Wind power short-term output prediction method based on multi-modal data

Abstract

The present application relates to the field of artificial intelligence and power system prediction technology, and discloses a wind power short-term output prediction method based on multi-modal data, which comprises the following steps: acquiring multi-modal data such as historical output, numerical weather forecast, measured meteorology of a wind tower, topography and geomorphology, and wind turbine operating state; performing sliding window segmentation on the output sequence and identifying the mutation interval; calculating the local optimal alignment path of each mode in the mutation interval based on the dynamic time warping algorithm; performing non-uniform resampling accordingly to generate a time-synchronized multi-modal alignment feature sequence; inputting the hybrid neural network composed of a gated recurrent unit and an attention mechanism, and outputting high-precision output prediction values for the next 15 minutes. The system comprises corresponding functional modules. Through dynamic time alignment and cross-modal feature fusion, the present application significantly improves the short-term prediction accuracy of wind power, reduces the root mean square error by 23.7% in the mutation scenario, and provides reliable support for power grid dispatching.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and power system prediction technology, specifically relating to a method for predicting short-term wind power output based on multimodal data. Background Technology

[0002] As the penetration rate of renewable energy in the power system continues to increase, wind power, as an important form of clean energy, poses a significant challenge to grid dispatch and stable operation due to its output volatility. To ensure the safe and economical operation of the power grid, high-precision short-term wind power output forecasting has become a fundamental support for key applications such as smart grids, electricity market transactions, and energy storage collaborative control. Current mainstream forecasting methods generally rely on single or multi-source data (such as numerical weather prediction, SCADA operation data, radar images, etc.) to construct mapping relationships through machine learning or deep learning models. However, these multimodal data have fundamental differences in their physical acquisition mechanisms, resulting in variations in their sampling frequency, timestamp accuracy, and update cycle, creating inherent temporal asynchrony.

[0003] Short-term wind power output prediction methods based on multimodal data aim to integrate heterogeneous information such as meteorological, equipment status, and environmental perception to capture the instantaneous impact of non-steady-state processes such as sudden wind speed changes and turbulent disturbances on power generation. These methods typically assume that different modal data can be aligned on a unified time axis before being input into a joint model for feature extraction and prediction. However, in actual operation, when sudden events such as rapid ramp-up or drop in wind power output occur, the temporal misalignment between low-frequency sampled meteorological data (such as 15-minute interval NWP) and high-frequency SCADA data (such as 1-second power records) is drastically amplified, leading to severe distortion of key dynamic features during the alignment process.

[0004] Global time interpolation or fixed-window moving average strategies are often used to handle time-asynchronous issues. However, while these methods can maintain basic consistency during periods of stable wind power output, they introduce a large number of spurious smoothing signals or phase shifts during periods of abrupt changes, causing distortion of the feature space. More seriously, global interpolation operations accumulate and propagate noise along the time axis, weakening the model's sensitivity to actual abrupt changes and ultimately causing a nonlinear surge in prediction errors at critical moments. Especially under extreme weather conditions or abnormal equipment operating conditions, such errors may directly lead to the failure of dispatch instructions or misjudgment of reserve capacity. Summary of the Invention

[0005] This invention provides a method for short-term wind power output prediction based on multimodal data. It addresses the time asynchrony problem caused by significant differences in sampling frequencies of heterogeneous data from multiple sources, such as meteorological, geographical, and operational status data, by constructing a multimodal feature fusion architecture with dynamic time alignment capabilities. Especially during critical time intervals where wind power output changes abruptly, this method can accurately align the semantic correspondence of each modality's data along the physical time axis, eliminating feature distortion introduced by interpolation or downsampling operations. This generates a high-fidelity joint feature representation, which drives a high-precision short-term output prediction model.

[0006] This invention provides a method for predicting short-term wind power output based on multimodal data, which includes:

[0007] Acquire historical power output sequences of wind farms, numerical weather prediction data, meteorological data measured by wind measurement towers, topographic and geomorphological feature data, and real-time operating status data of wind turbines;

[0008] The historical output sequence is divided into sliding window segments to generate output time sequence segments with a preset time step.

[0009] The numerical weather forecast data, meteorological data measured by the wind tower, topographic and geomorphological feature data, and real-time operating status data of the wind turbine are marked with timestamps at the original sampling frequency, and the original time index set of each modal data is constructed based on the time boundary of the output time series segment.

[0010] Based on the original time index set of each modal data and the time boundary of the power output time series segment, the dynamic time warping algorithm is used to calculate the local optimal alignment path of each modal data in the power output mutation interval;

[0011] Based on the local optimal alignment path, non-uniform resampling is performed on each modal data to generate a multimodal alignment feature sequence that is strictly time-synchronized with the output time sequence segment;

[0012] The multimodal aligned feature sequence is input into a hybrid neural network model composed of gated recurrent units and attention mechanisms to perform cross-modal feature interaction and spatiotemporal dependency modeling, and output the wind power output prediction value within a future preset time window.

[0013] As one embodiment of the present invention, the acquisition of historical power output sequences of wind farms, numerical weather prediction data, meteorological data measured by meteorological towers, topographic and geomorphological feature data, and real-time operating status data of wind turbines specifically includes:

[0014] Historical power output sequences are obtained through the wind farm monitoring system interface, with a sampling frequency of once per second. The recorded content includes active power, reactive power, and power factor.

[0015] Numerical weather forecast data is obtained through the meteorological service interface, with a time resolution of once every 15 minutes. The recorded content includes wind speed, wind direction, air pressure, temperature, humidity and cloud cover.

[0016] The meteorological data is obtained by the sensor array of the wind tower. The sampling frequency is once per second, and the recorded content includes the wind speed and direction at heights of 10 meters, 30 meters, 50 meters and 80 meters.

[0017] Topographic and geomorphological feature data are obtained through a geographic information system database. This data is static spatial data and includes information such as altitude, slope, aspect, surface roughness, and obstacle distribution.

[0018] The real-time operating status data of the wind turbine is obtained through the wind turbine controller communication bus. The sampling frequency is once per second, and the recorded content includes pitch angle, yaw angle, generator speed, converter temperature and fault codes.

[0019] As one embodiment of the present invention, the step of dividing the historical power output sequence into sliding window segments to generate power output time series segments with a preset time step specifically includes:

[0020] Set the sliding window length to 120 time steps, the step size to 10 time steps, and each time step to 1 second;

[0021] The historical power output sequence is traversed, and 120 consecutive seconds of active power data are extracted as the power output time sequence segment.

[0022] Identify whether there are abrupt changes in the power change rate exceeding a preset threshold in each output time segment, where the preset threshold is 5% of the rated power per second;

[0023] If a mutation point exists, the output time segment is marked as a mutation interval sample; if no mutation point exists, it is marked as a stationary interval sample.

[0024] As one embodiment of the present invention, the construction of the original time index set of each modality data specifically includes:

[0025] For numerical weather prediction data, the time index is formed by the forecast release timestamp and the forecast validity timestamp, with each forecast validity timestamp covering a 15-minute time period;

[0026] For meteorological data measured by the wind measurement tower, real-time operating status data of the wind turbine, and historical output sequence, the precise timestamp of the data acquisition time is used as the time index;

[0027] For topographic and geomorphological feature data, since it is static data, its time index is set to be the same as the start timestamp of the power output time series segment.

[0028] As one embodiment of the present invention, the calculation of the local optimal alignment path of each modal data in the power output mutation interval using the dynamic time warping algorithm specifically includes:

[0029] For each output time segment marked as a mutation interval, extract the time windows 30 seconds before and after it as alignment analysis intervals;

[0030] Use the timestamps of the output sequence within the alignment analysis interval as the reference sequence;

[0031] The original time indices of other modal data within this alignment analysis interval are used as the sequences to be aligned.

[0032] The distance metric function is defined as a weighted sum of Euclidean distance and time offset penalty term, where the weight coefficient of the time offset penalty term is 0.3;

[0033] A constrained dynamic time warping algorithm is used to calculate the cumulative distance matrix between the reference sequence and each sequence to be aligned, under the premise of satisfying monotonicity, continuity and boundary conditions.

[0034] By tracing back the path with the minimum cumulative distance in the cumulative distance matrix, the local optimal alignment path of each modality data within the mutation interval is obtained.

[0035] As one embodiment of the present invention, the non-uniform resampling based on the locally optimal alignment path specifically includes:

[0036] For each modal data, along its locally optimal alignment path, determine its mapping source time point at each time step of the output time sequence segment;

[0037] If the original data exists at the source time point of the mapping, then the data value is used directly;

[0038] If the source time point of the mapping is located between two original data points, the eigenvalue of that time point is calculated using cubic spline interpolation.

[0039] The feature values ​​of all modal data at each time step of the output time series segment are concatenated to form a multimodal aligned feature sequence with a dimension equal to the number of time steps multiplied by the total dimension of the modal features.

[0040] As one embodiment of the present invention, the hybrid neural network model composed of gated recurrent units and attention mechanisms specifically includes:

[0041] A bidirectional gated recurrent unit layer with 256 hidden units is used to extract temporal features from multimodal aligned feature sequences in both forward and backward directions.

[0042] A multi-head self-attention layer with 8 heads, each head having a dimension of 32, is used to capture long-range dependencies between different time steps within a sequence.

[0043] The cross-modal attention layer uses query vectors derived from historical output features and key and value vectors derived from features of other modalities to dynamically weight the contribution of each modal information to the current prediction task.

[0044] The fully connected decoding layer contains two hidden layers with 128 and 64 nodes respectively, and the activation function is a modified linear unit. The number of nodes in the output layer is equal to the length of the prediction time window, and the activation function is a linear function.

[0045] As one embodiment of the present invention, the output of the predicted wind power output value within a future preset time window specifically includes:

[0046] The prediction time window is set to the next 15 minutes, corresponding to 900 time steps;

[0047] After inputting the multimodal aligned feature sequence into the hybrid neural network model, the model outputs a continuous numerical sequence of length 900.

[0048] The power limit processing is applied to this numerical sequence to ensure that it does not exceed the rated installed capacity of the wind farm and is not lower than 0.

[0049] The processed numerical sequence will be used as the final short-term wind power output prediction result.

[0050] This invention provides a wind power short-term output prediction system based on multimodal data, which includes:

[0051] The multimodal data acquisition module is used to acquire historical power output sequences of wind farms, numerical weather forecast data, meteorological data measured by wind measurement towers, topographic and geomorphological feature data, and real-time operating status data of wind turbines.

[0052] The output time sequence segment generation module is used to perform sliding window segmentation on the historical output sequence to generate output time sequence segments with a preset time step.

[0053] The original time index construction module is used to mark the timestamps of each modal data at the original sampling frequency, and to construct the original time index set of each modal data according to the time boundary of the output time sequence segment;

[0054] The local optimal alignment path calculation module is used to calculate the local optimal alignment path of each modal data in the power output mutation interval based on the original time index set of each modal data and the time boundary of the power output time sequence segment, using the dynamic time warping algorithm.

[0055] The multimodal alignment feature sequence generation module is used to perform non-uniform resampling on each modal data according to the local optimal alignment path to generate a multimodal alignment feature sequence that is strictly time-synchronized with the output time sequence segment;

[0056] The wind power output prediction module is used to input the multimodal aligned feature sequence into a hybrid neural network model composed of a gated recurrent unit and an attention mechanism, to perform cross-modal feature interaction and spatiotemporal dependency modeling, and to output the wind power output prediction value within a preset future time window.

[0057] In one embodiment of the present invention, the multimodal data acquisition module is specifically configured as follows:

[0058] Historical power output sequences are obtained through the wind farm monitoring system interface, with a sampling frequency of once per second.

[0059] Numerical weather forecast data is obtained through a meteorological service interface, with a time resolution of once every 15 minutes.

[0060] The meteorological data is obtained by the sensor array of the wind measurement tower, with a sampling frequency of once per second;

[0061] Topographic and geomorphological feature data are obtained from a geographic information system database;

[0062] The real-time operating status data of the wind turbine is obtained through the wind turbine controller communication bus, with a sampling frequency of once per second.

[0063] In one embodiment of the present invention, the output timing segment generation module is specifically configured as follows:

[0064] Set the sliding window length to 120 time steps and the step size to 10 time steps;

[0065] The historical output sequence is traversed and extracted to generate output time sequence segments;

[0066] Calculate the power change rate within each output time segment. If there is a point where the change rate exceeds 5% per second of the rated power, mark it as a mutation interval sample.

[0067] In one embodiment of the present invention, the local optimal alignment path calculation module is specifically configured as follows:

[0068] For samples with mutation intervals, extract time windows of 30 seconds before and after them;

[0069] Using the output sequence timestamp as the reference sequence, and the time indices of other modal data as the sequences to be aligned;

[0070] A constrained dynamic time warping algorithm is adopted, which combines Euclidean distance and time offset penalty term to calculate the cumulative distance matrix and backtrack the minimum path.

[0071] In one embodiment of the present invention, the multimodal aligned feature sequence generation module is specifically configured as follows:

[0072] The mapping source time point for each time step is determined along the locally optimal alignment path;

[0073] Missing points were filled using cubic spline interpolation.

[0074] The features of each modality are concatenated at each time step to form a unified multimodal aligned feature sequence.

[0075] In one embodiment of the present invention, the wind power output prediction module is specifically configured as follows:

[0076] It includes a bidirectional gated recurrent unit layer, a multi-head self-attention layer, a cross-modal attention layer, and a fully connected decoding layer;

[0077] The number of hidden units in the bidirectional gated loop unit layer is 256;

[0078] The multi-head self-attention layer has 8 heads, and each head has 32 dimensions.

[0079] The cross-modal attention layer uses historical output features as queries and other modal features as keys and values.

[0080] The fully connected decoding layer outputs the predicted values ​​for the next 900 time steps and performs power limiting processing.

[0081] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0082] This invention introduces a dynamic time warping algorithm to perform local optimal alignment of multimodal data in the wind power output abrupt change range, avoiding feature ambiguity and information loss caused by traditional fixed frequency interpolation methods during the rapid power change phase.

[0083] The non-uniform resampling strategy ensures the semantic consistency of each modality's data on the physical time axis, significantly improving the fidelity of feature fusion.

[0084] The hybrid neural network model, which combines gated recurrent units with temporal attention mechanisms, not only effectively captures the dynamic evolution of wind power output, but also achieves adaptive weighted fusion of information from different sources through cross-modal attention mechanisms. Attached Figure Description

[0085] Figure 1 This is a schematic diagram of the overall technical solution architecture of the wind power short-term output prediction method based on multimodal data proposed in this invention;

[0086] Figure 2This is a schematic diagram of the core principle framework of the dynamic time alignment driven multimodal feature fusion architecture in this invention;

[0087] Figure 3 This is a flowchart illustrating the logical flow of the generation of output timing segments and the identification of abrupt change intervals in this invention.

[0088] Figure 4 This is a flowchart illustrating the logical flow of multimodal data local optimal alignment path calculation and non-uniform resampling in this invention.

[0089] Figure 5 This is a flowchart illustrating the logical flow of cross-modal interaction and spatiotemporal dependency modeling using a hybrid neural network model in this invention.

[0090] Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the wind farm, meteorological system, wind measurement tower and wind turbine controller in this invention. Detailed Implementation

[0091] Please refer to Figures 1 to 6 This invention provides a method for short-term wind power output prediction based on multimodal data. Its core lies in addressing the time asynchrony problem caused by significant differences in sampling frequencies among heterogeneous data from multiple sources, such as meteorological, geographical, and operational status data. Especially during critical time intervals where wind power output changes abruptly, traditional fixed-frequency interpolation or downsampling operations disrupt the semantic correspondence of various modal data along the physical time axis, causing feature alignment distortion and leading to a surge in prediction errors. To overcome this technical obstacle, this invention constructs a multimodal feature fusion architecture with dynamic time alignment capabilities. It guides non-uniform resampling through a locally optimal alignment path to generate a high-fidelity joint feature representation, which then drives a hybrid neural network model for high-precision short-term output prediction.

[0092] The method includes the following steps:

[0093] S1, acquire historical power output sequence of wind farm, numerical weather forecast data, meteorological data measured by wind measurement tower, topographic and geomorphological feature data, and real-time operating status data of wind turbines;

[0094] S2, the historical output sequence is divided into sliding window segments to generate output time sequence segments with a preset time step.

[0095] S3, the numerical weather forecast data, the meteorological data measured by the wind tower, the topographic and geomorphological feature data, and the real-time operating status data of the wind turbine are marked with timestamps at the original sampling frequency, and the original time index set of each modal data is constructed according to the time boundary of the output time sequence segment.

[0096] S4. Based on the original time index set of each modal data and the time boundary of the output time series segment, the dynamic time warping algorithm is used to calculate the local optimal alignment path of each modal data in the output change interval.

[0097] S5. Based on the local optimal alignment path, non-uniform resampling is performed on each modal data to generate a multimodal alignment feature sequence that is strictly time-synchronized with the output time sequence segment.

[0098] S6, the multimodal aligned feature sequence is input into a hybrid neural network model composed of a gated recurrent unit and an attention mechanism to perform cross-modal feature interaction and spatiotemporal dependency modeling, and output the wind power output prediction value within the future preset time window.

[0099] In step S1, the historical power output sequence of the wind farm, numerical weather prediction data, meteorological data measured by the wind measurement tower, topographic feature data, and real-time operating status data of the wind turbines are acquired. Specifically, the historical power output sequence is acquired through the wind farm monitoring system interface, with a sampling frequency of once per second, recording active power, reactive power, and power factor. This data stream is stored with timestamps accurate to the millisecond level to ensure the reference accuracy for subsequent time alignment. Numerical weather prediction data is acquired through the meteorological service interface, with a time resolution of once every 15 minutes, recording wind speed, wind direction, air pressure, temperature, humidity, and cloud cover. This data is identified by the effective timestamp of the forecast, with each timestamp representing the center point of a 15-minute time period, and its coverage extends forward and backward by 7 minutes and 30 seconds. Measured meteorological data is acquired through the wind measurement tower sensor array, with a sampling frequency of once per second, recording wind speed and direction at heights of 10 meters, 30 meters, 50 meters, and 80 meters.

[0100] Each altitude-level sensor independently collects data and adds a local system timestamp, which is then synchronized to the Coordinated Universal Time (UTC) standard via a network time protocol. Topographic feature data is obtained from a Geographic Information System (GIS) database; this is static spatial data, recording information including altitude, slope, aspect, surface roughness, and obstacle distribution. This data does not change over time, but it needs to be bound to a specific time window to maintain data structure consistency when participating in feature fusion. Real-time wind turbine operating status data is acquired via the wind turbine controller communication bus, with a sampling frequency of once per second, recording information including pitch angle, yaw angle, generator speed, converter temperature, and fault codes. All data is transmitted in real-time to the central data processing unit via industrial Ethernet and stored in a distributed time-series database categorized by modality; each record contains complete timestamp metadata.

[0101] In step S2, the historical power output sequence is segmented using a sliding window to generate power output time series segments with a preset time step size. The sliding window length is set to 120 time steps, with a step size of 10 time steps, each time step corresponding to 1 second. Therefore, each power output time series segment covers 120 consecutive seconds of active power data. The historical power output sequence is traversed starting from the beginning, extracting 120 seconds of continuous data each time to form a segment. The window is then moved forward by 10 seconds, and this process is repeated until the end of the sequence. For each generated power output time series segment, the power change rate between any two adjacent seconds within it is calculated, i.e., the absolute value of the difference between the active power of the current second and the active power of the previous second divided by the rated installed capacity of the wind farm. If the power change rate at any time step exceeds 5% per second, the segment is determined to contain a power output abrupt change event and is marked as an abrupt change interval sample; otherwise, it is marked as a stable interval sample. This marking information serves as the trigger condition for whether to initiate the dynamic time alignment process subsequently. High-precision alignment is only performed on abrupt change interval samples to balance computational overhead and prediction accuracy.

[0102] In step S3, the numerical weather forecast data, meteorological data measured by the wind tower, topographic feature data, and real-time wind turbine operating status data are each timestamped at their original sampling frequencies. Based on the time boundaries of the output time series segments, a set of original time indices for each modality of data is constructed. For numerical weather forecast data, the time index is formed by the forecast release time stamp and the forecast validity time stamp, where the forecast validity time stamp is used to determine the effective coverage interval of the data on the time axis. For example, if the validity time of a forecast data point is 12:00 on a certain day, its actual coverage interval is from 11:52:30 to 12:07:30. For meteorological data measured by the wind tower, real-time wind turbine operating status data, and historical output sequences, the precise timestamp of the data acquisition time is used as the time index, with each timestamp corresponding to an instantaneous sampling point. For topographic feature data, since it is static data and does not have temporal evolution characteristics, its time index is set to be the same as the start timestamp of the current output time series segment, thus logically including it in the same time window for feature fusion. The construction process of the original time index set ensures that each modal data can be accurately mapped to the time boundary defined by the output time series segment while retaining its original sampling characteristics, providing a basic index structure for subsequent dynamic alignment.

[0103] In step S4, based on the original time index set of each modal data and the time boundary of the output time series segment, a dynamic time warping algorithm is used to calculate the local optimal alignment path of each modal data in the output abrupt change interval. This step is only performed on the output time series segment marked as the abrupt change interval sample. First, a 30-second time window is extracted before and after the segment, forming a total alignment analysis interval of 180 seconds. Within this interval, the timestamp of the output sequence is used as the reference sequence R, with a length of 180. The original time index of other modal data within this interval is used as the sequence to be aligned Q, with a length varying depending on the modality: the length of the wind tower data and turbine status data is also 180, the length of the topographic data is 1, and the length of the numerical weather prediction data is usually 12 (because there are 12 points within 15 minutes, and 12 points within 180 seconds). A distance metric function is defined. The weighted sum of the Euclidean distance and the time offset penalty term is expressed as: Where r1 is the feature vector of the i-th time step of the reference sequence (only the active power value is used here). Let J be the feature vector of the sequence to be aligned at the j-th time step. and These are their corresponding timestamps. The time offset penalty weight coefficient is set to 0.3. This design, while ensuring feature similarity, suppresses excessive time jumps, ensuring the physical rationality of the alignment path. Subsequently, a constrained dynamic time warping algorithm is adopted to recursively calculate the cumulative distance matrix C, under the premise of satisfying monotonicity (the path can only move forward), continuity (it cannot skip multiple rows or columns), and boundary conditions (the path must start from (1,1) and end at (m,n)). Here, C(i,j) represents the minimum cumulative distance from (1,1) to (i,j). Finally, by backtracking the minimum cumulative distance path in the C matrix, the locally optimal alignment path P between the reference sequence and the sequence to be aligned is obtained. This path is represented in the form of a series of (i,k) pairs, indicating that the i-th step of the reference sequence should be aligned with the k-th step of the sequence to be aligned.

[0104] In step S5, based on the locally optimal alignment path, non-uniform resampling is performed on each modal data to generate a multimodal aligned feature sequence that is strictly time-synchronized with the power output time series segment. For each time step i (i from 1 to 120) in the power output time series segment, the corresponding mapping source time point is found along the alignment path P. If the original data exists at this time point (e.g., the meteorological tower or wind turbine status data is sampled at this second), the data value is directly used; if the time point is between two original data points (e.g., numerical weather prediction data is not sampled at non-integer 15-minute intervals), the feature value of this time point is calculated using cubic spline interpolation. Cubic spline interpolation constructs a piecewise cubic polynomial based on four adjacent data points, ensuring that the first and second derivatives of the interpolation curve are continuous at the nodes, thereby preserving the smooth variation characteristics of the meteorological data. For topographic and geomorphic feature data, since it remains constant throughout the entire alignment analysis interval, its feature vector is directly copied to each time step. After filling the feature values ​​of all modalities at each time step, the modal features are horizontally concatenated at time step i to form a joint feature vector of dimension M, where M is equal to the sum of the dimensions of all modal features. This process is repeated to generate a 120-row, M-column multimodal aligned feature sequence, which is completely synchronized with the output time series segment in the time dimension and retains the original dynamic details of each modal data within the abrupt change interval.

[0105] In step S6, the multimodal aligned feature sequence is input into a hybrid neural network model composed of a gated recurrent unit and an attention mechanism to perform cross-modal feature interaction and spatiotemporal dependency modeling, and outputs the predicted wind power output value within a preset future time window. The hybrid neural network model contains four core components.

[0106] The first component is a bidirectional gated recurrent unit layer with 256 hidden units. This layer receives a 120×M input sequence and extracts temporal features from both forward and backward directions, outputting a 120×512 context-aware feature sequence, where the feature vector at each time step incorporates past and future contextual information.

[0107] The second component is a multi-head self-attention layer with 8 heads, each with a dimension of 32. This layer models the internal temporal relationships of the output of the bidirectional gated recurrent unit. By calculating the attention weights between different time steps, it captures long-range dependencies, such as the impact of early wind speed changes on later power output.

[0108] The third component is a cross-modal attention layer, whose query vector comes from the historical output feature subspace, while the key and value vectors come from other modal feature subspaces. This layer dynamically evaluates the relevance of external information such as meteorology, terrain, and operational status to the current output prediction task, and then weights and fuses this information to achieve selective information interaction between modalities.

[0109] The fourth component is a fully connected decoding layer, which contains two hidden layers. The first layer has 128 nodes and the second layer has 64 nodes, and the activation function for both is a modified linear unit. The output layer has 900 nodes, corresponding to a prediction time window of 15 minutes (900 seconds) in the future, and the activation function is a linear function.

[0110] The model outputs a continuous numerical sequence of length 900. This sequence is then subjected to power limiting processing, setting all values ​​less than 0 to 0 and values ​​greater than the rated installed capacity to the rated capacity, ensuring the prediction results conform to physical constraints. The final processed sequence is the short-term wind power output prediction result.

[0111] The aforementioned method effectively addresses the feature distortion problem caused by the temporal asynchrony of multimodal data by introducing a dynamic time warping-driven local alignment mechanism within the abrupt change interval. The non-uniform resampling strategy avoids the smoothing effect of global interpolation on abrupt change details, preserving key information during rapid changes in wind power output. The hybrid neural network model, combining temporal modeling and attention mechanisms, achieves deep interaction and adaptive fusion of multi-source information, significantly improving prediction accuracy, especially in power ramp-up and sudden drop scenarios.

[0112] Furthermore, this invention also provides a wind power short-term output prediction system based on multimodal data for executing the above-described method. This system includes a multimodal data acquisition module, an output time-series segment generation module, an original time index construction module, a local optimal alignment path calculation module, a multimodal alignment feature sequence generation module, and a wind power output prediction module.

[0113] The multimodal data acquisition module is configured to acquire historical power output sequences via the wind farm monitoring system interface, with a sampling frequency of once per second; acquire numerical weather forecast data via the meteorological service interface, with a time resolution of once every 15 minutes; acquire measured meteorological data via the wind measurement tower sensor array, with a sampling frequency of once per second; acquire topographic feature data via the geographic information system database; and acquire real-time wind turbine operating status data via the wind turbine controller communication bus, with a sampling frequency of once per second. This module is deployed on the local server of the wind farm and communicates with external systems through a secure isolation gateway to ensure the real-time performance and security of data acquisition.

[0114] The power output time series segment generation module is configured with a sliding window length of 120 time steps and a step size of 10 time steps. It iterates through historical power output sequences, extracting segments to generate output time series data. The module calculates the power change rate within each segment; if a point exists where the change rate exceeds 5% per second of the rated power, it is marked as a sudden change interval sample. This module incorporates a high-efficiency sliding window algorithm and supports parallel processing of historical data streams from multiple wind turbines.

[0115] The raw time index construction module is configured to add a timestamp at the original sampling frequency to each modal data and construct a raw time index set for each modality based on the time boundaries of the output time series segments. This module maintains a unified time coordinate system to ensure that all modal data are comparable on the logical time axis.

[0116] The local optimal alignment path calculation module is configured to extract 30-second time windows before and after the abrupt change interval samples; using the output sequence timestamp as the reference sequence and the time indices of other modal data as the sequences to be aligned; employing a constrained dynamic time warping algorithm, combined with Euclidean distance and a time offset penalty term, to calculate the cumulative distance matrix and backtrack to the minimum path. This module is implemented using optimized dynamic programming and supports GPU acceleration to handle the computational load under high-dimensional features.

[0117] The multimodal alignment feature sequence generation module is configured to determine the mapping source time point for each time step along the locally optimal alignment path; fill missing points using cubic spline interpolation; and concatenate the modal features at each time step to form a unified multimodal alignment feature sequence. This module integrates a high-precision numerical computation library to ensure the stability and accuracy of the interpolation results.

[0118] The wind power output prediction module is configured to include a bidirectional gated recurrent unit layer, a multi-head self-attention layer, a cross-modal attention layer, and a fully connected decoding layer. The bidirectional gated recurrent unit layer has 256 hidden units; the multi-head self-attention layer has 8 heads, each with a dimension of 32; the cross-modal attention layer uses historical output features as queries and other modal features as keys and values; the fully connected decoding layer outputs predicted values ​​for the next 900 time steps and performs power limiting processing. This module is built on a deep learning framework, supporting online model updates and incremental training to adapt to the long-term evolution of wind farm operating characteristics.

[0119] In summary, this invention achieves high-precision time alignment and feature fusion of multimodal wind power data through the collaborative design of methods and systems, significantly improving the robustness and accuracy of short-term power output prediction in abrupt change scenarios.

Claims

1. A method for predicting short-term wind power output based on multimodal data, characterized in that, include: Acquire historical power output sequences of wind farms, numerical weather prediction data, meteorological data measured by wind measurement towers, topographic and geomorphological feature data, and real-time operating status data of wind turbines; The historical output sequence is divided into sliding window segments to generate output time sequence segments with a preset time step. The numerical weather forecast data, meteorological data measured by the wind tower, topographic and geomorphological feature data, and real-time operating status data of the wind turbine are marked with timestamps at the original sampling frequency, and the original time index set of each modal data is constructed based on the time boundary of the output time series segment. Based on the original time index set of each modal data and the time boundaries of the output time series segments, a dynamic time warping algorithm is used to calculate the local optimal alignment path of each modal data in the output mutation interval, including: For each output time series segment marked as a mutation interval, a 30-second time window before and after it is extracted as an alignment analysis interval. The timestamp of the output sequence within this alignment analysis interval is used as the reference sequence. The original time indices of other modal data within this alignment analysis interval are used as the sequences to be aligned. The distance metric function is defined as a weighted sum of Euclidean distance and a time offset penalty term, where the weight coefficient of the time offset penalty term is 0.

3. A constrained dynamic time warping algorithm is used to calculate the cumulative distance matrix between the reference sequence and each sequence to be aligned, under the premise of satisfying monotonicity, continuity, and boundary conditions. The minimum cumulative distance path in the cumulative distance matrix is ​​backtracked to obtain the local optimal alignment path of each modal data within this mutation interval. Based on the locally optimal alignment path, non-uniform resampling is performed on the modal data to generate a multimodal alignment feature sequence that is strictly time-synchronized with the output time sequence segment, including: For each modal data, along the locally optimal alignment path, determine the mapping source time point at each time step of the output time series segment; if the original data exists at the mapping source time point, the original data is used directly; if the mapping source time point is located between two original data points, the feature value of the mapping source time point is calculated using cubic spline interpolation; the feature values ​​of all modal data at each time step of the output time series segment are concatenated to form a multimodal alignment feature sequence with a dimension equal to the number of time steps multiplied by the total dimension of the modal features; The multimodal aligned feature sequence is input into a hybrid neural network model composed of gated recurrent units and attention mechanisms to perform cross-modal feature interaction and spatiotemporal dependency modeling, and output the wind power output prediction value within a future preset time window. The hybrid neural network model includes: A bidirectional gated recurrent unit layer with 256 hidden units is used to extract temporal features from multimodal aligned feature sequences in both forward and backward directions. A multi-head self-attention layer with 8 heads and 32 dimensions for each head is used to capture long-range dependencies between different time steps within the output sequence of the bidirectional gated recurrent unit layer. In the cross-modal attention layer, the query vector comes from historical output features, while the key and value vectors come from features from other modalities. This is used to dynamically weight the contribution of each modal information to the current prediction task. The fully connected decoding layer contains two hidden layers with 128 and 64 nodes respectively, and the activation function is a modified linear unit. The number of nodes in the output layer is equal to the length of the prediction time window, and the activation function is a linear function.

2. The wind power short-term output prediction method based on multimodal data according to claim 1, characterized in that, Acquire historical power output sequences, numerical weather prediction data, meteorological data measured by wind measurement towers, topographic and geomorphological feature data, and real-time operating status data of wind turbines, including: Historical power output sequences are obtained through the wind farm monitoring system interface, with a sampling frequency of once per second. The recorded content includes active power, reactive power, and power factor. Numerical weather forecast data is obtained through the meteorological service interface, with a time resolution of once every 15 minutes. The recorded content includes wind speed, wind direction, air pressure, temperature, humidity and cloud cover. The meteorological data measured by the wind tower is obtained through the sensor array of the wind tower. The sampling frequency is once per second, and the recorded content includes the wind speed and wind direction at heights of 10 meters, 30 meters, 50 meters and 80 meters. Topographic and geomorphological feature data are obtained through a geographic information system database. This data is static spatial data and includes information such as altitude, slope, aspect, surface roughness, and obstacle distribution. The real-time operating status data of the wind turbine is obtained through the wind turbine controller communication bus. The sampling frequency is once per second, and the recorded content includes pitch angle, yaw angle, generator speed, converter temperature and fault codes.

3. The wind power short-term output prediction method based on multimodal data according to claim 2, characterized in that, The historical power output sequence is segmented using a sliding window to generate power output time series segments with a preset time step, including: Set the sliding window length to 120 time steps, the step size to 10 time steps, and each time step to 1 second; The historical power output sequence is traversed, and 120 consecutive seconds of active power data are extracted as the power output time sequence segment. Identify whether there are abrupt changes in the power change rate exceeding a preset threshold in each output time segment, where the preset threshold is 5% of the rated power per second; If a mutation point exists, the output time segment is marked as a mutation interval sample; if no mutation point exists, it is marked as a stationary interval sample.

4. The wind power short-term output prediction method based on multimodal data according to claim 3, characterized in that, Construct a set of raw time indexes for each modality of data, including: For numerical weather prediction data, the time index is formed by the forecast release timestamp and the forecast validity timestamp, with each forecast validity timestamp covering a 15-minute time period; For meteorological data measured by the wind measurement tower, real-time operating status data of the wind turbine, and historical output sequence, the precise timestamp of the data acquisition time is used as the time index; For topographic and geomorphological feature data, since it is static data, its time index is set to be the same as the start timestamp of the power output time series segment.

5. The wind power short-term output prediction method based on multimodal data according to claim 4, characterized in that, Output the predicted wind power output within a future preset time window, including: The prediction time window is set to the next 15 minutes, corresponding to 900 time steps; After inputting the multimodal aligned feature sequence into the hybrid neural network model, the hybrid neural network model outputs a continuous numerical sequence of length 900. The continuous numerical sequence is subjected to power limiting processing to ensure that it does not exceed the rated installed capacity of the wind farm and is not lower than 0. The processed continuous numerical sequence is used as the final short-term wind power output prediction result.

6. The wind power short-term output prediction method based on multimodal data according to claim 5, characterized in that, The cubic spline interpolation method constructs a piecewise cubic polynomial based on four adjacent original data points, ensuring that the first and second derivatives of the interpolation curve are continuous at the nodes.

7. The wind power short-term output prediction method based on multimodal data according to claim 6, characterized in that, In the constrained dynamic time warping algorithm, the boundary condition is that the alignment path must start from the first element of the reference sequence and the sequence to be aligned, and end at the last element.

Images