Double-path ultra-short-term wind power prediction method based on numerical weather forecast and multi-order time sequence dynamic gating fusion

Abstract

A double-path ultra-short-term wind power prediction method based on numerical weather forecast and multi-order time sequence dynamic gating fusion comprises the following steps: acquiring wind power generation historical data and numerical weather forecast data of a wind power plant, and screening weather factors highly related to wind power by using an MIC; the CEEMDAN is adopted to decompose the power sequence into a plurality of intrinsic mode functions (IMF); a dual-path prediction architecture is constructed, one path adopts xLSTM to predict an intrinsic mode function (IMF), all subsequences are superposed, and a prediction result is obtained; in the other path, the XGBoost is combined with key meteorological characteristics of an intrinsic mode function (IMF) and a numerical weather forecast (NWP) for prediction, and all the subsequences are superposed to obtain a prediction result; the method comprises the following steps: designing an MT-DGFusion module through an enhanced attention and dynamic gating network; and fusing the dual-path prediction results through an MT-DGFusion module to obtain a final prediction result. According to the method, double breakthrough of prediction precision and stability is realized, and a new technical path is provided for a complex time sequence prediction task.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind power prediction, in particular to a double-path ultra-short-term wind power prediction method based on numerical weather prediction and multi-order time series dynamic gating fusion. BACKGROUND

[0002] With the global energy structure transforming to clean energy, wind power as a kind of renewable energy, the installed capacity and grid penetration rate are continuously increasing in the global range. However, the volatility and uncontrollability of wind energy lead to great uncertainty of wind power, which makes the grid-connection and consumption of wind power, grid dispatching and power system stability face severe challenges. Therefore, accurate and stable wind power prediction has important practical significance for the optimal operation of wind farms, the improvement of grid safety and the efficient use of new energy.

[0003] The prediction method of wind power can be mainly divided into two categories: physical modeling-based method and data-driven method. The traditional physical modeling method usually relies on wind turbine power curve, high-precision meteorological data and fluid dynamics model, and analyzes the meteorological conditions and wind turbine characteristics to derive the wind power. However, due to the complexity of wind farm terrain, significant local turbulence effect, and strong randomness of wind speed and direction, it is difficult for physical modeling to accurately depict the dynamic changes of wind power. In contrast, data-driven methods based on machine learning and deep learning can directly learn the nonlinear mapping relationship of wind power from historical data, thereby improving the prediction accuracy.

[0004] For machine learning methods, such as support vector regression (SVR), random forest (RF), and extreme gradient boosting (XGBoost), they perform well in high-dimensional feature extraction and pattern recognition, and are widely used in meteorological data analysis and wind power prediction. Among them, support vector regression (SVR) has strong generalization ability in dealing with small sample problems and nonlinear regression tasks, but its computational complexity is high and the training efficiency is low when facing large-scale data. Random forest (RF) can effectively reduce the risk of overfitting and perform well in handling high-dimensional features due to the advantage of ensemble learning, but its model interpretability is relatively low due to its tree-based structure. Extreme gradient boosting (XGBoost) optimizes tree models through weighted boosting strategy, and has significant advantages in computational efficiency and feature importance evaluation, so it shows good performance in wind power prediction tasks.

[0005] While deep learning methods such as Convolutional Neural Networks (CNN), Gated Recurrent Units (GRU), and Long Short-Term Memory Networks (LSTM) have shown remarkable advantages in modeling time series dependencies, they can effectively capture the temporal characteristics of wind power and improve prediction accuracy. Among them, Convolutional Neural Networks (CNN) can extract local spatio-temporal features in wind power data through local perception and weight sharing mechanism, and perform well in capturing short-term trends, but due to the lack of long-term dependency modeling capability, there are certain limitations in processing long time span data. Gated Recurrent Units (GRU) as an improved version of Recurrent Neural Networks (RNN), through the gating mechanism, reduces the gradient vanishing problem of traditional RNN, and is superior to Long Short-Term Memory Networks (LSTM) in terms of computational efficiency and memory occupancy, but in dealing with complex temporal dependencies, its performance is still slightly inferior to Long Short-Term Memory Networks (LSTM).

[0006] In deep learning, LSTM is widely used due to its ability to handle temporal dependencies, and its core innovation lies in the three-gate mechanism: input gate, forget gate, and output gate. This mechanism uses sigmoid function to control information flow and tanh function to generate candidate states, allowing the model to dynamically adjust the information storage and forgetting process of memory cells, thereby effectively alleviating the gradient vanishing problem of RNN and enhancing the ability to capture long-range dependencies of time series data. In wind power prediction, this mechanism has obvious advantages. The input gate can adaptively adjust the response amplitude to power mutations, the forget gate helps to retain historical information of periodic power fluctuations, and the output gate determines which power fluctuation features are passed to the next time. However, LSTM still has certain limitations, which affect its performance in more complex tasks. First, LSTM uses scalar cell state for information storage, which has information bottlenecks when modeling complex patterns and rare events, limiting the long-term memory capacity of the model. Second, the update of LSTM's hidden state depends on sequential computation between time steps, making it difficult to achieve efficient parallel processing, resulting in limited training efficiency on large-scale datasets. In addition, when input features are highly similar, LSTM's gating mechanism may have difficulty dynamically adjusting stored information, affecting the sensitivity to mutation features. To address these issues, researchers have proposed xLSTM, which optimizes LSTM structure to improve its modeling capacity and computational efficiency.

[0007] Long Short-Term Memory Networks (LSTM) effectively overcome the long-term dependency problem by introducing memory cells and gating structures, allowing it to accurately model long-term temporal relationships in wind power prediction tasks and improve prediction stability. Subsequently, various improved versions based on LSTM have been proposed.

[0008] In addition, the change of wind power not only depends on historical power data, but also is affected by future meteorological conditions such as wind speed, wind direction, temperature, humidity, etc. Numerical weather prediction (NWP) can provide meteorological data for a certain period in the future, which is an important data source for ultra-short-term wind power prediction. However, NWP data has high-dimensional characteristics, complex spatio-temporal dependence and error accumulation problems, which may lead to large uncertainty when directly using NWP data for prediction. Therefore, how to effectively integrate NWP data and historical wind power data has become one of the key challenges to improve prediction accuracy.

[0009] Wind power data usually exhibits strong nonlinearity and non-stationarity, and it is difficult to directly capture its complex patterns by relying solely on machine learning or deep learning models. Therefore, signal decomposition techniques are widely used in wind power prediction to enhance the feature extraction capability of the model. For non-stationary signal processing, empirical mode decomposition (EMD) has created an adaptive signal processing paradigm, but it is affected by mode mixing. Ensemble empirical mode decomposition (EEMD) introduces Gaussian white noise to alleviate this problem, but it leads to reconstruction error. Complementary ensemble empirical mode decomposition (CEEMD) improves completeness by positive and negative noise cancellation, but still has residual noise and computational efficiency problems. Adaptive noise complete ensemble empirical mode decomposition (CEEMDAN) uses an adaptive noise injection strategy to suppress mode mixing and improves reconstruction accuracy through phased residual decomposition and dynamic mode number determination mechanism, becoming one of the best signal decomposition solutions.

[0010] Although a single prediction model has its own advantages, it still has limitations in the complex scenario of wind power prediction. A single model is often difficult to capture both long-term trends and short-term fluctuations in wind power, and it is also difficult to effectively integrate historical power data and NWP information, thereby limiting the overall prediction capability. Especially in the case of sudden changes in wind speed or extreme weather conditions, the adaptability of the model to sudden changes is weak, and the prediction accuracy and stability are easily affected. To improve the robustness of prediction, multi-model fusion methods are widely studied, and the key technology is to build a complementary mechanism between heterogeneous models. Currently, multi-model fusion methods are mainly divided into two categories: series and parallel. In the series method, the pre-sequence model is responsible for feature extraction, and its output is used as the input of the subsequent model, such as the CNN-LSTM hybrid model proposed by researchers, which first uses CNN to extract local spatio-temporal features of wind power, and then combines LSTM for time series modeling to achieve efficient fusion of multi-dimensional information. In contrast, parallel fusion emphasizes the collaborative prediction of multiple base models, and some researchers have proposed a non-negative constraint fusion framework that determines static weight coefficients through optimization algorithms to integrate the outputs of multiple models. However, this framework lacks adaptive weight adjustment mechanism in the time series dynamic scenario, making it difficult to cope with the complexity of wind power changes over time.

[0011] In recent years, the integration method combining machine learning and deep learning has gradually become a research hotspot. A hybrid model based on CNN and LightGBM is proposed, which first extracts features from the power data of multiple wind farms using convolutional neural network (CNN), and then uses LightGBM for wind power prediction, making full use of the feature extraction ability of deep learning and the efficient regression ability of traditional machine learning. Although the multi-model fusion method has made certain progress in wind power prediction, it still faces many problems: 1) The dynamic association mechanism between models is imperfect. The existing method mainly calculates the weight based on historical data, and usually uses static fusion strategies such as weighted average or voting, which lacks adaptive adjustment ability and is difficult to fully exert the advantages of each model; 2) The generalization ability under extreme weather scenarios is weak. The existing prediction framework highly depends on historical time series data, and the fusion of real-time weather forecast information is still insufficient, resulting in a lag in response when dealing with sudden weather changes, and limiting the prediction stability and accuracy. SUMMARY

[0012] In view of the above technical problems, the present application provides a dual-path ultra-short-term wind power prediction method based on numerical weather prediction and multi-order time series dynamic gating fusion. First, the maximum information coefficient (MIC) is used to select weather factors highly related to wind power, and the power sequence is decomposed into multiple intrinsic mode functions (IMF) using adaptive complete ensemble empirical mode decomposition (CEEMDAN); Then, a dual-path prediction architecture is constructed, one path uses extended long short-term memory network (xLSTM) to predict intrinsic mode functions (IMF), and the other path uses extreme gradient boosting (XGBoost) to combine IMF and key meteorological features of numerical weather prediction (NWP) for prediction; Finally, the prediction results of each IMF component are superimposed to obtain the dual-path prediction result. In addition, the method designs a multi-order time series dynamic gating fusion (MT-DGFusion) module to dynamically fuse the dual-path prediction results through enhanced attention and gating network.

[0013] The technical scheme adopted by the present application is:

[0014] The dual-path ultra-short-term wind power prediction method based on numerical weather prediction and multi-order time series dynamic gating fusion comprises the following steps:

[0015] Step 1: Obtain the historical data of wind power generation of the wind farm and the numerical weather prediction data, and use the maximum information coefficient MIC to select weather factors highly related to wind power;

[0016] Step 2: Use adaptive complete ensemble empirical mode decomposition CEEMDAN to decompose the power sequence into multiple intrinsic mode functions IMF;

[0017] Step 3: Construct a double-path prediction architecture, one path uses an extended long short-term memory network (xLSTM) to predict the intrinsic mode function IMF, and superimposes all sub-sequences to obtain the prediction result;

[0018] Another path uses XGBoost to combine the intrinsic mode function IMF and the key meteorological features of numerical weather prediction NWP to make predictions, and superimposes all sub-sequences to obtain the prediction result;

[0019] Step 4: Design the MT-DGFusion module through enhanced attention and dynamic gating network;

[0020] Step 5: Fuse the double-path prediction results through the MT-DGFusion module to obtain the final prediction result.

[0021] Step 6: Use the mean absolute error MAE, mean square error MSE, and determination coefficient R 2 These three evaluation indexes are used to compare and analyze the prediction results.

[0022] In step 1, the basic idea of the maximum information coefficient MIC is to grid the data and calculate the mutual information under each grid division, and then find the maximum value of the mutual information divided by the complexity of the grid division. The complexity of grid division is usually measured by the number of grids or the fineness of grid division. The calculation formula of MIC is represented as:

[0023]

[0024] In formula (1), MIC(X, Y) represents the maximum information coefficient of variables X and Y, X and Y are two variables, a x b represents the grid division method, that is, X is divided into a intervals and Y is divided into b intervals, and the value of a x b is less than a given upper limit B to avoid overfitting; I(X; Y| G ) represents the mutual information between X and Y under the given grid G division; log2min(a, b) is the complexity measure of grid division, which is used to normalize the mutual information.

[0025] The heat map of the maximum information coefficient is shown in Figure 5 As can be seen from Figure 5 , the weather factor highly related to wind power is wind speed. In step 2, CEEMDAN is used as the power decomposition method, which is an improved version of EEMD, aiming to overcome the mode aliasing and reconstruction error problems caused by noise introduction in EEMD. Its core idea is to improve the decomposition quality through adaptive noise injection and step-by-step residual decomposition. The algorithm steps are as follows:

[0026] S2.1: Set the original sequence as x(t), the noise standard deviation coefficient as ε0, the set number as M, the initial residual error as r0(t) = x(t), and the decomposition order as k = 1;

[0027] S2.2: Generate M noisy signals:

[0028] x (i) (t) = r0(t) + ε0·E1(ω (i) (t)), i = 1, 2, …, M (2);

[0029] In equation (2), x (i) (t) represents the i th noisy signal, ω (i) (t) is the i th white noise, E1(·) represents the first IMF obtained by EMD decomposition of the noise; and M represents the number of noisy signals.

[0030] S2.3: Perform EMD decomposition on each x (i) (t) to extract the first IMF thereof. Calculate the average value:

[0031]

[0032] In equation (3), IMF1(t) represents the first subsequence obtained by CEEMDAN decomposition; represents the i th subsequence obtained by EMD decomposition of x (i) (t).

[0033] Update the residual error:

[0034] r1(t) = r0(t) - IMF1(t) (4);

[0035] In equation (4), r1(t) represents the updated residual error.

[0036] S2.4: For k = 2, generate M noisy residual errors:

[0037]

[0038] In equation (5),: represents the k-1 th residual error generated after the i th noise addition, r k-1 (t) represents the residual error of the original signal after decomposition of the first k-1 orders, ε k-1 represents the adaptive noise amplitude coefficient, E k (·) is the k th IMF obtained by EMD decomposition of the noise; ω (i) (t) represents the i th added Gaussian white noise, ε k = ε0·σ(r k-1 ), and σ is the standard deviation.

[0039] S2.5: For each noisy residual EMD decomposition is performed to extract the first The average value is calculated:

[0040]

[0041] In formula (6): IMF k (t) represents the ensemble average result of the kth order intrinsic mode function.

[0042] Update the residual:

[0043] r k (t) = r k-1 (t) - IMF k (t) (7).

[0044] In formula (7): r k-1 (t) represents the remaining residual signal after k-1 order decomposition.

[0045] When the residual r k (t) is a monotonic function or has less than 2 extreme points, stop decomposition. The final signal decomposition is:

[0046]

[0047] In formula (8): X(t) represents the decomposition signal; r K (t) represents the final residual.

[0048] In step 3, the extended long short-term memory network (xLSTM) is a deep neural network architecture constructed based on a residual module, which is composed of scalar long short-term memory network (sLSTM) and matrix long short-term memory network (mLSTM) module stacking. As two innovative extensions of LSTM, sLSTM reconstructs the gating mechanism of traditional LSTM by introducing an exponential function to replace the Sigmoid activation function, including the formula: i t ,f t ,o t = exp(W·[h t-1 ,x t ]+b), i t ,f t ,o t are the input gate, the forgetting gate, and the output gate, respectively, exp() is the exponential function, W is the weight matrix, h t-1 is the hidden state at the previous moment, and x tis the input vector at the current time, b is the bias term; to enhance the expression ability of the gating signal, and at the same time, the scalar memory mechanism across time steps is fused in the memory unit, which significantly improves the control ability of the model to the timing information. And the mLSTM extends the scalar memory unit to a matrix structure, and combines the parameterized memory correlation matrix, so that the hidden state update can be fully parallelized in the sequence dimension, including the formula: M t is the memory matrix at the current time, M t-1 is the memory matrix at the previous time, F t , I t , O t are the forget gate matrix, input gate matrix and output gate matrix respectively, is the candidate memory matrix, A is the correlation matrix, is the element-wise multiplication, tanh(·) is the hyperbolic tangent activation function, H t is the hidden state matrix at the current time; thereby greatly improving the calculation efficiency. The independent module structure of the two is shown in detail in Figure 1 .

[0049] The architecture design of xLSTM realizes functional complementation by alternately stacking sLSTM and mLSTM modules: the sLSTM module focuses on fine timing modeling, while the mLSTM module uses matrix calculation to break through the sequential limitation of traditional RNN. This stacking mode realizes gradient propagation across layers through residual connection, and the hierarchical combination mode and parameter sharing mechanism are presented in Figure 2 . In addition, xLSTM supports dynamically adjusting the proportion of the two types of modules according to task requirements: when the task requires strong sequence dependency, increase the proportion of sLSTM, and when high parallelism is required, increase the proportion of mLSTM. The present application adopts the SMS architecture, and the proportion of sLSTM and mLSTM is 2:1. By embedding the configurable hybrid architecture into the residual framework, xLSTM not only retains the core modeling capability of LSTM, but also significantly improves the training efficiency through the matrix parallelization of mLSTM. At the same time, the residual design effectively alleviates the degradation problem of deep network, providing an efficient solution for complex sequence tasks.

[0050] In step 3, the principle process of the sLSTM module is as follows:

[0051] The traditional LSTM uses a Sigmoid function to regulate the input gate and the forget gate, and its output range is limited to the [0, 1] interval, which causes the dynamic range of the gating signal to be compressed, affecting the flexible regulation of information flow. sLSTM has made an innovative improvement: first, replace the gating activation function with an exponential function to expand the gating signal range to [0, +∞). The specific formula is i t ,f t ,o t =exp(W·[ht-1 ,x t ]+b), i t ,f t ,o t They are input gate, forget gate, output gate, exp() is exponential function, W is weight matrix, h t-1 is the hidden state of the previous moment, x t is the input vector at the current moment, and b is the bias term; this significantly enhances the nonlinear regulation capability of the gating mechanism on information flow;

[0052] Secondly, by introducing the normalized state mechanism, the weight relationship between the input gate and the forget gate is dynamically balanced during the calculation process. The specific formula is: is the normalized input gate, is the normalized forget gate, Ct is the memory unit at the current moment, C t-1 is the memory unit of the previous moment, is the candidate memory, and ⊙ is the element-by-element multiplication; this effectively alleviates the problem of tight gate coupling in traditional LSTM.

[0053] In addition, sLSTM breaks through the traditional single memory unit architecture, supports the parallel operation of multiple independent memory units, and realizes cross-unit memory mixing through recursive connections. The core content is: each memory unit It has its own input gate, forget gate, and output gate, which independently updates its own memory according to traditional LSTM rules and calculates the interaction weights between all memory units. E is a learnable weight matrix, Softmax ensures that the sum of the weights of each row is 1, A ij Represents the proportion of memory contribution of unit j to unit i. Each unit receives the weighted sum of the memory of other units: The diagonal terms (A kk ) retains the unit's own memory, and the off-diagonal items (A kj ) integrates the memories of other units; this design enables the model to establish memory associations at different time scales, achieving efficient capture of complex patterns and long-range state tracking capabilities in time series modeling. Its forward transfer process can be systematically described by the following formula:

[0054]

[0055] In formula (9): z t is the candidate unit state, represents the transpose of the input weight matrix, x t The input vector at the current moment, r z represents the recursive weight matrix, h t-1 represents the hidden state of the previous moment, b z represents the bias term.

[0056]

[0057] in Equation (10): i t is an input gate, denotes the transpose of an input weight matrix, r i denotes a recurrent weight coefficient, h t-1 denotes a hidden state at a previous time, b i denotes a bias term, and exp(·) denotes an exponential function.

[0058]

[0059] in Equation (11): f t is a forget gate, denotes the transpose of an input weight matrix, r f denotes a recurrent weight coefficient, b f denotes a bias term.

[0060]

[0061] in Equation (12): o t is an output gate, denotes the transpose of an input weight matrix, r o denotes a recurrent weight coefficient, b o denotes a bias term.

[0062] m t = max(log(f t ) + m t-1 , log(i t )) (13);

[0063] in Equation (13): m t is a stabilizer applied to i t and f t to prevent gradient explosion from generating i′ t and f t ′; log(f t ) denotes a logarithm of a forget gate, m t-1 denotes a stabilizer at a previous time, and log(i t ) denotes a logarithm of an input gate.

[0064] i′ t = exp(log(i t ) - m t ) (14);

[0065] in Equation (14): i′ t denotes a normalized input gate.

[0066] f t= exp(log(f t )+m t-1 -m t ) (15);

[0067] In formula (15), f t ' represents the normalized forget gate.

[0068] c t = f t '·c t-1 +i′ t ·z t (16);

[0069] In formula (16), c t is a cell state, c t-1 represents a cell state at a previous moment, and z t represents a candidate cell state.

[0070] n t = f t '·n t-1 +i′ t (17);

[0071] In formula (17), n t is a normalized state, and n t-1 represents a normalized state at a previous moment.

[0072] h t = o t (c t / n t )(18);

[0073] In formula (18), h t is a hidden state.

[0074] In step 3, the principle process of the mLSTM module is as follows:

[0075] The mLSTM significantly expands the storage capacity of the model in the parameter space dimension by expanding the scalar memory cell into a high-dimensional matrix memory structure, and improves the information association and extraction efficiency through the matrix key-value pair association mechanism. The specific formula is C t is a matrix memory state at a current moment, v t is a value vector, and k t is a key vector.

[0076] The design innovatively integrates the core idea of bidirectional associative memory, adopts an updating rule based on a covariance matrix, and realizes fast storage and retrieval of key-value pairs through implicit coding. Specifically, in the memory matrix updating process, mLSTM establishes a dynamic association between input features and hidden states through covariance operation, and realizes accurate retrieval based on similarity calculation of the attention mechanism. The formula is output = C t ·q t ; q t is a query vector, and C t is a memory matrix.

[0077] To realize a hardware-friendly computing paradigm, mLSTM eliminates the memory mixing mechanism of traditional LSTM, completely eliminates the time sequence dependency constraint in sequence processing, and makes the hidden state update of all time steps realize global parallelization calculation through matrix multiplication operation. The formula is: Where cumsum represents the cumulative sum. Its forward propagation process can be described by the following formula:

[0078] q t = W q x t +b q (19);

[0079] In formula (19), q t is a query input, W q is the weight matrix corresponding to q t , and b q represents a bias term.

[0080]

[0081] In formula (20), k t is a key input, W k is the weight matrix corresponding to k t , d represents the dimension of the key vector, and b k represents a bias term.

[0082] v t = W v x t +b v (21);

[0083] In formula (21), v t is a value input, W v is the weight matrix corresponding to v t , x t represents the input vector at the current time, and b v represents a bias term.

[0084]

[0085] In formula (22), i t represents an activation value of an input gate, represents a transpose of a weight vector of an input gate, b i represents a bias term.

[0086] represents taking an exponential of a result of a linear transformation.

[0087]

[0088] In formula (23), f t represents an activation value of a forget gate, represents a transpose of a weight vector of a forget gate, b f represents a bias term.

[0089] o t = σ(w o x t +b o )(24);

[0090] In formula (24), o t represents an activation value of an output gate, w o represents a weight vector of an output gate, b o represents a bias term.

[0091] m t = max(log(f t )+m t-1 , log(i t ))(25);

[0092] In formula (25), m t represents a stabilizer applied to i t and f t , log(f t ) represents a logarithm of an activation value of a forget gate, m t-1 represents a logarithm memory state of a previous time, and log(i t ) represents a logarithm of an activation value of an input gate.

[0093] i t ′ = exp(log(i t )-m t ) = exp(i t -m t )(26);

[0094] In formula (26), i t ′ represents a normalized input gate.

[0095] f t ′ = exp(log(f t )+mt-1 -m t )(27);

[0096] f in formula (27) is: t ′ represents a normalized forget gate.

[0097]

[0098] C in formula (28) is: t represents a memory matrix at the current moment, f t ′ represents a normalized forget gate, C t-1 represents a memory matrix at the previous moment, i′ t represents a normalized input gate, v t represents a value vector, represents the transpose of a key vector.

[0099] n t = f t ′·n t-1 +i′ t ·k t (29);

[0100] n in formula (29) is: t represents the normalized state at the current moment, n t-1 represents the normalized state at the previous moment, k t represents the current input key vector.

[0101]

[0102] h in formula (30) is: t is a hidden state, o t represents an output gate, C t represents a memory matrix at the current moment, q t represents a query vector, represents the normalized state at the current moment, to ensure numerical stability and avoid gradient explosion.

[0103] The hidden state calculation includes calculating a context-aware vector using the query vector q t and the normalized state n t , and then performing element-wise multiplication with the output gate signal o t and the normalized version of the cell state C t , denoted by the symbol ⊙, to finally generate the hidden state output with both temporal dependence and feature selectivity.

[0104] In step 3, XGBoost is an ensemble learning algorithm based on gradient boosting trees. The prediction output of XGBoost is integrated by K decision trees, and the sample xi The predicted value is:

[0105]

[0106] In formula (31): is the model prediction output, f k is the function of the kth tree, x i is the i-th sample.

[0107] The objective function of XGBoost consists of a loss function and a regularization term:

[0108]

[0109] In the above formula, represents the overall objective function, is the mean square error between the predicted value and the actual value; i represents the true value of the i-th sample, n represents the total number of samples, Ω(f k ) represents the regularization term of the k-th tree, T k is the number of leaf nodes of the kth tree; w k is the leaf node weight; γ and λ are the leaf number penalty coefficient and L2 regularization coefficient respectively.

[0110] At the tth iteration, Xgboost has fixed the first t-1 trees and needs to optimize the tth tree f t The objective function is simplified to:

[0111]

[0112] In formula (34): represents the overall objective function at the tth iteration, represents the cumulative prediction value of the first t-1 trees for sample i, f t (x i ) represents the predicted output of the t-th tree for sample i; represents the loss function value of sample i in round t, Ω(f t ) represents the regularization term of the t-th tree.

[0113] Perform a second-order Taylor expansion on the loss function to For expansion point:

[0114]

[0115] In formula (35): represents the prediction error of the first t-1 trees on sample i; g i represents the first-order gradient; h i represents the second-order gradient; First and second order gradients of the loss function, respectively.

[0116] Group samples by leaf node, set I i = {i | x i ∈ j}; I i represents the set of all sample indexes divided into leaf j; i represents the sample index; j represents the leaf node number.

[0117] The objective function can be rewritten as:

[0118]

[0119] In equation (36), T is the number of leaf nodes of the current tree; w j represents the output weight of leaf j; h i represents the second order gradient; γ represents the leaf split penalty coefficient.

[0120] Derive w j and set the derivative to zero to get the optimal weight:

[0121]

[0122] In equation (37), w represents the optimal weight of leaf j, I j represents the set of samples divided into leaf j.

[0123] Substitute w into the objective function to get the minimum loss:

[0124]

[0125] In equation (38), L represents the minimum loss value of the t-th round, and T represents the total number of leaf nodes.

[0126] In step 4, the MT-DGFusion module constructs multi-order feature representation with time sequence correlation for the fusion requirements of the double prediction model; given the outputs of two prediction models at time t and and the corresponding actual observation value y t , construct time sequence dynamic features by introducing the lag operator L k . Specifically, for each prediction model, construct two historical features with lag steps;

[0127]

[0128] In equation (39), h represents the historical prediction of model a in the previous 1 step, represents the historical prediction of model a in the previous 2 steps, represents the historical prediction of model b one step ahead, Represents the historical prediction of model b 2 steps ahead. represents the prediction of model a at time t-1, represents the prediction of model a at time t-2, represents the prediction of model b at time t-1, represents the prediction of model b at time t-2, represents the predicted output of model a at time t-1, represents the predicted output of model a at time t-2, represents the predicted output of model b at time t-1, Represents the predicted output of model b at time t-2.

[0129] The input feature matrix constructed in this way contains the current moment prediction value and its historical information:

[0130]

[0131] In formula (40): X t represents the original feature vector, represents the prediction of model a at time t, represents the prediction of model b at time t, represents the t-1 time prediction of model a, represents the t-2 time prediction of model a, represents the prediction of model b at time t-1, Represents the prediction of model b at time t-2.

[0132] Feature Matrix Extract high-order time series features through multi-layer nonlinear transformation:

[0133]

[0134] H (2) =LayerNorm(W2H (1) +b2)(42);

[0135] H=Dropout(H (2) ) (43);

[0136] In the above formula, H (1) represents the first layer hidden state, H (2) Represents the second layer hidden state, H represents the final output feature, W1 and W2 are learnable weight matrices, LeakyReLU is an activation function with a negative slope, LayerNorm is a layer normalization operation, and Dropout is used for regularization; represents an input feature matrix, b1 represents a first layer bias term, and b2 represents a second layer bias term. Subsequently, the features are decomposed into multi-head attention mechanisms:

[0137] MultiHead(H)=Concat(head1,...,head h )W O (44);

[0138]

[0139] In the above formula, MultiHead(H) represents the final output after the input features H are processed by the multi-head attention mechanism;

[0140] head1,...,head h represent the output results of the first to the hth attention heads, respectively;

[0141] Concat(head1,...,head h ) represents concatenating the outputs of the h attention heads in the feature dimension;

[0142] head i represents the calculation result of the ith attention head; represents the calculation process of the ith attention head;

[0143] is a projection matrix for each head, and W O is an output projection matrix.

[0144] wherein the calculation of each attention head adopts scaled dot-product attention:

[0145]

[0146] In formula (46), Attention(Q, K, V) represents scaled dot-product attention, Q represents a query matrix, K represents a key matrix, and V represents a value matrix, K T represents the transpose of the key matrix, d k is a dimension scaling factor of the key vector.

[0147] The multi-head attention output A is dynamically gated and fused with the original features H:

[0148] G=σ(W g [H;A]+b g ) (47);

[0149] F=G⊙H+(1-G)⊙A (48);

[0150] wherein G is a gating weight, W gis the gating weight matrix, is the Sigmoid function, and represents element-wise multiplication g is the gating bias term, W g [H; A] represents the feature linear transformation, and F represents the fused feature.

[0151] Meanwhile, the auxiliary prediction path is constructed to retain the original feature information:

[0152]

[0153] In equation (49), R represents the auxiliary feature representation, W r is the projection weight matrix, is the original input feature, and b r is the auxiliary path bias.

[0154]

[0155] In equation (50), R is the auxiliary prediction output, W r2 is the output weight matrix, ReLU(R) represents the activated feature, R represents the auxiliary feature representation, and b r2 is the output bias.

[0156] The final prediction is realized by the superposition of the main path and the auxiliary path:

[0157]

[0158] wherein: is the final prediction output, and MLP(F) represents the main path prediction.

[0159] The MLP adopts a dimension reduction structure to suppress noise interference through a bottleneck design. This dual-path structure not only retains the abstract expression ability of deep features but also maintains the physical interpretability of original features. The overall structure is shown in Figure 3 .

[0160] In step 6, to quantitatively evaluate the differences in model prediction performance, the present application selects the mean absolute error (MAE), the root mean square error (RMSE), and the coefficient of determination (R 2 ) as multi-dimensional evaluation indexes, and the specific mathematical expressions are as follows:

[0161]

[0162] In the above equation, y i represents the true value, represents the predicted value, N is the sample number, and y represents the mean of the true value.

[0163] The technical effects of a double-path ultra-short-term wind power prediction method based on numerical weather prediction and multi-order time series dynamic gate fusion are as follows:

[0164] 1) A physical information constrained data preprocessing framework is constructed, MIC meteorological characteristics are used, and CEEMDAN is used to realize adaptive separation of power signal noise and effective components;

[0165] 2) A double-channel heterogeneous prediction architecture is designed, the xLSTM network is used in the time series channel to capture the long-term dependence of the power, the XGBoost enhanced tree coupled with the NWP multi-element feature is developed in the meteorological channel, and the rapid response of the meteorological mutation scene is realized;

[0166] 3) A multi-order time series dynamic gate fusion mechanism (MT-DG Fusion) is proposed, a lag time series feature module is constructed, multi-scale correlation is captured through multi-head attention, and a double-path gate network is designed to dynamically fuse the xLSTM and XGBoost prediction results. 4) Ablation and comparative experiments are carried out on two data sets to verify the effectiveness and superiority of the proposed method. The ablation experiment systematically evaluates the contribution of each component, such as the MT-DG module, the double-path structure and the NWP data. At the same time, a comparative analysis is carried out on multiple basic models, and the objective performance indicators show the superiority of the model in multiple scenarios.

[0167] 5) The present application proposes an extended long short-term memory network (xLSTM), which significantly improves the ability of the model to capture long time series dependence by optimizing the memory unit and the gating mechanism compared with the traditional LSTM. Multiple benchmark test results show that the extended long short-term memory network (xLSTM) far exceeds the traditional LSTM in performance, which provides a more advanced prediction method for wind power prediction. BRIEF DESCRIPTION OF DRAWINGS

[0168] The present application will be further described below in conjunction with the drawings and examples:

[0169] Figure 1 It is the internal structure diagram of sLSTM and mLSTM unit of the present application.

[0170] Figure 2 It is the xLSTM module structure diagram of the present application.

[0171] Figure 3 It is the multi-order time series dynamic gate fusion module structure diagram of the present application.

[0172] Figure 4 It is the flowchart of the proposed prediction method of the present application.

[0173] Figure 5 It is the maximum information coefficient thermogram of the present application.

[0174] Figure 6 The power CEEMDAN decomposition result diagram of data set 1 and data set 2 of the application.

[0175] Figure 7 The fusion result and contribution degree analysis diagram of data set 1 and 2 of the application.

[0176] Figure 8 The prediction result comparison and evaluation index diagram of data set 1 of the application.

[0177] Figure 9 The prediction result comparison and evaluation index diagram of data set 2 of the application.

[0178] Figure 10 The comparison test result diagram of different benchmark models of the application in data set 1;

[0179] Figure 11 The comparison test result diagram of different benchmark models of the application in data set 2. DETAILED DESCRIPTION

[0180] The application will be further described in detail below in combination with the drawings:

[0181] The application proposes a dual-path collaborative prediction framework based on NWP data and a dynamic gating fusion module. First, historical power data and NWP data are collected from a wind farm, and correlation analysis is performed thereon to explore the relationship between meteorological factors and wind power. On this basis, the wind power data is decomposed into multiple modes to better capture different scale features in the data. Next, the application constructs a dual-path prediction architecture, in which the time series path is predicted by an xLSTM model to capture the temporal dependence in the power sequence; while the meteorological path is predicted by an XGBoost model, which models the key meteorological features in the power subsequence and NWP. Finally, the MT-DG module fuses the prediction results of the two paths, optimizes the fusion strategy through a dynamic gating mechanism, and further improves the prediction accuracy. The prediction process is shown in Figure 4 , which specifically includes the following steps:

[0182] Step a: Obtain the historical wind power data and numerical weather prediction data of the wind farm. Considering that the historical information of the wind farm includes meteorological information such as wind speed, air temperature, air pressure, and historical power data, in order to avoid input redundancy, the MIC is used to select the associated meteorological data. The MIC values of various factors and power are shown in Figure 5 , it can be seen from Figure 5 that among the main factors affecting wind power, the correlation of wind speed is the largest, so the predicted wind speed is selected as a multi-factor variable.

[0183] The calculation formula of MIC is expressed as:

[0184]

[0185] Where X and Y are two variables, a×b represents the grid division method, that is, X is divided into a intervals, Y is divided into b intervals, and the value of a×b is less than a given upper limit B (to avoid overfitting), I(X; Y| G ) represents the mutual information between X and Y under a given grid G ​​partitioning, log2min(a,b) is the complexity measure of the grid partitioning, which is used to normalize the mutual information.

[0186] Step b: Use CEEMDAN to decompose the power sequence into multiple IMFs. Through parameter pre-experiments (noise amplitude ε = 0.2, number of collections N = 200), the decomposition parameters are optimized and the optimal number of IMFs is determined to be 9. The decomposition results are as follows: Figure 6 As shown, the algorithm steps are as follows:

[0187] Assume that the original sequence is x(t), the noise standard deviation coefficient is ε0, the number of collections is M, the initial residual r0(t) = x(t), and the decomposition order k = 1.

[0188] Generate M noise signals:

[0189] x (i) (t)=r0(t)+ε0·E1(ω (i) (t)),i=1,2,…,M

[0190] where ω (i) (t) is the i-th white noise, and E1(·) represents the first IMF obtained by EMD decomposition of the noise.

[0191] For each x (i) (t) Perform EMD decomposition and extract its first IMF Calculate the average:

[0192]

[0193] Update residuals:

[0194] r1(t)=r0(t)-IMF1(t)

[0195] For k=2, M noisy residuals are generated:

[0196]

[0197] where E1(·) is the kth order IMF obtained by EMD decomposition of the noise, ε k =ε0·σ(r k-1), σ is the standard deviation.

[0198] For each noisy residual EMD decomposition is performed to extract the first IMF The average value is calculated:

[0199]

[0200] Update the residual:

[0201] r k (t) = r k-1 (t) - IMF k (t)

[0202] When the residual r k (t) is a monotonic function or has less than 2 extreme points, stop decomposition. The final signal decomposition is:

[0203]

[0204] Step c: Build a double-path prediction architecture, one path uses xLSTM to predict IMF and superimposes all sub-sequences to get the prediction result. xLSTM builds a time series prediction model based on the IMF sub-sequences decomposed by CEEMDAN, uses a 10-time step sliding window input (SMS layer structure, batch size 32, learning rate 0.001, 100 rounds of MSE optimization), and generates a prediction result through IMF reconstruction.

[0205] Step d: The other path uses XGBoost to combine IMF and NWP key meteorological features for prediction, and superimposes all sub-sequences to get the prediction result; NWP-XGBoost parallel fusion of meteorological features and IMF components to build a multi-factor prediction model (100 decision trees, depth 5, learning rate 0.001, 100 rounds of MSE optimization), and outputs the prediction result after IMF reconstruction.

[0206] Step e: The double-path prediction results are fused through the MT-DGFusion module, a fusion training set is built based on the validation set data, the input features include the prediction values of the two models and their lag items (t-1, t-2 time), and the target value is the actual historical power. The time series feature interaction layer uses 4 heads of attention mechanism (hidden layer dimension 32) to capture the cross-period correlation of the prediction sequence, and the dynamic gate decision layer generates adaptive weights through the Sigmoid function to adjust the prediction contribution of the two models in real time. Key parameters are determined through pre-experiment verification, including initial learning rate 0.001, 300 iterations of training and 64 batches of data division.

[0207] Step f: Use the mean absolute error MAE, root mean square error RMSE and determination coefficient R 2The three evaluation indexes are used for comparative analysis of the prediction results.

[0208] Step f1: comparing the proposed model with the dual-path model.

[0209] The test set fusion results are shown in Figure 7 Compared with the NWP-XGBoost model, the average RMSE and MAE of the fusion model on the two data sets are significantly reduced by 50.78% and 51.38%, respectively, and the R 2 is improved by 4.33 percentage points; compared with the xLSTM model, the RMSE and MAE are reduced by 29.33% and 24.49%, respectively, and the R 2 is improved by 1.54 percentage points, which indicates that the fusion mechanism is significantly effective. Figure 7 The contribution degree analysis of the b part and the d part of shows that the xLSTM plays a leading role in wind power prediction, and the global contribution degrees of the data sets 1 and 2 are 58.64% and 73.79%, respectively, while the global contribution degrees of the NWP-XGBoost are 41.36% and 26.21%, respectively. When the wind power undergoes a significant mutation or the power generation is in a small amount interval, the contribution degree of the NWP-XGBoost increases significantly; when the power is stable, the contribution weight of the xLSTM remains at a high level, which verifies the advantage of capturing long-term time series dependence. The analysis shows that the xLSTM mainly performs time series prediction based on historical power data, while the NWP-XGBoost can integrate real-time weather forecast information for dynamic response, so the fusion model achieves significant improvement in prediction accuracy and efficiency. The experimental results verify that the adaptive fusion mechanism designed in the present application can effectively realize the collaborative optimization of prediction accuracy and robustness by dynamically adjusting the model output weight.

[0210] Table 1 evaluation indexes

[0211]

[0212] Step f2: design an ablation experiment to verify the role of the core module of the proposed model.

[0213] To further verify the effectiveness of the fusion method designed in the present application, four groups of comparative experiments are constructed by the control variable method: XGBoost, NWP-XGBoost, xLSTM and NWP-xLSTM, and the experimental results are shown in Figure 8 , Figure 9 The quantitative evaluation indexes are shown in Table 2.

[0214] Table 2 evaluation indexes of ablation experiment

[0215]

[0216]

[0217] The experimental results show that the sensitivities of different models to meteorological factors are significantly different. In the XGBoost series model, the introduction of NWP data significantly improves the performance: the average MAE decreases by 10.5%, the average RMSE decreases by 10.3%, and the R 2 score increases by 1.7 percentage points. This verifies that the tree model can effectively capture the nonlinear relationship between meteorological factors and target variables through the feature splitting mechanism. In contrast, the performance of the xLSTM model degrades after the fusion of NWP data: the average MAE increases by 24.1%, and the average RMSE increases by 19.1%. The proposed fusion model reduces the MAE by 56.2% and 56.9% compared with NWP-XGBoost and the basic xLSTM, respectively, and reduces the RMSE by 57.0% and 60.1%, respectively. These results show that directly fusing multi-source data can cause unpredictable fluctuations in model performance, and the dual-path fusion framework designed by the present application provides a reliable solution to this problem.

[0218] Step f3: Design a comparative experiment to verify the superiority of the proposed model.

[0219] To verify the effectiveness and advancement of the multi-modal fusion model proposed by the present application, a benchmark model comparison system across algorithm categories is constructed. Specifically, LightGBM, LSTM, GRU, KELM and Transformer are selected as the control group, and the prediction results of each model on two data sets are shown in Figure 10 and Figure 11 , and the core evaluation indicators are shown in Table 3.

[0220] Table 3 Comparison indicators of different benchmark models

[0221]

[0222] From the above Figure 10 , Figure 11 , Table 3, the multi-modal fusion model shows significant advantages on both data sets. In data set 1, compared with the KELM model with the best performance, the absolute error of the fusion model is reduced by 50.4%. Although KELM approaches the fusion model in the R 2 index, its MAE value is still 2.02 times that of the fusion model, which reveals the inherent defect of kernel methods in local error suppression. When dealing with long-range temporal dependence in data set 2, the MAE of the Transformer model is 206.5% higher than that of the fusion model, indicating that simply relying on the self-attention mechanism cannot effectively capture multi-scale temporal patterns.

[0223] The multi-modal fusion model proposed by the present application can effectively capture the nonlinear relationship between meteorological factors and target variables through the dynamic weight allocation mechanism, while maintaining the R2 At the same time, the MAE and RMSE are controlled below 1.0 and 1.2, respectively, achieving a breakthrough in both prediction accuracy and stability. This improvement verifies the effectiveness of the cross-modal feature interaction module and the adaptive fusion gating unit, providing a new technical path for complex time series prediction tasks.

Claims

1. A dual-path ultra-short-term wind power forecasting method based on the fusion of numerical weather forecast and multi-order time series dynamic gating is characterized by The following steps are involved: Step 1: Obtain historical wind power generation data and numerical meteorological forecast data of the wind farm, and use the maximum information coefficient (MIC) to screen out weather factors that are highly correlated with wind power; Step 2: Adopt the adaptive noise complete ensemble empirical mode decomposition (CEEMDAN) to decompose the power sequence into multiple intrinsic mode functions (IMFs); Step 3: Build a dual-path prediction architecture, where one path uses the extended long short-term memory network xLSTM to predict the intrinsic mode function IMF and superimpose all subsequences to obtain the prediction result; The other path uses XGBoost to combine the intrinsic mode function (IMF) with the key meteorological characteristics of numerical weather forecast (NWP) for prediction, and superimposes all subsequences to obtain the prediction results. Step 4: Design the MT-DGFusion module through enhanced attention and dynamic gating network; Step 5: The dual-path prediction results are fused through the MT-DGFusion module to obtain the final prediction result.

2. The dual-path ultra-short-term wind power forecasting method based on the fusion of numerical weather forecast and multi-order time series dynamic gating according to claim 1 is characterized by: Also includes step 6: using the mean absolute error MAE, mean square error MSE and determination coefficient R 2 These three evaluation indicators are used to compare and analyze the prediction results; The mean absolute error (MAE), root mean square error (RMSE) and coefficient of determination (R 2 ) as a multidimensional evaluation index, the specific mathematical expressions are: In the above formula, y i represents the true value, Represents the predicted value, N is the number of samples, and y represents the mean of the true value.

3. The dual-path ultra-short-term wind power forecasting method based on the fusion of numerical weather forecast and multi-order time series dynamic gating according to claim 1 is characterized by: In step 1, the calculation formula of the maximum information coefficient MIC is expressed as: In formula (1), MIC(X, Y) represents the maximum information coefficient of variables X and Y, X and Y are two variables, a×b represents the grid division method, that is, X is divided into a intervals, Y is divided into b intervals, and the value of a×b is less than a given upper limit B to avoid overfitting; I(X; Y| G ) represents the mutual information between X and Y under a given grid division G; log2min(a,b) is a complexity measure of the grid division, which is used to normalize the mutual information; the weather factor that is highly correlated with wind power is wind speed.

4. The dual-path ultra-short-term wind power forecasting method based on the fusion of numerical weather forecast and multi-order time series dynamic gating according to claim 1 is characterized by: In step 2, CEEMDAN is used as a power decomposition method, which includes the following steps: S2.1: Let the original sequence be x(t), the noise standard deviation coefficient be ε0, the number of collections be M, the initial residual r0(t) = x(t), and the decomposition order k = 1; S2.2: Generate M noise signals: x (i) (t)=r0(t)+ε0·E1(ω (i) (t)),i=1,2,...,M(2); In formula (2): x (i) (t) represents the i-th noise signal, ω (i) (t) is the i-th white noise, E1(·) represents the first IMF obtained by EMD decomposition of the noise; M represents the number of noise-added signals; S2.3: For each x (i) (t) Perform EMD decomposition and extract its first Calculate the average: In formula (3): IMF1(t) represents the first subsequence obtained by CEEMDAN decomposition; Indicates x (i) (t) The i-th subsequence obtained by EMD decomposition; Update residuals: r1(t)=r0(t)-IMF1(t)(4); In formula (4): r1(t) represents the updated residual; S2.4: For k=2, generate M noisy residuals: In formula (5): represents the k-1th order residual generated after the i-th noise addition, r k-1 (t) represents the residual of the original signal after the first k-1 order decomposition, ε k-1 Represents the adaptive noise amplitude coefficient, E k (·) is the kth-order IMF obtained by performing EMD decomposition on the noise; ω (i) (t) represents the Gaussian white noise added for the i-th time, ε k =ε0·σ(r k-1 ), σ is the standard deviation; S2.5: For each noise residual Perform EMD decomposition and extract the first Calculate the average: In formula (6): IMF k (t) represents the ensemble average result of the k-th order intrinsic mode function; Update residuals: r k (t)=r k-1 (t)-IMF k (t)(7); In formula (7): r k-1 (t) represents the residual signal after the first k-1 order decomposition; When the residual r k Stop decomposition when (t) is a monotonic function or has less than 2 extreme points; the final signal decomposition is: In formula (8): X(t) represents the decomposition signal; r K (t) represents the final residual.

5. The dual-path ultra-short-term wind power forecasting method based on the fusion of numerical weather forecast and multi-order time series dynamic gating according to claim 1 is characterized by: In step 3, the extended long short-term memory network xLSTM is composed of a stack of long short-term memory network sLSTM and matrix long short-term memory network mLSTM modules; sLSTM reconstructs the gating mechanism of traditional LSTM and introduces exponential function to replace Sigmoid activation function, including formula: t ,f t ,o t =exp(W·[h t-1 ,x t ]+b), i t ,f t ,o t They are input gate, forget gate, output gate, exp() is exponential function, W is weight matrix, h t-1 is the hidden state of the previous moment, x t is the input vector at the current moment, and b is the bias term; mLSTM expands the scalar memory unit into a matrix structure and combines it with a parameterized memory association matrix to enable full parallelization of hidden state updates in the sequence dimension, including the formula: H t =O t ⊙tanh(M t A · ), M t is the memory matrix at the current moment, M t-1 is the memory matrix of the previous moment, F t , I t , O t They are respectively the forget gate matrix, the input gate matrix, and the output gate matrix. is the candidate memory matrix, A is the association matrix, ⊙ is the element-by-element multiplication, tanh(·) is the hyperbolic tangent activation function, H t is the hidden state matrix at the current moment; In addition, xLSTM supports dynamic adjustment of the ratio of the two types of modules according to task requirements. When the task requires strong sequential dependencies, the proportion of sLSTM is increased, while when high parallelism is required, the proportion of mLSTM is increased. Using the SMS architecture, the ratio of sLSTM to mLSTM is 2:

1.

6. The dual-path ultra-short-term wind power forecasting method based on the fusion of numerical weather forecast and multi-order time series dynamic gating according to claim 5 is characterized by: In step 3, the principle process of the sLSTM module is as follows: sLSTM introduces a normalized state mechanism to dynamically balance the weight relationship between the input gate and the forget gate during the calculation process. The specific formula is: is the normalized input gate, is the normalized forget gate, Ct is the memory unit at the current moment, C t-1 is the memory unit of the previous moment, is the candidate memory, ⊙ is the element-by-element multiplication; sLSTM supports the parallel operation of multiple independent memory units and realizes cross-unit memory mixing through recursive connections; each memory unit It has its own input gate, forget gate, and output gate. It updates its own memory independently according to traditional LSTM rules and calculates the interaction weights between all memory units: Where: E is a learnable weight matrix, Softmax ensures that the sum of the weights of each row is 1, A ij represents the proportion of memory contribution of unit j to unit i; each unit receives the weighted sum of the memory of other units: Among them, the diagonal term A kk Retain the unit's own memory, the off-diagonal item A kj Integrate memories from other units.

7. The dual-path ultra-short-term wind power forecasting method based on the fusion of numerical weather forecast and multi-order time series dynamic gating according to claim 6 is characterized by: The sLSTM forward pass process is described by the following formula system: In formula (9): z t is the candidate unit state, represents the transpose of the input weight matrix, x t The input vector at the current moment, r z represents the recursive weight matrix, h t-1 represents the hidden state of the previous moment, b z represents the bias term; In formula (10): i t is the input gate, represents the transpose of the input weight matrix, r i represents the recursive weight coefficient, h t-1 represents the hidden state of the previous moment, b i represents the bias term, exp(·) represents the exponential function; In formula (11): f t For the Gate of Forgetfulness, represents the transpose of the input weight matrix, r f represents the recursive weight coefficient, b f represents the bias term; In formula (12): t is the output gate, represents the transpose of the input weight matrix, r o represents the recursive weight coefficient, b o represents the bias term; m t =max(log(f t )+m t-1 ,log(i t ))(13); In formula (13): m t Is applied to i t and f t stabilizer to prevent gradient explosion from generating i′ t and f′ t ;log(f t ) represents the logarithm of the forget gate, m t-1 represents the stabilizer at the previous moment, log(i t ) represents the logarithm of input gates; i t ′=exp(log(i t )-m t )(14); In formula (14): i t ' represents the normalized input gate; f t ′=exp(log(f t )+m t-1 -m t )(15); In formula (15): f t ' represents the normalized forget gate; c t =f t ′·c t-1 +i t ′·z t (16); In formula (16): c t is the cell state, c t-1 represents the unit state at the previous moment, z t Indicates the candidate unit status; n t =f t ′·n t-1 +i t ′(17); In formula (17): n t is the normalized state, n t-1 represents the normalized state at the previous moment; h t =o t (c t / n t )(18); In formula (18): h t It is a hidden state.

8. The dual-path ultra-short-term wind power forecasting method based on the fusion of numerical weather forecast and multi-order time series dynamic gating according to claim 7 is characterized by: In step 3, the mLSTM module principle process is as follows: mLSTM significantly expands the storage capacity of the model in the parameter space dimension by expanding the scalar memory unit into a high-dimensional matrix memory structure. At the same time, it improves the efficiency of information association and extraction through the matrix key-value pair association mechanism. The specific formula is: C t is the matrix memory state at the current moment, v t is a value vector, k t is the key vector; In the process of updating the memory matrix, mLSTM establishes a dynamic association between input features and hidden states through covariance operation, and realizes accurate retrieval based on the similarity calculation of the attention mechanism. The formula is: Output = C t ·q t ;q t is the query vector, C t is the memory matrix; mLSTM enables global parallel calculation of hidden state updates at all time steps through matrix multiplication operations, and the formula is: Where cumsum represents the cumulative sum; its forward propagation process can be described by the following formula: q t =W q x t +b q (19); In formula (19): q t is the query input, W q q t The corresponding weight matrix, b q represents the bias term; In formula (20): k t For key input, W k k t The corresponding weight matrix, d represents the dimension of the key vector, b k represents the bias term; v t =W v x t +b v (21); In formula (21): v t Enter the value, W v v t The corresponding weight matrix, x t represents the input vector at the current moment, b v represents the bias term; In formula (22): i t represents the activation value of the input gate, represents the transpose of the weight vector of the input gate, b i represents the bias term; Indicates taking the exponential of the linear transformation result; In formula (23): f t represents the activation value of the forget gate, represents the transpose of the weight vector of the forget gate, b f represents the bias term; the t =σ(w o x t +b o (24); In formula (24): t represents the activation value of the output gate, w o represents the weight vector of the output gate, b o represents the bias term; m t =max(log(f t )+m t-1 ,log(i t ))(25); In formula (25): m t Indicates that it applies to i t and f t The stabilizer, log(f t ) represents the logarithm of the activation value of the forget gate, m t-1 Represents the logarithmic memory state of the previous moment, log(i t ) represents the logarithm of the activation value of the input gate; i t ′=exp(log(i t )-m t )=exp(i t -m t )(26); In formula (26): i t ′ represents the normalized input gate; f t ′=exp(log(f t )+m t-1 -m t )(27); In formula (27): f t ′ represents the normalized forget gate; In formula (28): C t Represents the memory matrix at the current moment, f t ′ represents the normalized forget gate, C t-1 Represents the memory matrix of the previous moment, i t ′ represents the normalized input gate, v t represents a value vector, represents the transpose of the key vector; n t =f t ′·n t-1 +i t ′·k t (29); In formula (29): n t Represents the normalized state at the current moment, n t-1 represents the normalized state of the previous moment, k t The key vector representing the current input; In formula (30): h t is the hidden state, o t represents the output gate, C t Represents the memory matrix at the current moment, q t represents the query vector, represents the normalized state at the current moment, Ensure numerical stability and avoid exploding gradients.

9. The dual-path ultra-short-term wind power forecasting method based on the fusion of numerical weather forecast and multi-order time series dynamic gating according to claim 8 is characterized by: In step 3, the prediction output of XGBoost is integrated by K decision trees, and the prediction output of sample x is i The predicted value is: In formula (31): is the model prediction output, f k is the function of the kth tree, x i is the i-th sample; The objective function of XGBoost consists of a loss function and a regularization term: In the above formula, represents the overall objective function, is the mean square error between the predicted value and the actual value; i represents the true value of the i-th sample, n represents the total number of samples, Ω(f k ) represents the regularization term of the k-th tree, T k is the number of leaf nodes in the kth tree; w k is the leaf node weight; γ and λ are the leaf number penalty coefficient and L2 regularization coefficient respectively; At the tth iteration, Xgboost has fixed the first t-1 trees and needs to optimize the tth tree f t ; The objective function is simplified to: In formula (34): represents the overall objective function at the tth iteration, represents the cumulative prediction value of the first t-1 trees for sample i, f t (x i ) represents the predicted output of the t-th tree for sample i; represents the loss function value of sample i in round t, Ω(f t ) represents the regularization term of the t-th tree; Perform a second-order Taylor expansion on the loss function to For expansion point: In formula (35): represents the prediction error of the first t-1 trees on sample i; g i represents the first-order gradient; h i represents the second-order gradient; are the first-order and second-order gradients of the loss function, respectively; Group the samples by leaf nodes, let I i ={i|x i ∈j};I i represents the set of all sample indexes that are divided into leaf j; i represents the sample index; j represents the leaf node number; Then the objective function can be rewritten as: In formula (36): T is the number of leaf nodes in the current tree; w j represents the output weight of leaf j; h i represents the second-order gradient; γ represents the leaf splitting penalty coefficient; to w j Taking the derivative and setting it to zero gives the optimal weight: In formula (37): represents the optimal weight of leaf j, I j represents the set of samples divided into leaf j; Will Substituting into the objective function, we get the minimum loss: In formula (38): represents the minimum loss value of the tth round, and T represents the total number of leaf nodes.

10. The dual-path ultra-short-term wind power forecasting method based on the fusion of numerical weather forecast and multi-order time series dynamic gating according to claim 9 is characterized by: In step 4, the MT-DGFusion module constructs a multi-order feature representation with temporal correlation; given the output of two prediction models at time t, and And the corresponding actual observation value y t , by introducing the lag operator L k Construct time series dynamic features; specifically, construct historical features with two lag steps for each prediction model; In formula (39): represents the historical prediction of model a one step ahead, represents the historical prediction of model a 2 steps ago, represents the historical prediction of model b one step ahead, represents the historical prediction of model b 2 steps ahead; represents the prediction of model a at time t-1, represents the prediction of model a at time t-2, represents the prediction of model b at time t-1, represents the prediction of model b at time t-2, represents the predicted output of model a at time t-1, represents the predicted output of model a at time t-2, represents the predicted output of model b at time t-1, Represents the predicted output of model b at time t-2; The input feature matrix constructed in this way contains the current moment prediction value and its historical information: In formula (40): X t represents the original feature vector, represents the prediction of model a at time t, represents the prediction of model b at time t, represents the t-1 time prediction of model a, represents the t-2 time prediction of model a, represents the prediction of model b at time t-1, represents the prediction of model b at time t-2; Feature Matrix Extract high-order time series features through multi-layer nonlinear transformation: H (2) =LayerNorm(W2H (1) +b2)(42); H=Dropout(H (2) )(43); In the above formula, H (1) represents the first layer hidden state, H (2) Represents the second layer hidden state, H represents the final output feature, W1 and W2 are learnable weight matrices, LeakyReLU is an activation function with a negative slope, LayerNorm is a layer normalization operation, and Dropout is used for regularization; Represents the input feature matrix, b1 represents the first layer bias term, and b2 represents the second layer bias term; the feature is then decomposed into a multi-head attention mechanism: MultiHead(H)=Concat(head1,...,head h )W O (44); head i =Attention(HW i Q ,HW i K ,HW i V )(45); In the above formula, MultiHead(H) represents the final output after the input feature H is processed by the multi-head attention mechanism; head1,...,head h Represent the output results of the 1st to hth attention heads respectively; Concat(head1,...,head h ) means concatenating the outputs of h attention heads in the feature dimension; head i Indicates the calculation result of the i-th attention head; Attention(HW i Q ,HW i K ,HW i V ) represents the calculation process of the i-th attention head; W i Q , W i K , W i V is the projection matrix of each head, W O is the output projection matrix; The calculation of each attention head uses scaled dot product attention: In formula (46), Attention(Q,K,V) represents scaled dot product attention, Q represents query matrix, K represents key matrix, V represents value matrix, K represents T represents the key matrix transpose, d k is the dimension scaling factor of the key vector; Dynamically gate the multi-head attention output A with the original feature H: G=σ(W g [H;A]+b g (47); F=G⊙H+(1-G)⊙A(48); Among them, G is the gate weight, W g is the gate weight matrix, σ is the Sigmoid function, ⊙ represents element-by-element multiplication; b g represents the gate bias term, W g [H; A] represents the linear transformation of features, and F represents the fused features; At the same time, an auxiliary prediction path is constructed to retain the original feature information: In formula (49): R represents the auxiliary feature representation, W r represents the projection weight matrix, represents the original input features, b r Indicates auxiliary path bias; In formula (50): represents the auxiliary prediction output, W r2 Represents the output weight matrix, ReLU(R) represents the activated feature, R represents the auxiliary feature representation, b r2 Indicates output bias; The final prediction is achieved by superimposing the main path and the auxiliary path: in: represents the final prediction output, MLP(F) represents the main path prediction, and F represents the fused feature.

Images