CN-122022052-A - Short-term wind speed prediction method and system

CN122022052ACN 122022052 ACN122022052 ACN 122022052ACN-122022052-A

Abstract

The invention relates to a short-term wind speed prediction method and a system, belonging to the field of time sequence prediction, wherein the method comprises the steps of S1 data acquisition of a historical wind speed sequence and an exogenous meteorological element sequence, double-path modal decomposition of the wind speed sequence, modal screening and feature construction, hidden state weighted aggregation through time attention, attention-sLSTM modeling, obtaining a model for multi-step short-term wind speed prediction by adopting a mean square error loss optimization network parameter, and predicting wind speeds of a plurality of time steps in the future by using a trained model and outputting six main steps.

Inventors

SHU BIN
LUO QI
ZHAO YAPING
LUO JIN
LI YONGQIAN
Miao Yingqing
HE QIAN
ZHANG GUOXING

Assignees

云南省大气探测技术保障中心
国家气象中心(中央气象台、中国气象局气象导航中心)
云南省迪庆藏族自治州气象局
云南省嵩明县气象局

Dates

Publication Date: 20260512
Application Date: 20260204

Claims (10)

1. A short-term wind speed prediction method is characterized by comprising the following steps, S1, acquiring data, namely acquiring a historical wind speed sequence and an exogenous meteorological element sequence of a target wind power plant; s2, dual-path modal decomposition, namely respectively inputting the wind speed sequence into two decomposition branches of a variation modal decomposition VMD and a characteristic modal decomposition FMD to obtain a first sub-modal set and a second sub-modal set; s3, mode screening and feature construction, namely calculating the correlation between each sub-mode and target predicted quantity based on the Pearson correlation coefficient, and eliminating low correlation modes; S4, modeling the Attention-sLSTM, namely inputting the high-dimensional feature sequence 567 into a scalar long-short-term memory network sLSTM with an Attention mechanism, obtaining a hidden state sequence through stable gating and normalized memory updating, and carrying out weighted aggregation on the hidden states through time Attention; S5, performing supervision training based on a sample window, and optimizing network parameters by adopting mean square error loss to obtain a model for multi-step short-term wind speed prediction; s6, outputting, namely predicting and outputting wind speeds of a plurality of time steps in the future by using the trained model.
2. The short-term wind speed prediction method according to claim 1, wherein the variational modal decomposition VMD in S2 is performed by constraining the problem by: s2.1, decomposing and optimizing targets of variation modes: s2.1, changing a modal decomposition constraint condition: In the formula, Representing the number of target modalities, Representing the first intrinsic mode function IMF, Representing the center frequency corresponding to the first modality, symbol ∗ represents the convolution operation, To construct the hilbert core of the resolved signal, Representing the bandwidth measure, i.e. the derivative with respect to time, A variable of the time is represented and, Is an imaginary unit.
3. The short-term wind speed prediction method according to claim 1, wherein the characteristic pattern decomposition FMD branch in S2 performs filter adaptive iterative optimization based on a principle of maximum correlation kurtosis MCKD deconvolution, and wherein the correlation kurtosis CK is calculated by: where x (n) is a signal sequence, n represents a sample number, T is a period, and M represents a correlation order.
4. Short-term wind speed prediction method according to claim 1, characterized in that in step S3, the low correlation modes are rejected, in particular, a correlation threshold τ is set in the mode screening, and the modes with correlation lower than τ are rejected.
5. The method of claim 1, wherein the exogenous weather element comprises at least one of barometric pressure, wind direction, air temperature, dew point temperature, relative humidity, water vapor pressure, precipitation, and 0-320 cm multi-layer ground temperature.
6. A short-term wind speed prediction method according to claim 1, wherein, in S4, The sLSTM adopts exponential gating and introduces a stabilization state, performs stabilization constraint on an input gate and a forgetting gate, and performs normalization on a cell state and a hidden state to inhibit numerical divergence, so as to obtain a hidden state sequence, and specifically comprises the following sub-steps: s4.1, calculating the steady state quantity: Wherein, the Is the steady state quantity at the current moment; the steady state at the previous moment in time, And Is the forget gate output and forget gate bias term in the traditional LSTM; S4.2, calculating a stabilization index gate: Wherein, the And Respectively represent an input gate and a forgetting gate after stabilization treatment, Is an input gate in a conventional LSTM; s4.3, calculating a normalization state: s4.4, calculating normalization factors: s4.5, normalized hidden state output calculation: Wherein, the Is a candidate state for use in the method, Is the cell state quantity after normalization, Is a normalization factor that is used to normalize the data, Is a hidden state.
7. The method according to claim 1, wherein in S4, the attention mechanism is time attention, weights are assigned to hidden states of each time step and the hidden states are weighted and aggregated to alleviate information bottlenecks only depending on end time, and the method specifically comprises the following sub-steps: s4.6, calculating the time attention weight: S4.7, calculating a time attention expression vector: 。
8. the short-term wind speed prediction method according to claim 1, wherein after the VMD and the FMD-derived sub-modes are screened, high-dimensional features are combined to form a training wind speed prediction model.
9. A system for performing the short-term wind speed prediction method according to any one of claims 1 to 8, comprising, The data acquisition module is used for connecting a SCADA system or a meteorological database of the wind power plant data acquisition terminal to acquire a historical wind speed sequence and exogenous meteorological elements of the target wind power plant; the modal decomposition module comprises a VMD decomposition unit and an FMD decomposition unit which are arranged in parallel, wherein the VMD unit is configured to decompose a wind speed sequence into a low-frequency trend mode; The mode screening module is used for calculating the pearson correlation coefficient of each decomposition mode and the original sequence and automatically eliminating invalid modes according to a preset threshold or a sequencing result; The feature construction module is used for aligning and splicing the screened modal set and the preprocessed exogenous weather features in the time dimension to construct a multi-channel high-dimensional feature matrix; the time sequence prediction module is internally provided with an Attention-sLSTM network structure and comprises a stabilizing gate control sub-layer, a normalized memory updating sub-layer and a time Attention sub-layer, and is used for carrying out feature extraction and time sequence modeling on an input feature matrix; The training and reasoning module is used for loading training data to carry out iterative optimization on the model, calculating MSE loss and updating weight, and receiving real-time data and outputting future multi-step wind speed predicted values in the reasoning stage.
10. Use of the system of claim 9 in short term wind speed prediction.

Description

Short-term wind speed prediction method and system Technical Field The invention belongs to the field of time sequence prediction, and particularly relates to a short-term wind speed prediction method and a short-term wind speed prediction system. Background Wind energy is a clean, renewable low-carbon energy source and gradually plays an important role in global energy structure transformation. However, the uncertainty and intermittent characteristic of wind energy cause larger fluctuation of wind power output, and the improvement of economic benefit of a wind farm and the safety of power grid operation are severely restricted. Accurate short-term wind speed prediction is one of the key links of improving wind power utilization rate and optimizing power grid dispatching strategies. Wind speed prediction faces a number of technical challenges, mainly in the following aspects. First, wind speed sequences have highly non-stationary characteristics, and wind speeds have significant periodic and non-periodic wave characteristics on different time scales, and traditional single prediction models tend to have difficulty in capturing these complex modes effectively, thereby reducing prediction accuracy. Secondly, wind speed data has obvious multi-scale characteristics, namely wind speed change contains information of a plurality of frequency scales, such as fluctuation characteristics of day scale, hour scale and even minute scale, and a single-scale model is difficult to fully characterize all scale information. In addition, wind speed sequences often have complex nonlinear correlations with other meteorological factors, and neglecting these correlations tends to result in reduced prediction accuracy. Therefore, how to effectively extract multi-scale modal information and comprehensively utilize the space and time characteristics in meteorological data becomes an important research direction for improving wind speed prediction accuracy. Disclosure of Invention In order to overcome the problems in the background art, the invention provides a short-term wind speed prediction method and a short-term wind speed prediction system, which can effectively extract characteristics of different scales in wind speed data, reduce errors of non-stationary signal processing and improve the accuracy of wind speed prediction. In order to achieve the above purpose, the invention is realized by the following technical scheme: the short-term wind speed prediction method comprises the following steps: S1, acquiring data, namely acquiring a historical wind speed sequence and an exogenous meteorological element sequence of a target wind power plant; s2, dual-path modal decomposition, namely respectively inputting the wind speed sequence into two decomposition branches of a variation modal decomposition VMD and a characteristic modal decomposition FMD to obtain a first sub-modal set and a second sub-modal set; s3, mode screening and feature construction, namely calculating the correlation between each sub-mode and target predicted quantity based on the Pearson correlation coefficient, and eliminating low correlation modes; S4, modeling the Attention-sLSTM, namely inputting the high-dimensional feature sequence 567 into a scalar long-short-term memory network sLSTM with an Attention mechanism, obtaining a hidden state sequence through stable gating and normalized memory updating, and carrying out weighted aggregation on the hidden states through time Attention; S5, performing supervision training based on a sample window, and optimizing network parameters by adopting mean square error loss to obtain a model for multi-step short-term wind speed prediction; s6, outputting, namely predicting and outputting wind speeds of a plurality of time steps in the future by using the trained model. Further, the variant modal decomposition VMD described in S2 proceeds by constraining the problem as follows: s2.1, decomposing and optimizing targets of variation modes: s2.1, changing a modal decomposition constraint condition: In the formula, Representing the number of target modalities,Representing the first intrinsic mode function IMF,Representing the center frequency corresponding to the first modality, symbol ∗ represents the convolution operation,To construct the hilbert core of the resolved signal,Representing the bandwidth measure, i.e. the derivative with respect to time,A variable of the time is represented and,Is an imaginary unit. Further, the feature pattern decomposition FMD branch in S2 performs filter self-adaptive iterative optimization based on a principle of deconvolution of a maximum correlation kurtosis MCKD, wherein the correlation kurtosis CK is calculated by the following method: wherein x (n) is a signal sequence, n represents a sample sequence number, T is a period, and M represents a related order; further, in step S3, the mode with low correlation is specifically selected by setting a correlation threshold τ in the mode screeni