As shown, this method includes three stages: data preprocessing, regional map meteorological forecasting, and meteorological condition state-space power forecasting. The specific steps are as follows:

[0071] Step S1, obtain historical weather fields (including (etc.), wind turbine static information and SCADA power operation data.

[0072] Step S2, perform time alignment and derived computation ( , , ), anomaly marking, standardization and by / Windowed mode.

[0073] Step S3: Construct a regional meteorological map based on historical meteorological field data, topographic data, and wind turbine coordinates. Specifically, it includes a set of nodes. (Three types of nodes) and edge set Initialization of (four types of edges), and calculation of adjacency relationships and edge weights.

[0074] Step S4: Input the regional meteorological map and historical meteorological field data into the regional map meteorological prediction model and perform variable grouping and coding. Area Map Message Passing Future weather forecast The four sub-steps of wind turbine location meteorological decoding output .

[0075] Step S5: Transfer historical SCADA data The input sequence is fed into the power operation sequence embedding unit, and after processing by input projection, temporal convolution, and gated recurrent units, the power hidden state is output. .

[0076] Step S6, according to Generate meteorological condition embedding ,and Meteorological condition representations are generated through gating fusion alignment. .

[0077] Step S7, by Output from meteorological condition gating parameter generation unit It includes five types of gating parameters.

[0078] Step S8, with For the initial state, As a conditional parameter, execute Step-gated selective state-space power evolution to generate initial predictions .

[0079] Step S9: Generate spatial coupling correction quantities based on future meteorological variables, wind turbine coordinates, and historical power correlations. After obtaining spatial correction, the prediction is... .

[0080] Step S10: Generate a dynamic power feasible region based on future meteorological variables and wind turbine operating status, and then... Perform soft constraints during training and hard pruning during inference to output the final power point prediction. Quantile prediction and risk output.

[0081] 1. Data Acquisition and Preprocessing Module

[0082] 1.1 Meteorological Field Data Acquisition

[0083] Historical weather fields were obtained using ERA5 / WeatherBench2, with a spatial resolution of [missing information]. Variables include wind speed ,temperature air pressure , specific moisture Potential height (unit ), roughness and topographic standard deviation .

[0084] Meteorological grid coverage extends outwards from the wind field Scope, composition ,typical ( (Grid).

[0085] 1.2 Wind Turbine Operation Data Acquisition

[0086] SCADA time resolution is 10 minutes, and the field is active power. No merit Rotation speed Yaw angle Pitch angle cabin temperature and status codes.

[0087] Static information includes altitude Wheel hub height Impeller diameter ,radius ,area and rated power .

[0088] 1.3 Time Alignment and Interpolation

[0089] ERA5 and SCADA have different time granularities, so time alignment is required.

[0090] ERA5 data was converted to 10-minute granularity using linear interpolation:

[0091] ;

[0092] The interpolation results are matched with SCADA by timestamp, with a maximum allowed time difference of 5 minutes.

[0093] 1.4 Calculation of Derived Variables

[0094] Air density (estimated from air pressure, temperature, and specific humidity):

[0095] ;

[0096] in The gas constant for dry air. This is a virtual temperature.

[0097] Surface wind speed and direction:

[0098] ;

[0099] Power change rate (climbing characteristics):

[0100] ;

[0101] in This refers to the normalized active power.

[0102] Wheel hub height wind speed correction (corrected from 10-meter height to wheel hub height) ):

[0103] ;

[0104] in The wind profile power index is derived from surface roughness. Logarithmic wind profile estimation under neutral conditions:

[0105] .

[0106] 1.5 Abnormal Status Marking

[0107] Abnormal periods are only marked, not deleted:

[0108] Shutdown indicator: Active power And rotational speed Lasting more than 30 minutes, marked .

[0109] Power rationing indicator: Operating status code is 2, or active power is limited by high wind speed ( When the power is significantly lower than the rated power ( ),mark .

[0110] Sensor fault flag: If the value remains constant or exceeds the physical range for three or more consecutive time steps, a fault flag is set. .

[0111] The above-mentioned marker fields serve as runtime efficiency. The calculation basis is used to generate gating parameters in the second stage of power prediction.

[0112] 1.6 Data Standardization and Windowing

[0113] Standardization: Applying the training set mean to meteorological variables and standard deviation Perform Z-score standardization: Active power is normalized to rated power. Other power-related variables are normalized to their respective rated values ​​or ranges.

[0114] Slide the chart in 10-minute increments to capture... Length segment, before Step is input, then The goal is to achieve a step-by-step approach; the data is divided by time and is not randomly shuffled.

[0115] 1.7 Module Input / Output Data Dimensions

[0116] Input: ERA5 raw weather field ， Input: SCADA raw sequence 7 raw fields Output: Historical weather field tensor Input the first stage Output: Historical power tensor Enter the second stage Output: Target power tensor Training supervision labels

[0117] 2. Regional Map Weather Forecasting Module

[0118] The first phase structure is as follows: Figure 3 As shown.

[0119] 2.1 Construction of Regional Meteorological Maps

[0120] Regional meteorological maps are defined as , For a set of nodes, Let be the set of edges. It is an adjacency matrix.

[0121] Node set :

[0122] ERA5 mesh node, features include , , , , and latitude and longitude; typical , coverage expansion scope.

[0123] : Terrain node, characterized by , , , ,quantity .

[0124] Target nodes and quantity of wind turbines The data includes coordinates, hub height, impeller area, and rated power, and is only used for decoding.

[0125] edge set There are four categories in total:

[0126] 8-neighborhood, edge features , , Weight , km.

[0127] : Connect to the most recent Each meteorological grid node has edge characteristics of distance and elevation difference.

[0128] Each wind turbine is connected to the nearest Each grid node has a weight equal to a bilinear interpolation coefficient.

[0129] :right ,like ,in For wind direction, For nodes relatively The displacement, and ( (grid spacing), then build the edge Features .

[0130] 2.2 Grouping and coding of meteorological variables

[0131] Original weather channel Including wind speed Temperature, air pressure, specific humidity, and geopotential height; derived quantities are coded in five groups. The following... To expand the number of features, and distinguish:

[0132] Wind field variable group ( ):original Components, derived horizontal wind speed Wind angle ;

[0133] Thermodynamic variable set ( ): Original temperature , derived virtual potential temperature ;

[0134] Water vapor variable group ( Original specific humidity ;

[0135] Group of air pressure variables ( Original ground air pressure Derived normalized pressure ;

[0136] Topographic / Surface Variables Group ( Original potential height Terrain nodes provide terrain elevation. and surface roughness .

[0137] For the Group variables undergo linear embedding and layer normalization:

[0138] ;

[0139] in Output dimension .

[0140] Embedding vectors and temporal location encodings (Sine and cosine encoding of hours and dates, dimension 4) and spatial location encoding (Latitude and longitude sine and cosine encoding, dimension 4) After concatenation, the initial node features are obtained through a linear fusion layer:

[0141] ;

[0142] 2.3 Area Map Message Passing

[0143] In regional meteorological maps Execution layer( The residual graph uses message passing to update the hidden state of nodes at each level.

[0144] Edge message calculation: For any directed edge in the graph , No. The layer edge message is:

[0145] ;

[0146] in It consists of two layers of MLP + ReLU. Embed edge attributes (distance, azimuth, edge type).

[0147] Node aggregation and update: for nodes Aggregate all incident edge messages and update the hidden state:

[0148] ;

[0149] ;

[0150] ;

[0151] Add LayerNorm to MLP For terrain correction coding.

[0152] 2.4 Future Weather Forecast

[0153] through After the layer message is passed, the meteorological grid node is in a hidden state. It has integrated local meteorological field, topographic disturbance, and wind direction propagation information. Based on this, a residual incremental prediction method is used to output the future meteorological conditions:

[0154] ;

[0155] in For the first Step-by-step meteorological increment prediction is achieved using a lightweight GRU or linear recursive structure:

[0156] ;

[0157] As conditional coding for the prediction step.

[0158] 2.5 Meteorological Decoding of Wind Turbine Locations

[0159] The future hidden states of the meteorological grid nodes are interpolated to the wind turbine target nodes using a graph-to-target decoder.

[0160] ;

[0161] in For pre-calculated bilinear interpolation weights ( ), For wind turbine The Neighboring meteorological grid nodes, .

[0162] The following variables are output by the multi-head decoding layer for the interpolated hidden states of the wind turbine nodes:

[0163] Hub height wind speed (corrected by power law): ;

[0164] Air density (derived from predicted air pressure, temperature, and specific humidity): ;

[0165] Meteorological reliability: According to the fan and prediction steps The output is obtained by a weighted combination of the following three normalization factors via a sigmoid mapping:

[0166] ;

[0167] in , , These respectively characterize intervals, abrupt changes, and historical errors; As weight, For bias, Limit the results to ; The larger it is, the more credible it is.

[0168] 1.2.6 Module Input / Output Data Dimensions

[0169] Input: Historical weather fields ， Internal: Initial characteristics of nodes Initial embedding of three types of nodes Internal: Node characteristics after message passing through Layer update Output: ( Wind direction cosine, wind direction sine, , , , ) Output: Wind turbine node meteorological hidden state Output: Meteorological reliability (independent tensor)

[0170] 3. Meteorological Condition Spatial Power Prediction Module

[0171] The second phase and For input, see the structure. Figure 4 gate control Figure 5 .

[0172] 3.1 Power Operation Sequence Embedding

[0173] Includes One channel: reactive power, speed , , Cabin temperature, status code and .

[0174] Input projection: Perform a linear projection of the feature vector at each time step to the hidden dimension. :

[0175] ;

[0176] Perform a one-dimensional convolution with kernel 3 and stride 1 along the time axis to output the dimension. .

[0177] After aggregation along the time axis by a single-layer GRU, the final hidden state is:

[0178] ;

[0179] 3.2 Meteorological Condition Embedding and Alignment

[0180] Meteorological conditions embedded: The first stage output Fusion encoding:

[0181] ;

[0182] Will Pooling ,Will Pooling ;and spliced ​​as ,through Mapped to .

[0183] Weather-power conditions are aligned as follows:

[0184] ;

[0185] ;

[0186] It is sigmoid. For element-wise product; Decrease increases Weights, retaining historical state when increased. .

[0187] Cross-attention can be used as an alternative, to Query, For keys and values.

[0188] 3.3 Generation of meteorological condition gating parameters

[0189] Each prediction step outputs a set of gates. ,overall .

[0190] No. Step General , (from) No. (step) and Input shared MLP:

[0191]

[0192] For the first Step position encoding, and add and .

[0193] Five-way branch projection: Four types of modulation gating (input selection, state hold, state update, output response) are mapped by independent projectors via sigmoid to... :

[0194] ;

[0195] Interval width gating The value must be greater than 1 to expand the interval and less than 1 to narrow the interval, and it must always be positive. Therefore, an exponential form of a naturally positive scaling factor is used.

[0196] ;

[0197] Depend on have to When the confidence level is low High reliability , When =0.7, the range is approximately .

[0198] The meaning and function of the five types of gating parameters:

[0199] Input gating Acting on the Step input ; during mutation Increase.

[0200] Keep Gated modulation ; The larger it is, the stronger it remains. The younger the child, the faster they forget, and... A maintain-update relationship is formed.

[0201] Update the gate control modulation It increases when the weather changes.

[0202] Output gating modulation Adjust the gain according to the prediction step.

[0203] Interval Gating Adjust the quantile width; when Below hour Increase, when reliable Decrease.

[0204] 3.4 Gated Selective State-Space Power Evolution

[0205] Perform multi-step state evolution as follows:

[0206] State-space fundamental equations: for the th One prediction step ( ):

[0207] ;

[0208] ;

[0209] For the first Step state ( ), For the first Step input, For the first Step output; Depend on and After projection filter:

[0210] .

[0211] Gating modulation rules:

[0212] ,in , Learnable; To stabilize the diagonal matrix, Modulation attenuation. ,in Based on the matrix, For low-rank residuals, Modulation update. , Modulated output; fixed. Acting on , , , Modulate separately , , , Adjust the quantile interval.

[0213] Training complexity Reasoning complexity This invention is not limited to a specific state space implementation.

[0214] Output decoding: for each prediction step Output The power point prediction and fundamental quantile prediction are output after two layers of MLP decoding:

[0215] ;

[0216] through Scaling by median: Summary Step to get .

[0217] 3.5 Wind Turbine Spatial Coupling Correction

[0218] Upstream influence weighting calculation: For wind turbines The upstream influence weight is defined as:

[0219] ;

[0220] in This is the unit vector of the average wind direction within the prediction window. ( (Impeller diameter) Only when lie in Downwind side .

[0221] Spatial correction calculation:

[0222] ;

[0223] ;

[0224] in For wind turbine Historical average power This is the learnable global wake effect coefficient (initialized to 0.05, updated during training).

[0225] 3.6 Dynamic Physical Consistency Constraints

[0226] Dynamic physical feasible domain generation: for wind turbines In the predicted time step Based on the predicted future meteorological variables and SCADA operational status data from the first phase, the physically available power is calculated:

[0227] ;

[0228] Obtained from air pressure, temperature, and humidity; For wind turbine Sweeped area; Depend on and Confirmed, below the rated wind speed, Optimization, theoretical upper limit Above the rated wind speed Increase and decrease; For hub wind speed; For yaw efficiency; Wake reduction; For operational efficiency, normally Power rationing Shutdown .

[0229] , and From ; and These are static parameters; Determined by the incoming flow geometry. When there is no future scheduling, take... ( ); , Depend on Estimation by looking up a table; Retrieve the most recent state or scheduling value.

[0230] Upper bound of the power feasible region:

[0231]

[0232] Constraints ; Upper Realm Follow Change, distinct from .

[0233] Soft-constraint loss during training (for spatially corrected predictions) Apply, guide the model to learn physically consistent output):

[0234] ;

[0235] Hard constraint correction during inference phase (spatial correction of prediction) Trimming to the dynamic feasible region yields the final power point prediction. ):

[0236] ;

[0237] when Time output ,otherwise .

[0238] 3.7 Multi-task output

[0239] Power point prediction Fan In the predicted time step The predicted value of active power.

[0240] Wind farm-level aggregate power: ,in This is the availability weight for the wind turbine (0 when it is down, 1 when it is working normally).

[0241] Quantile prediction , , forming the prediction interval .

[0242] probability of climbing Depend on through Received; tagged as and with Output binary risk.

[0243] Risk level (Low / Medium / High): Based on comprehensive risk score Determined by threshold mapping. The overall score is a weighted sum of all risk factors:

[0244] ;

[0245] in Due to the low reliability of meteorological data, For the probability of climbing, This indicates a physical conflict (predicted power is close to the upper bound of the feasible region). Scoring for historical distribution drift, As weight.

[0246] Drift rating Using normalized Mahalanobis distance:

[0247] ;

[0248] in This is the current weather-power joint feature vector (such as future wind speed, air density, historical power statistics, etc.). and These are the mean vector and covariance matrix of the feature in the training set, respectively. The normalization scaling constant is taken as the 95th quantile of the Mahalanobis distance in the training set, and the result is truncated to... The larger the Mahalanobis distance, the more the current input deviates from the training distribution, and the less reliable the model's predictions are.

[0249] The risk level mapping rule is as follows: hour (Low), hour (middle), hour (High), default , .

[0250] 3.8 Module Input / Output Data Dimensions

[0251] enter: enter: (First stage output) enter: Weather hidden state enter: Meteorological reliability (independent tensor) middle: Hidden state of power history middle: Meteorological-Power Alignment Representation middle: Five types of gating parameters (generated step-by-step) middle: State space hidden state / output middle: State space initial prediction middle: Prediction after spatial coupling correction Output: Final power point prediction after physical clipping Output: Quantile prediction Output: Probability of climbing the hill One of the risk output tensors Output: Meteorological low-confidence markers One of the risk output tensors Output: Risk Level Value One of the risk output tensors

[0252] 4. Model Training and Optimization

[0253] Two-stage models can be trained end-to-end or in stages.

[0254] 4.1 Loss Function Design

[0255] The combined loss is:

[0256] ;

[0257] Point prediction loss :

[0258] ;

[0259] Quantile loss :

[0260] ;

[0261] Physical loss Apply before cutting If Cut to Subsequent calculations will be invalid:

[0262] ;

[0263] meteorological consistency loss :

[0264] ;

[0265] Slope weighted loss :

[0266] ;

[0267] ;

[0268] Slope classification loss Supervision , tag as :

[0269] ;

[0270] >1 mitigates category imbalance (default) =5.

[0271] Default weights: , , , , , .

[0272] 4.2 Optimization Strategy

[0273] Optimizer: AdamW, initial learning rate Weight decay .

[0274] Learning rate scheduling: cosine annealing, 20 epochs period, minimum learning rate .

[0275] Batch size: Historical Window Prediction window Hidden Dimensions .

[0276] 4.3 Evaluation Indicators

[0277] MAE RMSE NMAE MAE divided by rated power to eliminate model size differences. NRMSE RMSE divided by rated power Coefficient of determination, measures goodness of fit Pinball Loss Quantile Coverage Quality Comprehensive Evaluation PICP Predicted interval coverage, target (80% confidence interval) PINAW The narrower the normalized average width of the prediction interval, the better. Hill Climbing Recall Rate The proportion of actual climbing events covered by prediction Weighted MAE during high-risk periods MAE is calculated separately for periods with medium / high risk levels.

[0278] 4.5 Module Input / Output Data Dimensions

[0279] Input: Spatial correction prediction Used for calculation (Before cutting) Input: Final point prediction Used for calculation and Input: Quantile prediction Used for calculation Input: Dynamic upper bound Used for calculation Input: Meteorological physical variables Used for calculation Input: probability of climbing Used for calculation Input: Slope tag Supervision label Input: Actual power Supervision label Output: Total loss scalar Backpropagation optimization objective

[0280] 5. Examples

[0281] 5.1 Implementation process using SDWPF dataset as an example

[0282] The following uses the SDWPF dataset to illustrate the implementation process and parameter configuration.

[0283] SDWPF includes 10-minute SCADA data and coordinates / elevation for 134 wind turbines; anomalies are marked as described in Section 1; ERA5 / WeatherBench2 resolution. , Hours, coverage extends outward by 0.5° ( (Grid).

[0284] Implementation parameter settings: Historical input window: (8 hours, corresponding to 48 10-minute time steps); Prediction output window: (2 hours, corresponding to 12 10-minute time steps); Number of fans: Number of meteorological grid nodes: ( Hidden Dimensions: ; Message passing layers in the diagram: Quantile count: (0.1, 0.5, 0.9); Batch size: Optimizer: AdamW, initial learning rate Weight decay Loss weights: , , , , , , , .

[0285] Data partitioning: Divide the data into training, validation, and test sets according to time sequence. Random shuffling is prohibited to prevent leakage of time-series information. For example, the first 60% of a continuous time period can be used as the training set, the middle 20% as the validation set, and the last 20% as the test set.

[0286] Graph parameters: Adjacent edges 8-neighborhood The closest point to the terrain edge A grid, mapping the nearest edge Each grid is assigned a bilinear weight; the wind direction edge is dynamically updated.

[0287] The test set was compiled over two hours with 134 wind turbines, and the MAE, RMSE, NMAE, and NRMSE were statistically analyzed. Pinball Loss, PICP, PINAW, and hill-climb recall.

[0288] 5.2 Optional Construction Methods for the First-Stage Meteorological Output Interface

[0289] Phase 1 Interface The second-stage input can be constructed from either model inference or reanalysis data, and the inputs are consistent in both methods:

[0290] Method 1: Model outputs future Step , and ,in It is estimated based on the interval width and historical error.

[0291] Method 2: Use ERA5 interpolated meteorological data as... Meteorological field embedded as and set .

[0292] Figure 1 Overall presentation process: Historical meteorology and power Input; First stage construction Output , , Power encoding is , and merge into ,Depend on generate After state-space evolution, we get Spatial correction Physical constraints . modulation Output power, range, and risks.

[0293] Figure 2 middle Depend on (wind speed) , ), ( )and ( ) constitutes; the edges are ( ), ( ), ( )and Together constitute .

[0294] Figure 3 process: Obtained through group coding ,pass Layer messages And predict ,through Neighborhood decoding. (By...) get (Correction) );Depend on and Derivation. Output. ( , including and ), , .

[0295] Figure 4 middle Power Embedding ; Weather Alignment Gating generation . No. Step by , , and calculate Output sequentially , and ; modulation .

[0296] Figure 5 middle and , , Input gate unit, press generate , , , , ; Acting on , , , modulation , according to Adjustment range. The state equation is: and initial value ,enter ; fixed.

[0297] Figure 6 Depend on , , , , , , , , generate and with Determine the upper bound; input Feasible region calculation includes , , , and Hard correction to Output , , , and risk score .

[0298] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A wind power prediction method integrating regional map meteorological forecasting and meteorological condition state spatial power prediction, characterized in that, Includes the following steps: Acquire historical meteorological field data, topographic data, wind turbine spatial location data, wind turbine operating status data, and wind turbine power time series data of the target wind farm; A regional meteorological map is constructed based on the historical meteorological field data, topographic data, and wind turbine spatial location data. The regional meteorological map includes a set of nodes and a set of edges. The node set includes meteorological grid nodes, topographic nodes, and wind turbine target nodes. The edge set includes meteorological adjacency edges, topographic influence edges, grid-to-wind turbine mapping edges, and wind direction propagation edges. The regional meteorological map is input into a regional meteorological prediction model. Through multi-layer message passing in a graph neural network, the model predicts meteorological variables, hidden meteorological states, and meteorological confidence levels for wind turbine locations at multiple future time steps. The historical power time series data is encoded as a power hidden state. This hidden state is then fused with the meteorological variables, hidden meteorological states, and meteorological confidence levels to generate a meteorological-power condition-aligned representation. Based on the meteorological-power condition-aligned representation, the meteorological characteristics and confidence levels of each prediction step, and the location encoding, meteorological condition gating parameters are generated step by step. Using the hidden power state as the initial state of the state-space model, the state transition matrix, input matrix, and output matrix of the state-space model are modulated by the meteorological condition gating parameters. Multi-step state evolution is performed, and the initial power prediction result is output. The initial power prediction results are corrected by wind turbine spatial coupling to obtain the spatially corrected power prediction. A dynamic power feasible region is generated based on future meteorological variables and wind turbine operating status. The dynamic power feasible region is then used to apply dynamic physical consistency constraints to the spatially corrected power prediction, and the final power point prediction, prediction interval, and risk level are output.

2. The method according to claim 1, characterized in that, The meteorological grid nodes carry time-varying meteorological features, the terrain nodes carry static terrain features, and the wind turbine target nodes carry static wind turbine information. The meteorological adjacency edges connect adjacent meteorological grid nodes, and their weights decay exponentially with the distance between nodes, used to express the spatial continuity of the meteorological field. The terrain influence edges connect terrain nodes with neighboring meteorological grid nodes, used to model the influence of terrain on the local wind field. The grid-to-wind turbine mapping edges connect meteorological grid nodes with wind turbine target nodes, and their weights are interpolation coefficients, used to decode the meteorological hidden state to the wind turbine location. The wind direction propagation edges are directed edges, pointing from the upwind node to the downwind node, used to model the propagation directionality of meteorological disturbances along the wind direction.

3. The method according to claim 1, characterized in that, The regional map meteorological prediction model uses a residual incremental prediction method to output future meteorological variables, that is, to predict the change in meteorological state relative to the current time. The meteorological reliability is obtained by comprehensively evaluating the prediction quantile interval width, the degree of meteorological abrupt change within the historical window, and the historical prediction error of the wind turbine on the validation set.

4. The method according to claim 1, characterized in that, The meteorological condition gating parameters include input selection gating, state preservation gating, state update gating, output response gating, and interval width gating. The input selection gating acts on the input features of the current prediction step, controlling the effective amplitude of meteorological driving features entering the state update. The state preservation gating modulates the overall decay rate of the state transition matrix, controlling the degree of retention of historical power inertia. The state update gating modulates the overall amplitude of the input matrix, controlling the response speed to new meteorological driving forces. The output response gating modulates the gain of the output matrix, controlling the output amplitude of different prediction steps. The interval width gating acts on the interval width of quantile prediction, adaptively adjusting the width of the prediction interval based on meteorological confidence.

5. The prediction method according to claim 4, characterized in that, The interval width gate is constructed in exponential form, and its expression is: ； in, Let be the interval width gate value for the k-th prediction step. These are intermediate features of the gated generative network. This is the scaling hyperparameter. and The parameters are learnable; when the meteorological reliability is lower than a preset threshold, the interval width gate is automatically increased to output a wider prediction interval.

6. The prediction method according to claim 1, characterized in that, The state update equation and output equation of the state-space model are as follows: ； ； in, This is the hidden power state at step k. This is equal to the power hidden state. For the input features at step k, The output hidden state at step k. Let be the meteorological condition gating parameters for the k-th prediction step. , , These are the state transition matrix, input matrix, and output matrix, respectively, modulated by the meteorological condition gating parameters. It is a fixed feedforward matrix that is not subject to gating modulation.

7. The prediction method according to claim 1, characterized in that, The wind turbine spatial coupling correction specifically includes: Based on the average wind direction unit vector within the future forecast window, the relative displacement vector between each wind turbine, and the wake attenuation distance, the upstream influence weight of the upstream wind turbine on the downstream wind turbine is calculated. Based on the upstream influence weight, the learnable global wake influence coefficient, and the historical average power of each wind turbine, the spatial correction is calculated; the spatial correction is then superimposed on the initial power prediction result to obtain the spatially corrected power prediction.

8. The prediction method according to claim 1, characterized in that, The dynamic power feasible region is calculated using the following formula to determine the upper bound of power availability: ； ； in, Let i be the aerodynamic capture power of the i-th wind turbine in the k-th prediction step. air density, For impeller swept area, For power coefficient, Wind speed at wheel hub height For yaw efficiency, This is the wake reduction factor. For operational efficiency, Rated power, This is the upper bound of the dynamic power feasible region; the dynamic physical consistency constraint is applied in the form of a soft constraint loss function during the training phase and in the form of a hard constraint pruning during the inference phase.

9. The method according to claim 1, characterized in that, The risk level is obtained through a comprehensive risk score mapping, which is calculated by weighting the following risk factors: low meteorological confidence, probability of climbing events, physical conflict indicator factor, and historical distribution drift score; wherein, the historical distribution drift score uses normalized Mahalanobis distance based on training set statistics to measure the degree of deviation of the current input feature from the training set distribution.

10. A wind power prediction system implementing the prediction method according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire historical meteorological field data, topographic data, wind turbine spatial location data, wind turbine operating status data, and wind turbine power time series data of the target wind farm; The regional map meteorological prediction module is used to construct a regional meteorological map based on the historical meteorological field data, topographic data, and wind turbine spatial location data, and to predict the meteorological variables, hidden states, and meteorological reliability of wind turbine locations at multiple future time steps through a graph neural network. The meteorological condition state space power prediction module is used to encode the historical power sequence into a hidden power state, and to fuse it with the meteorological prediction results to generate meteorological condition gating parameters, which are used to modulate the state space model for power evolution and output the initial power prediction. The spatial coupling correction module is used to perform wind turbine spatial coupling correction on the initial power prediction; the dynamic physical consistency constraint module is used to generate a dynamic power feasible region based on future meteorological variables and wind turbine operating status, perform constraint correction on the power prediction, and output the final power point prediction, prediction interval, and risk level.

Images