Offshore wind power prediction method based on multi-source meteorological data and terminal

Abstract

The invention discloses an offshore wind power prediction method and terminal based on multi-source meteorological data, and the method comprises the steps: compensating the timeliness difference of different data sources through dynamic weight fusion, and enabling input data to be closer to the actual sea condition; the spatial terrain influence and meteorological information are jointly embedded into features by means of a graph neural network, and the situation that wake flow, terrain disturbance and other factors are ignored in a traditional method is avoided; a model more suitable for the current time period and weather is automatically selected through a task awareness routing mechanism, so that the prediction model is not fixed any more and is automatically switched along with the change of weather characteristics; and finally, through probability distribution parameter output, power prediction has a confidence interval, and upper and lower bound quantification of risks can be provided for a scheduling end. According to the method, the physical credibility, the adaptability and the stability of the prediction data can be remarkably improved, the prediction deviation can be reduced and the possible power fluctuation risk can be identified in advance especially in the complex weather or the multi-fan interference environment, and the overall operation safety of the wind power plant is improved.

Description

Technical Field

[0001] This invention relates to the field of wind power observation technology, and in particular to a method and terminal for predicting offshore wind power based on multi-source meteorological data. Background Technology

[0002] Offshore wind farms are significantly affected by marine meteorological conditions. Meteorological elements such as wind speed, wind direction, air pressure, and temperature all exhibit strong temporal variability and spatial heterogeneity, thus placing high demands on wind power forecasting. Existing wind power forecasting methods typically rely on a single meteorological data source, such as using only numerical weather prediction data or satellite remote sensing data as model input. However, different meteorological data sources vary significantly in terms of timeliness, coverage, and observation accuracy. A single data source often fails to fully reflect the true state of offshore wind fields, leading to substantial deviations in forecast results as weather conditions change.

[0003] On the other hand, most traditional forecasting methods are based on fixed model architectures and lack the ability to adapt to different forecast durations, weather types, and operational scenarios. In offshore wind farms, ultra-short-term forecasts are highly sensitive to temporal details, while short-term forecasts rely more heavily on changes in the meteorological system. Furthermore, the performance of conventional models deteriorates significantly during severe weather events. Due to the lack of the ability to automatically select models based on mission characteristics and weather patterns, existing forecasting systems struggle to maintain stable forecasting performance in complex scenarios.

[0004] Furthermore, offshore wind farms have a high turbine density, and spatial factors such as the wake effect between turbines and the interference of seabed topography on the wind flow field often have a continuous impact on power output. Traditional methods usually treat power prediction as a single-point problem, ignoring the spatial correlation between turbines, which can easily lead to systematic errors in wind farms with complex spatial structures, especially in areas with frequent wind direction changes or significant topographic effects.

[0005] Regarding the expression of prediction results, existing technologies generally output a single power prediction value without quantifying the uncertainty of the prediction results. Due to the rapid changes in marine weather and large observation errors, single-point predictions often fail to meet the risk assessment requirements in actual dispatching. For example, in the event of sudden changes in meteorological conditions or when the model is sensitive to changes in input, the prediction error may amplify rapidly, but existing systems lack corresponding confidence intervals or risk warnings, making it difficult for power grid dispatching to make timely adjustments. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method and terminal for predicting offshore wind power based on multi-source meteorological data, which can significantly improve the accuracy of the prediction data.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for predicting offshore wind power based on multi-source meteorological data includes the following steps: S1. Acquire weather forecast data, satellite remote sensing data and wind turbine real-time operation data to form multi-source heterogeneous data, and use an adaptive weighted fusion algorithm to dynamically assign weights and fuse the multi-source heterogeneous data to generate dynamic fused data. S2: Obtain micro-geographic information of the wind farm, combine it with the dynamic fusion data, construct a meteorological geographic coupled map using a graph neural network, and extract meteorological enhancement features; S3. Obtain the time attribute of the prediction task, combine it with the meteorological enhancement features, use the task-aware routing mechanism to perform optimal model matching and invocation, and determine the target prediction model. S4. Use the target prediction model and the meteorological enhancement features to perform time-series prediction and output point prediction data; S5. Based on the point prediction data and the dynamic fusion data, uncertainty is quantified using a probability distribution parameter generation method to obtain confidence interval data. S6. Integrate the point prediction data and the confidence interval data, perform risk rule matching and early warning signal generation, and form a power prediction result with risk warning.

[0008] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is as follows: A terminal for predicting offshore wind power based on multi-source meteorological data includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps in a method for predicting offshore wind power based on multi-source meteorological data.

[0009] The beneficial effects of this invention are as follows: It provides a method and terminal for offshore wind power prediction based on multi-source meteorological data. By achieving multi-source meteorological data fusion, geographic information enhancement, adaptive model selection, and uncertainty quantification within the same process, it elevates offshore wind power prediction from a single model input method to a dynamic, multi-dimensional information-driven mode. Its core principle lies in: utilizing dynamic weight fusion to compensate for the timeliness differences of different data sources, making the input data closer to actual sea conditions; furthermore, using graph neural networks to embed spatial topographic influences and meteorological information into features, avoiding the neglect of factors such as wake and topographic disturbances in traditional methods; automatically selecting a model more suitable for the current time period and weather through a task-aware routing mechanism, so that the prediction model is no longer fixed but automatically switches according to changes in weather characteristics; finally, through probability distribution parameter output, the power prediction has a confidence interval, providing upper and lower bound quantification of risk for the dispatching end. Overall, this method can significantly improve the physical reliability, adaptability, and stability of the prediction data, especially in complex weather or multi-turbine interference environments, reducing prediction deviations and identifying potential power fluctuation risks in advance, thereby improving the overall operational safety of wind farms. Attached Figure Description

[0010] Figure 1 This is a flowchart of a method for predicting offshore wind power based on multi-source meteorological data, according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating a method for predicting offshore wind power based on multi-source meteorological data, according to an embodiment of the present invention. Figure 3 This is a schematic diagram of an offshore wind power prediction terminal based on multi-source meteorological data according to an embodiment of the present invention; Label Explanation: 1. A terminal for predicting offshore wind power based on multi-source meteorological data; 2. Memory; 3. Processor. Detailed Implementation

[0011] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0012] Before detailing the embodiments of this application, some related concepts will first be explained: To at least solve the above problems, please refer to Figure 1 and Figure 2 This invention provides a method for predicting offshore wind power based on multi-source meteorological data, including the following steps: S1. Acquire weather forecast data, satellite remote sensing data and wind turbine real-time operation data to form multi-source heterogeneous data, and use an adaptive weighted fusion algorithm to dynamically assign weights and fuse the multi-source heterogeneous data to generate dynamic fused data. S2. Obtain micro-geographic information of the wind farm, combine it with the dynamic fusion data, construct a meteorological geographic coupled map using a graph neural network, and extract meteorological enhancement features; S3. Obtain the time attribute of the prediction task, combine it with the meteorological enhancement features, use the task-aware routing mechanism to perform optimal model matching and invocation, and determine the target prediction model. S4. Use the target prediction model and the meteorological enhancement features to perform time-series prediction and output point prediction data; S5. Based on the point prediction data and the dynamic fusion data, uncertainty is quantified using a probability distribution parameter generation method to obtain confidence interval data. S6. Integrate the point prediction data and the confidence interval data, perform risk rule matching and early warning signal generation, and form a power prediction result with risk warning.

[0013] As described above, the beneficial effects of this invention are as follows: by achieving multi-source meteorological data fusion, geographic information enhancement, model adaptive selection, and uncertainty quantification within the same process, offshore wind power prediction is upgraded from a single model input method to a dynamic, multi-dimensional information-driven mode. Its core principle lies in: utilizing dynamic weight fusion to compensate for the timeliness differences of different data sources, making the input data closer to actual sea conditions; then, using graph neural networks to embed spatial topographic influences and meteorological information into features, avoiding the neglect of factors such as wakes and topographic disturbances by traditional methods; automatically selecting a model more suitable for the current time period and weather through a task-aware routing mechanism, so that the prediction model is no longer fixed but automatically switches according to changes in weather characteristics; finally, through the output of probability distribution parameters, the power prediction has a confidence interval, providing upper and lower bounds of risk quantification for the dispatching end. Overall, this method can significantly improve the physical reliability, adaptability, and stability of prediction data, especially in complex weather or multi-turbine interference environments, reducing prediction deviations and identifying potential power fluctuation risks in advance, thereby improving the overall operational safety of wind farms.

[0014] In some implementations, step S1 specifically includes: The multi-source heterogeneous data is spatiotemporally registered and unified to the same spatiotemporal grid system. An adaptive weighted fusion algorithm is used to dynamically allocate weights and fuse data in the same spatiotemporal grid system to generate dynamically fused data.

[0015] As described above, spatiotemporal registration of multi-source heterogeneous data ensures consistency between data from different sources within the same spatial grid and time scale, reducing feature shifts caused by resolution differences. The principle is that data can only achieve effective weighted fusion after spatial and temporal alignment; otherwise, the risk of misalignment arises, such as "the same wind turbine location corresponding to different sea areas." Dynamic weight fusion based on a unified grid system ensures that weighted calculations are performed only between truly corresponding data points, thereby improving the comparability and stability of the fusion results. The registered data makes the subsequent node feature inputs of the graph neural network more accurate, avoiding erroneous associations caused by spatial misalignment, thus improving the reliability of the overall prediction chain.

[0016] In some implementations, step S1 involves using an adaptive weighted fusion algorithm to dynamically assign weights and fuse the multi-source heterogeneous data to generate dynamically fused data. Specifically, this includes: A timeliness weight is established based on the forecast timeliness of the weather forecast data; quality weights are determined for the weather forecast data, satellite remote sensing data, and wind turbine real-time operation data based on the data timeliness, spatial coverage, and observation accuracy of the multi-source heterogeneous data; an adaptive weighted fusion algorithm is constructed based on the timeliness weights and quality weights, and the contribution of the weather forecast data, satellite remote sensing data, and wind turbine real-time operation data in the multi-source heterogeneous data is corrected using the adaptive weighted fusion algorithm, and the weather forecast data, satellite remote sensing data, and wind turbine real-time operation data with corrected contribution are combined to generate dynamic fused data.

[0017] As described above, the dynamic fusion results can automatically adjust the data proportions based on real-time weather changes through a dual-component logic of time-sensitivity weight and quality weight. The principle is as follows: weather forecast data is generally more valuable in the future but contains errors, while satellite data is closer to reality but has a limited update frequency, and wind turbine data reflects the local actual condition. By compensating for the time decay of forecast data with time-sensitivity weight and identifying issues such as noise and insufficient coverage with quality weight, the contribution of the three types of data can be adjusted in real time. When the quality of a data source declines, the algorithm will proactively reduce its weight to avoid a situation where "erroneous data dominates the prediction input," thereby effectively improving the credibility of the fused data, making subsequent model processing more stable, and reducing the accumulation of prediction bias.

[0018] In this embodiment, step S1 specifically involves establishing a multi-source heterogeneous data acquisition and standardization preprocessing system. This system acquires raw data from three independent data sources: a numerical weather prediction model provides weather forecast data, including time series of meteorological elements such as wind speed, wind direction, air pressure, and temperature; a satellite remote sensing system provides satellite remote sensing data of inversion products such as sea surface wind speed and sea surface temperature; and a wind farm monitoring system collects real-time wind turbine operation data. Because the data sources differ significantly in spatiotemporal resolution, data format, and accuracy, system standardization preprocessing is required. Preprocessing includes using Kriging interpolation to fill in missing data, the mathematical expression of which is: ,in These are the estimated values ​​for the points to be interpolated. Given the values ​​at the observed points. The weighting coefficients are determined through semi-variogram optimization. Simultaneously, outlier data is removed based on statistical principles, when data points satisfy... When it is determined to be an outlier, among which The sample mean. This represents the sample standard deviation.

[0019] Based on the completed data preprocessing, the core computational framework of the adaptive weighted fusion algorithm is constructed. This embodiment designs a dynamic weight allocation mechanism based on dual constraints of timeliness and quality weights. For any spatial location within the target region, at a specific time t, the fusion weights of different data sources are calculated using the following dynamic weight function: , ,in and These represent the fusion weights of numerical weather prediction and satellite remote sensing data at time t, respectively. Indicates the forecast lead time. Function To reflect the impact of forecast lead time on data reliability, an exponential decay method is used: Where k is the attenuation coefficient; function Responsible for adjusting the weights based on the real-time data quality assessment results.

[0020] Establish a multi-dimensional data quality assessment system and data quality assessment indicators. The specific expression, calculated using the weighted comprehensive evaluation model, is as follows: ,in It represents the time continuity index, which is obtained by calculating the integrity and stability of data over time series; Indicators representing spatial integrity reflect the performance of data in terms of spatial coverage; This indicates the accuracy of the data, derived from a comparative analysis with measured data. This indicates a data reliability metric, determined through historical performance evaluation. , , , These are the weight coefficients for each indicator, calculated using the entropy weight method, and satisfying the following conditions: The constraints.

[0021] After obtaining the basic weights, a real-time data quality feedback correction mechanism is introduced. Based on the deviation analysis between real-time observation data and various data sources, the initial weights are dynamically adjusted, and the correction formula is as follows: , in and The corrected weights, and These represent the deviations between each data source and the real-time observations. The adjustment coefficients are obtained through training on historical data. This correction mechanism can effectively cope with sudden changes in the quality of data sources, ensuring that the weight of a data source is reduced in a timely manner when an anomaly occurs, thus maintaining the stability of the fusion result.

[0022] Perform weight normalization and multi-source data fusion calculation, and then normalize the corrected weights: , Based on the normalized weights, a weighted average method is used to fuse multi-source data. ,in This represents the merged wind speed data. and These represent wind speed data from numerical weather prediction and satellite remote sensing, respectively. This fusion process is performed separately for all meteorological elements, ultimately generating a dynamic fused dataset containing complete meteorological data. The entire fusion system employs a closed-loop control architecture, dynamically adjusting algorithm parameters by real-time monitoring of the consistency between the fusion results and measured data, ensuring the fusion algorithm can adapt to changes in data characteristics under different weather conditions.

[0023] In some implementations, step S2 specifically includes: Based on the preset wind turbine spatial distribution map, a node map with wind turbine locations as nodes is constructed, and the micro-geographical information of the wind farm is obtained and injected into the node map; The dynamically fused data is mapped to the nodes in the node graph to form a spatiotemporal feature matrix; A graph neural network is used to extract the local relationships and topographic influences between nodes in the node graph, and a meteorological-geographical coupled map is constructed. Meteorological enhancement features are generated using the meteorological-geographical coupled map and the spatiotemporal feature matrix.

[0024] As described above, by constructing a node graph and injecting micro-geographic information, this method structurally represents the spatial relationships, wake effects, and topographic interference among wind turbines. The principle is that wind power is affected not only by local weather but also by the mutual obstruction of neighboring turbines, downstream wake effects, and wind speed changes caused by seabed topography. By mapping dynamically fused data to nodes and then extracting local relationships through a graph neural network, the model can automatically learn the correlation patterns between wind turbines without requiring pre-determined influence weights. The resulting enhanced meteorological features better reflect the overall operating environment of the wind farm, enabling the prediction model to better capture the true physical causes of power changes, thereby reducing systematic bias.

[0025] In this embodiment, step S2 specifically includes: A pre-defined spatial distribution map of wind turbines is constructed, defining the location of each wind turbine generator in the map as a node, thus forming a set of nodes. , where N is the total number of wind turbines in the wind farm. Each node It contains two types of feature vectors: meteorological feature vectors. It consists of meteorological elements corresponding to spatial locations in dynamically fused data, including wind speed, wind direction, temperature, and air pressure; geographic feature vectors This includes micro-geographic information such as wind turbine coordinates, altitude, seabed topography elevation, and distance from the coastline. The initial representation of node features is achieved through feature concatenation: ,in This represents a vector concatenation operation. For nodes The initial feature representation, This represents the total dimension of the features.

[0026] In terms of edge structure construction, two types of edges are defined based on the spatial layout of the wind farm and the characteristics of wind propagation. Spatial adjacency edges are established based on the Euclidean distance between wind turbine nodes. and satisfy Establish a connection at that time, where This refers to the distance between the wind turbines. The spatial adjacency radius is used. The wake influence edges are established based on wind direction and wind turbine layout. Considering the asymmetry of wake influence in actual wind farms, the edge weights are calculated as follows: ,in This is the distance attenuation coefficient. For the node Pointing to node azimuth angle, This represents the current prevailing wind direction. This weighting calculation method accurately reflects the intensity and directional characteristics of the wake effect between wind turbines.

[0027] Based on the constructed graph structure, a multi-layer graph convolutional network is used for spatial feature extraction. In the ll-th layer of graph convolution, the node features are updated using the following formula: ,in Represents a node The set of neighboring nodes, and They are nodes and The degree normalization coefficient, and For the trainable parameters of the l-th layer, for Activation function. This graph convolution operation can effectively aggregate meteorological and geographical information from neighboring nodes, capturing spatial dependencies within the wind farm.

[0028] To further improve the accuracy of feature extraction, a graph attention mechanism is introduced to adaptively learn the importance weights between nodes. In the attention layer, nodes... and The attention coefficient between them is calculated as follows: ,in To share the weight matrix of the linear transformation, Let be the parameter vector of the attention mechanism, and LeakyReLU be the non-linear activation function. The final attention weights are obtained by standardizing the attention coefficients. The feature aggregation formula based on attention weights is: .

[0029] This attention mechanism can dynamically learn the contribution of different neighboring nodes to the feature update of the central node, and can more accurately model the mutual influence between wind turbines, especially under complex terrain conditions.

[0030] Enhanced meteorological features are generated through multi-level feature fusion, which integrates the output features of different graph convolutional layers: Where MLP stands for Multilayer Perceptron, used for feature dimensionality reduction and fusion, and L is the number of graph convolutional layers. The resulting enhanced features... It not only includes raw meteorological information, but also incorporates spatial correlation patterns and geographical environmental influences, providing richer and more discriminative feature representations for subsequent power prediction.

[0031] In some implementations, step S3 specifically includes: Obtain the time attributes of the prediction task; The current weather model is obtained based on the analysis of the aforementioned meteorological enhancement features; Obtain the historical performance of candidate models from pre-established model performance records; The task-aware routing mechanism is used to match the time attribute with the current weather model to the corresponding candidate model. The matching degree of the candidate model's historical performance is calculated, and the candidate model with the highest matching degree is selected as the target prediction model.

[0032] As described above, by comprehensively matching the time requirements of the forecasting task, meteorological characteristics, and historical model performance, the model selection is no longer fixed but automatically chooses the most suitable forecasting model for the current conditions. The principle is that different forecast durations have varying sensitivities to model features, and model performance differs significantly under different weather patterns. By analyzing time attributes and meteorological patterns, it's possible to identify whether the current data is in a stationary state or undergoing rapid change, thus avoiding the use of unsuitable models. Calculating the matching degree by combining the historical performance of candidate models effectively improves the targeting of model selection, making the final forecast results generally more stable and closer to actual change patterns.

[0033] In some implementations, the task routing awareness mechanism in step S3 includes: When the prediction task meets the ultra-short-term requirements, a time-series prediction model based on an attention mechanism is matched. When the prediction task is short-term and the current meteorological model is in a stable meteorological state, a lightweight gradient booster model is matched. When the current meteorological model indicates drastic weather changes, an extreme weather prediction model is used.

[0034] As described above, the task routing mechanism clarifies the model selection rules under different meteorological and forecasting scenarios, enabling the forecasting system to have an interpretable model switching mechanism. The principle is as follows: ultra-short-term forecasts mainly rely on details of temporal changes, therefore attention models are better able to capture rapid dynamics; short-term stable weather is suitable for models with simple structures and fast computation, improving efficiency; drastic weather changes easily amplify the errors of conventional models, thus requiring specialized extreme weather models to capture abrupt changes. By clarifying scenario-based routing rules, the system can automatically select the optimal model, reducing sudden forecast distortions caused by model mismatch and improving risk identification capabilities.

[0035] In this embodiment, step S3 specifically includes: Establish a quantitative representation system for the attributes of the prediction task. The attributes of the prediction task are represented by feature vectors. The (prediction task attribute vector) is mathematically represented, and this vector consists of three key dimensions: time-scale features. Describing the time span of the forecast, a segmented coding method is used to distinguish between ultra-short-term, short-term, and medium-to-long-term forecast needs; meteorological model characteristics. (Meteorological model feature vector) is extracted from enhanced meteorological features, including indicators such as wind speed variation coefficient, wind direction stability, and atmospheric stratification stability; operational demand features. The (operational requirement feature vector) comprehensively considers practical operational factors such as prediction accuracy requirements, computational timeliness constraints, and system resource limitations. The fusion of these features is achieved through a linear transformation: ,in The feature transformation weight matrix, Let r be the time-scale feature, and r be the runtime requirement feature vector. The bias vector, sign This represents a vector concatenation operation. The dimension and number of columns of the corresponding output vector T This corresponds to the total dimension of the input concatenated vector. This representation system can comprehensively and accurately describe the essential characteristics of the prediction task, providing a reliable basis for subsequent model selection.

[0036] In terms of model performance evaluation, a candidate model library containing multiple prediction algorithms was constructed. Each candidate model They are all associated with a set of performance metric vectors , express These are real vectors of dimension; these metrics are continuously updated and improved through offline testing and online performance monitoring using historical data. For a given task attribute T, the model... The expected performance is calculated using a comprehensive evaluation function: ,in It is a matching function between task attributes and model performance, using a Gaussian kernel function. To measure the degree of fit between task requirements and model strengths; The weight vectors for each performance metric are determined jointly by domain expert knowledge and historical data. It is a regularization term used to balance the prediction accuracy and computational complexity of the model; It is the regularization coefficient, which controls the strength of the influence of model complexity on the overall evaluation.

[0037] Routing decision-making based on multi-objective optimization is a core component of task-aware mechanisms. Considering that practical applications require the simultaneous optimization of multiple conflicting objectives, a multi-objective optimization problem is established: The objective function includes a prediction accuracy target. and computational efficiency objectives The problem also considers other possible objectives such as model stability and resource consumption. By solving this multi-objective optimization problem, a Pareto-optimal solution set is obtained. Then, based on the specific application scenario requirements, the most suitable model weight allocation scheme is selected from the solution set. .

[0038] The final routing decision is generated by comprehensively considering factors such as task matching degree, real-time system status, and model weights: ,in It is a task-model matching function, which uses the softmax form to ensure the probabilistic interpretability of the output; `s` is the system state adaptation function, considering real-time constraints such as current computing resource utilization, memory availability, and inference time budget; `s` is the system state vector. The entire routing mechanism also includes an online learning component that dynamically adjusts routing parameters by continuously monitoring the actual performance of the selected model. This ensures that the system can adapt to environmental changes and data distribution shifts, and always maintains optimal model selection performance.

[0039] In some implementations, step S4 specifically includes: The meteorological enhancement features are standardized and continuous time-series input samples are constructed. The time-series input samples are input into the target prediction model, and the target prediction model is used to output point prediction data.

[0040] As described above, standardizing the enhanced features and constructing time-series input samples allows the prediction model to obtain continuous, stable, and scale-consistent input data, which helps reduce training bias caused by differences in feature dimensions. The principle is that time-series prediction relies on a continuous window structure, and standardized data avoids numerical fluctuations affecting the model gradient; constructing continuous samples enables the model to capture the cumulative trend of meteorological changes. After inputting the processed sequence into the target prediction model, the output point prediction data is smoother and more reliable, and it also provides stable input for subsequent uncertainty estimation.

[0041] In this embodiment, step S4 specifically includes: The enhanced meteorological features are subjected to temporal reconstruction and standardization preprocessing, and the enhanced meteorological features output by the graph neural network are then processed. The input samples are organized chronologically. Each sample contains a feature sequence with a historical time step of L as input, predicting the power value at a future time step of H. The input sequence is standardized using a dynamic normalization method. ,in and Let represent the mean and standard deviation of the feature within the sliding window, respectively. This is a small constant added for numerical stability. This dynamic normalization method can adapt to the non-stationary nature of meteorological characteristics while preserving the short-term fluctuation patterns of the sequence.

[0042] In the model forward computation phase, temporal feature extraction is performed based on the target prediction model architecture selected in step 3. For attention-based models, the core computation process is implemented through a multi-head self-attention mechanism: ,in Let X represent the query, key, and value matrices, respectively, and X be the standardized input feature sequence. It is a learnable projection matrix. It is the dimension of the key vector. This attention mechanism can automatically learn the dependencies between different time steps, capturing complex temporal patterns of meteorological elements and power output.

[0043] For a selected target prediction model, its complete time series prediction process can be formally represented as: ,in The parameter is Predictive models, It is the standardized historical feature sequence, and T is the task attribute vector from step 3, used to guide the model's prediction behavior. It is the power prediction sequence for the next H time steps output by the model. The specific architecture of the model depends on the choice of the task-aware routing mechanism: for ultra-short-term prediction tasks, a lightweight architecture combining temporal convolutional networks and gating mechanisms is usually adopted to ensure inference speed; for medium- and long-term prediction tasks, a deeper encoder-decoder structure is adopted, equipped with stronger sequence modeling capabilities.

[0044] In terms of output layer design, considering the physical constraints of wind power, a constrained output converter is adopted: Where z is the hidden representation of the last layer of the model, and These are the weights and bias parameters of the output layer. As a linear combination, the hidden layer feature z is transformed into an unscaled predicted value. This is the rated power of the wind farm. The Sigmoid activation function ensures that the prediction results fall within a reasonable physical range. Meanwhile, prior knowledge of the wind turbine power curve is introduced, and the physical rationality of the prediction is further improved through post-processing calibration.

[0045] Post-processing and quality checks of the prediction results are performed. The original prediction sequence output by the model is checked for temporal consistency and smoothed to eliminate non-physical abrupt changes. Simultaneously, an anomaly detection mechanism is established based on historical prediction error distribution to mark and correct prediction results that significantly deviate from the normal pattern. The entire forward computation process is executed within a unified inference framework, ensuring that different model architectures can generate compliant point prediction data, providing a reliable benchmark prediction for subsequent uncertainty quantification.

[0046] In some implementations, step S5 specifically includes: An uncertainty analysis framework is constructed, in which the probability distribution form is determined based on the dynamic fusion data, and the probability distribution parameters are initialized based on the point prediction data. Confidence intervals are generated using quantile regression based on the probability distribution form and the probability distribution parameters.

[0047] As described above, by determining the probability distribution based on dynamically fused data and initializing distribution parameters using point prediction results, uncertainty quantification gains physical meaning and data-driven characteristics. The principle is that dynamically fused data reflects the complexity of real external weather conditions and can be used to determine whether the error distribution is too wide or too narrow, while point prediction data is used to initialize the center location. Quantile regression further generates confidence intervals, ensuring that the prediction results are not just single-value outputs but have upper and lower bounds, describing the prediction reliability. This helps the scheduling system take contingency plans in advance during periods of high uncertainty, improving the overall system resilience.

[0048] In this embodiment, step S5 specifically includes: Construct an uncertainty analysis framework (also known as a probability distribution parameter generation network) that uses point prediction data and dynamic fusion data As input, the uncertainty characteristics of the predicted values ​​are learned through a multi-layer neural network structure. The network outputs the probability distribution parameters for each prediction time step, including the mean. and variance Specifically, the process of generating the probability distribution parameters can be represented as follows: ,in It is based on the feature transformation function of a neural network, and uses logarithmic variance output to ensure positive definiteness of the variance. This design enables the model to dynamically adjust the prediction uncertainty according to the characteristics of the input data, accurately reflecting the changes in prediction reliability under different meteorological conditions.

[0049] After obtaining the probability distribution parameters, quantile regression is used to generate confidence intervals. For a given confidence level... The corresponding quantiles are calculated using the inverse transform of the cumulative distribution function: , in It is the inverse cumulative distribution function of the standard normal distribution. To accommodate the asymmetric error distribution commonly found in wind power forecasting, a skewed normal distribution assumption is further introduced, using an additional skewness parameter. This adjusts the asymmetry of the confidence interval, enabling more accurate interval estimates when the power is close to zero or at rated power.

[0050] During the model training phase, a combined optimization strategy of quantile loss and negative log-likelihood loss is adopted: The first term ensures the statistical calibration of the confidence interval, while the second term promotes the generation of confidence intervals to be as sharp as possible. This combined loss function guarantees the statistical properties of the confidence intervals while avoiding overly conservative interval estimations.

[0051] The generated confidence intervals are post-processed and optimized. Considering the physical constraints of wind power, boundary constraints are applied to the confidence intervals to ensure that the power is non-negative and does not exceed the rated power. Simultaneously, time-series smoothing techniques are used to eliminate abnormal fluctuations in the confidence intervals, ensuring the continuity of interval estimates between adjacent time steps. Through this series of processing steps, high-quality confidence interval data that satisfies both statistical characteristics and physical constraints are finally obtained.

[0052] In some implementations, step S6 specifically includes: The prediction result is formed by integrating the point prediction data with the confidence interval data. Risk rule matching is performed according to a preset early warning threshold system. When the prediction result meets the conditions of the preset early warning threshold system, a corresponding early warning signal is generated, and a power prediction result containing prediction data and early warning markers is output.

[0053] As described above, by integrating point predictions with confidence intervals and then matching them with risk rules, the prediction results gain a directly usable risk warning function. The principle is that point predictions are used to determine future trends, while confidence intervals are used to identify potential risks caused by uncertainty. When both meet preset threshold conditions, an early warning can be proactively triggered. By outputting power prediction results with accompanying warning tags, wind farms and grid dispatch can make adjustments in advance, such as increasing reserve capacity or adjusting load dispatch, thereby reducing operational risks caused by sudden fluctuations and improving overall operational safety.

[0054] In this embodiment, step S6 specifically includes: A comprehensive risk rule base and early warning level system have been established. Based on power system safety operation standards and actual wind farm operation experience, a rule base containing multiple dimensions of risk judgment criteria has been constructed. This rule base comprehensively considers factors such as power fluctuation characteristics, prediction uncertainty, and the impact of extreme weather, classifying risk levels into four levels: normal, attention, early warning, and critical. Each risk level corresponds to different threshold parameters, which are dynamically adjusted according to the wind farm's installed capacity, grid connection requirements, and seasonal characteristics. Power fluctuation risk is mainly identified by calculating the power change rate between adjacent time periods. When the change rate exceeds a set threshold, the corresponding level of early warning is triggered. Prediction results are evaluated based on the relative ratio of the confidence interval width to the point prediction value. When the confidence interval is too wide, it indicates insufficient prediction reliability, requiring an upgrade of the early warning level. In addition, the system also considers special risks under extreme weather conditions, such as typhoons, thunderstorms, and other severe weather events that may cause drastic power fluctuations.

[0055] During the risk indicator quantification calculation phase, the system automatically calculates multiple risk assessment indicators based on point prediction sequences and confidence interval data. The power volatility indicator is derived by calculating the magnitude of predicted power changes within a specific time window, reflecting the stability of power output. The prediction uncertainty indicator is based on the difference between the upper and lower bounds of the confidence interval, combined with point prediction values ​​for normalization, quantifying the reliability of the prediction results. The extreme event identification indicator comprehensively judges whether the point prediction values ​​are close to the operating boundaries of the wind turbine and whether there is abnormal expansion of the confidence interval. The calculation results of each risk indicator will serve as input for determining the warning level. The system also establishes a correlation analysis mechanism between indicators; when multiple indicators show abnormalities simultaneously, the final risk level will be appropriately increased to ensure the comprehensiveness and accuracy of risk identification.

[0056] Based on the quantitative results of risk indicators, the system performs early warning signal generation and decision analysis. A multi-level fuzzy comprehensive evaluation method is employed to map the values ​​of each risk indicator to corresponding early warning levels. First, the membership degree of each risk indicator to different early warning levels is calculated. Then, a weighted synthesis is performed according to pre-set weighting coefficients to ultimately determine the comprehensive early warning level. During early warning signal generation, the system considers not only the current risk status but also the development trend of the risk. When a sustained increase in the risk level is detected within a short period, the system generates a trend warning, prompting relevant personnel to pay attention to the evolution of the risk. Simultaneously, the system automatically generates corresponding handling suggestions and control strategies based on different early warning levels, including suggested backup capacity, possible adjustments to scheduling measures, and the activation level of emergency plans.

[0057] The system standardizes the encapsulation and output of forecast results and early warning information, integrating point forecast data, confidence interval data, early warning signals, and handling suggestions into a unified structured data product. Data encapsulation uses the standard JSON-LD format to ensure machine readability and semantic clarity. The output product includes complete metadata descriptions, including forecast generation time, effective period, data source, model version, and quality control indicators. Simultaneously, the system generates visual display interfaces tailored to different user groups, providing grid dispatchers with intuitive power forecast curves and prominent risk warnings, and wind farm operators with detailed forecast uncertainty analysis and operational recommendations. All output content adheres to power system data exchange standards, ensuring seamless integration with existing dispatch automation systems and providing strong support for reliable wind power absorption and the safe and stable operation of the power grid.

[0058] Please refer to Figure 3 A marine wind power prediction terminal 1 based on multi-source meteorological data includes a memory 2, a processor 3, and a computer program stored in the memory 2 and running on the processor 3. When the processor 3 executes the computer program, it implements the steps in a marine wind power prediction method based on multi-source meteorological data.

[0059] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for predicting offshore wind power based on multi-source meteorological data, characterized in that, Including the following steps: S1. Acquire weather forecast data, satellite remote sensing data and wind turbine real-time operation data to form multi-source heterogeneous data, and use an adaptive weighted fusion algorithm to dynamically assign weights and fuse the multi-source heterogeneous data to generate dynamic fused data. S2. Obtain micro-geographic information of the wind farm, combine it with the dynamic fusion data, construct a meteorological geographic coupled map using a graph neural network, and extract meteorological enhancement features; S3. Obtain the time attribute of the prediction task, combine it with the meteorological enhancement features, use the task-aware routing mechanism to perform optimal model matching and invocation, and determine the target prediction model. S4. Use the target prediction model and the meteorological enhancement features to perform time-series prediction and output point prediction data; S5. Based on the point prediction data and the dynamic fusion data, uncertainty is quantified using a probability distribution parameter generation method to obtain confidence interval data. S6. Integrate the point prediction data and the confidence interval data, perform risk rule matching and early warning signal generation, and form a power prediction result with risk warning.

2. The method for predicting offshore wind power based on multi-source meteorological data according to claim 1, characterized in that, Step S1 specifically includes: The multi-source heterogeneous data is spatiotemporally registered and unified to the same spatiotemporal grid system. An adaptive weighted fusion algorithm is used to dynamically allocate weights and fuse data in the same spatiotemporal grid system to generate dynamically fused data.

3. The method for predicting offshore wind power based on multi-source meteorological data according to claim 1, characterized in that, In step S1, an adaptive weighted fusion algorithm is used to dynamically assign weights and fuse the multi-source heterogeneous data to generate dynamically fused data, specifically including: A timeliness weight is established based on the forecast timeliness of the weather forecast data; quality weights are determined for the weather forecast data, satellite remote sensing data, and wind turbine real-time operation data based on the data timeliness, spatial coverage, and observation accuracy of the multi-source heterogeneous data; an adaptive weighted fusion algorithm is constructed based on the timeliness weights and quality weights, and the contribution of the weather forecast data, satellite remote sensing data, and wind turbine real-time operation data in the multi-source heterogeneous data is corrected using the adaptive weighted fusion algorithm, and the weather forecast data, satellite remote sensing data, and wind turbine real-time operation data with corrected contribution are combined to generate dynamic fused data.

4. The method for predicting offshore wind power based on multi-source meteorological data according to claim 1, characterized in that, Step S2 specifically includes: Based on the preset wind turbine spatial distribution map, a node map with wind turbine locations as nodes is constructed, and the micro-geographical information of the wind farm is obtained and injected into the node map; The dynamically fused data is mapped to the nodes in the node graph to form a spatiotemporal feature matrix; A graph neural network is used to extract the local relationships and topographic influences between nodes in the node graph, and a meteorological-geographical coupled map is constructed. Meteorological enhancement features are generated using the meteorological-geographical coupled map and the spatiotemporal feature matrix.

5. The method for predicting offshore wind power based on multi-source meteorological data according to claim 1, characterized in that, Step S3 specifically includes: Obtain the time attributes of the prediction task; The current weather model is obtained based on the analysis of the aforementioned meteorological enhancement features; Obtain the historical performance of candidate models from pre-established model performance records; The task-aware routing mechanism is used to match the time attribute with the current weather model to the corresponding candidate model. The matching degree of the candidate model's historical performance is calculated, and the candidate model with the highest matching degree is selected as the target prediction model.

6. The method for predicting offshore wind power based on multi-source meteorological data according to claim 5, characterized in that, In step S3, the task routing awareness mechanism includes: When the prediction task meets the ultra-short-term requirements, a time-series prediction model based on an attention mechanism is matched. When the prediction task is short-term and the current meteorological model is in a stable meteorological state, a lightweight gradient booster model is matched. When the current meteorological model indicates drastic weather changes, an extreme weather prediction model is used.

7. The method for predicting offshore wind power based on multi-source meteorological data according to claim 1, characterized in that, Step S4 specifically includes: The meteorological enhancement features are standardized and continuous time-series input samples are generated. The time-series input samples are input into the target prediction model, and the target prediction model is used to output point prediction data.

8. The method for predicting offshore wind power based on multi-source meteorological data according to claim 1, characterized in that, Step S5 specifically includes: An uncertainty analysis framework is constructed, in which the probability distribution form is determined based on the dynamic fusion data, and the probability distribution parameters are initialized based on the point prediction data. Confidence intervals are generated using quantile regression based on the probability distribution form and the probability distribution parameters.

9. The method for predicting offshore wind power based on multi-source meteorological data according to claim 1, characterized in that, Step S6 specifically includes: The prediction result is formed by integrating the point prediction data with the confidence interval data. Risk rule matching is performed according to a preset early warning threshold system. When the prediction result meets the conditions of the preset early warning threshold system, a corresponding early warning signal is generated and the power prediction result is output.

10. A marine wind power forecasting terminal based on multi-source meteorological data, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps in the offshore wind power prediction method based on multi-source meteorological data as described in any one of claims 1-9.

Images