Wind speed prediction method and device based on fft-rankpooling-lstm

Abstract

This invention relates to the field of wind speed prediction technology, and discloses a wind speed prediction method and apparatus based on FFT-Rankpooling-LSTM. The method includes: acquiring multi-source time-series wind speed data; generating frequency domain features through normalization and FFT decomposition; selecting core features using a ranking pooling algorithm and constructing a spatiotemporal correlation data matrix; fusing features and reducing dimensionality through a multi-channel attention mechanism to generate a time-series feature encoding vector; building a prediction evolution model based on LSTM network modeling and optimizing hidden state propagation; and finally generating the result through multi-step rolling prediction. The apparatus includes modules such as frequency domain decomposition, feature pooling, spatiotemporal modeling, and a prediction engine, supporting data smoothing, visualization, and adaptive parameter optimization. This invention improves the accuracy and generalization ability of wind speed prediction by integrating frequency domain analysis, hierarchical feature pooling, and spatiotemporal correlation modeling, and is applicable to multi-step wind speed prediction scenarios in fields such as wind power and meteorology.

Claims

1. A wind speed prediction method based on FFT-Rankpooling-LSTM, characterized in that, Includes the following steps: Step T1: Obtain multi-source wind speed time-series data for the target area; The multi-source wind speed time series data are normalized and preprocessed, and then frequency domain feature decomposition is performed to generate wind speed frequency domain feature data. Step T2: Perform hierarchical feature pooling based on wind speed frequency domain feature data and filter key patterns to obtain the core feature set of wind speed; Step T3: Perform multi-scale temporal window truncation on the core feature set of wind speed to collect multi-dimensional time segments; perform spatial correlation alignment on the multi-dimensional time segments, and then perform dynamic weight allocation to construct a spatiotemporal correlation data matrix; Step T4: Perform multi-channel feature fusion on the spatiotemporal correlation data matrix and hierarchical feature dimensionality reduction to generate temporal feature encoding vectors; Step T5: Model the temporal feature encoding vector using a long short-term memory network, optimize the hidden state propagation, and construct a wind speed prediction evolution model; Step T6: Perform multi-step rolling prediction simulation on the wind speed prediction evolution model to generate wind speed prediction simulation response data; Wind speed prediction results for the target area are generated based on simulation response data for wind speed prediction.

2. The wind speed prediction method based on FFT-Rankpooling-LSTM according to claim 1, characterized in that, The specific steps of step T1 are as follows: Step T11: Collect raw wind speed time-series data of the target area through a meteorological sensor network and perform data cleaning to eliminate outliers; Step T12: Perform sliding window segmentation on the cleaned wind speed time series data to generate a continuous time segment sequence; Step T13: Perform a Fast Fourier Transform on each time segment sequence to extract the frequency domain energy distribution features; Step T14: Divide the frequency domain energy distribution characteristics into frequency bands and calculate the energy proportion weight of each frequency band; Step T15: Weighted fusion of frequency domain features based on energy proportion weight to generate wind speed frequency domain feature data.

3. The wind speed prediction method based on FFT-Rankpooling-LSTM according to claim 1, characterized in that, The specific steps of step T2 are as follows: Step T21: Perform time-frequency joint analysis on the wind speed frequency domain characteristic data to generate a time-frequency joint distribution map; Step T22: Sort the time-frequency joint distribution map by feature importance based on the sorting pooling algorithm, and select the top K significant features; Step T23: Perform spatial topological association analysis on salient features and construct a dependency graph between features; Step T24: Cluster the features according to the dependency graph to generate feature subset clusters; Step T25: Redundant features are removed from the feature subset clusters, retaining the feature combinations with the strongest independence; Step T26: Generate a core feature set of wind speed based on the combination of independent features.

4. The wind speed prediction method based on FFT-Rankpooling-LSTM according to claim 1, characterized in that, The specific steps of step T3 are as follows: Step T31: Perform multi-scale truncation of the core wind speed feature set according to the preset time window length to generate short-term, medium-term and long-term time segments; Step T32: Perform spatial gridding on time segments of different scales and calculate the feature similarity of grid nodes; Step T33: Align time segments across scales based on feature similarity and construct a time segment association mapping table; Step T34: Dynamically assign weights to time segments based on the association mapping table to generate a weighted time segment set; Step T35: Reorganize the weighted time segment set into a matrix to construct a spatiotemporal correlation data matrix.

5. The wind speed prediction method based on FFT-Rankpooling-LSTM according to claim 4, characterized in that, The specific steps of step T35 are as follows: Calculate the spatial grid node coordinates for each time segment; Extend the time dimension of the weighted time segment set to generate a three-dimensional spatiotemporal tensor. Spatial interpolation filling of the 3D spatiotemporal tensor is performed based on node coordinates to eliminate sparse data regions. Channel separation and merging operations are performed on the filled spatiotemporal tensor to generate a spatiotemporal correlation data matrix.

6. The wind speed prediction method based on FFT-Rankpooling-LSTM according to claim 1, characterized in that, The specific steps of step T4 are as follows: Step T41: Perform multi-channel feature extraction on the spatiotemporal correlation data matrix to generate channel feature vectors; Step T42: Perform cross-channel attention calculation on the channel feature vectors to generate channel attention weights; Step T43: Weight and fuse the channel feature vectors based on attention weights to generate a fused feature vector; Step T44: Perform hierarchical principal component analysis on the fused feature vector and retain the first N principal components; Step T45: Normalize and standardize the principal components to generate temporal feature encoding vectors.

7. The wind speed prediction method based on FFT-Rankpooling-LSTM according to claim 1, characterized in that, The specific steps of step T5 are as follows: Step T51: Divide the temporal feature encoding vector into time step blocks to generate input sequence blocks; Step T52: Construct a multi-layer LSTM network and set the parameters for the forget gate, input gate, and output gate; Step T53: Initialize the hidden state of the input sequence block and propagate the hidden state step by step; Step T54: Dynamically update the hidden state through a gating mechanism to generate a hidden state evolution sequence; Step T55: Map the hidden state evolution sequence to a fully connected layer to construct a wind speed prediction evolution model.

8. The wind speed prediction method based on FFT-Rankpooling-LSTM according to claim 1, characterized in that, The specific steps of step T6 are as follows: Step T61: Configure multi-step prediction parameters for the wind speed prediction evolution model and set the rolling prediction step size; Step T62: Iteratively update the input sequence based on the prediction results at the current time step to generate the input for future time steps; Step T63: Perform model inference for each future time step input to generate a single-step prediction result; Step T64: Accumulate single-step prediction results and update the historical data window to generate a multi-step rolling prediction sequence; Step T65: Post-process and smooth the multi-step rolling prediction sequence to generate wind speed prediction simulation response data.

9. A wind speed prediction device based on FFT-Rankpooling-LSTM, characterized in that, A method for performing wind speed prediction based on FFT-Rankpooling-LSTM as described in any one of claims 1 to 8, comprising: The frequency domain decomposition module acquires multi-source wind speed time-series data and performs normalization preprocessing, and generates wind speed frequency domain feature data through fast Fourier transform. The feature pooling module performs hierarchical filtering and redundancy removal of frequency domain features based on the sorted pooling algorithm to generate a core feature set of wind speed. The spatiotemporal modeling module performs multi-scale truncation and spatial alignment of the core feature set, constructs a spatiotemporal correlation data matrix, and performs feature fusion. The prediction engine module models the temporal feature encoding vector through a long short-term memory network and performs multi-step rolling prediction to generate wind speed prediction results.

10. The wind speed prediction device based on FFT-Rankpooling-LSTM according to claim 9, characterized in that, Also includes: The post-processing module performs data smoothing and error correction on the prediction results; The visualization module maps the prediction results to a geographic information system and generates a wind speed distribution heatmap. The adaptive optimization module dynamically adjusts model parameters and feature weights based on real-time prediction errors.

Images