A wind power prediction method combining MCG-STE, Time-TK and FNO joint optimization

Abstract

The application discloses a kind of wind power prediction methods of combining MCG-STE, Time-TK and FNO joint optimization, comprising the following steps: step S1: modal decomposition and intrinsic mode generation based on VMD;Step S2: complexity quantification and entropy weight construction based on RCMFDE;Step S3: space-time feature coding based on STAE;Step S4: obtain standardized input matrix, and carry out multiple offset time embedding and offset subsequence construction, and input multiple offset subsequence generated into MI-KAN module, and carry out nonlinear feature mapping to each offset branch;Step S5: multi-head self-attention operation is carried out to each offset branch respectively, and the correlation between different time positions under the same offset perspective is learned;Step S6: MOTE based on FNO and MI-KAN key parameter joint optimization;Step S7: the optimal parameter of output is solved back to MOTE module and MI-KAN module, and prediction result is generated.The application integrates signal smoothing processing, time correlation modeling, nonlinear representation enhancement, global interaction fusion and key parameter joint optimization into one, effectively improves prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction, and in particular to a wind power prediction method that combines MCG-STE, Time-TK and FNO for joint optimization. Background Technology

[0002] Against the backdrop of high-proportion grid integration of new energy sources and the continuous development of new power systems, the accuracy of wind power forecasting directly affects grid dispatching plans, reserve capacity configuration, power balance control, and the formulation of new energy consumption strategies. Especially in scenarios involving wind farm cluster access, coordinated operation of power generation, grid, load, and storage, and cloud-edge collaborative forecasting, wind power forecasting not only impacts the optimization of individual wind farm operations but also relates to regional power balance, safety margin assessment, and dispatch response efficiency. Because wind power is influenced by multiple factors such as wind speed, wind direction, air density, terrain disturbances, meteorological changes, wake effects, and unit operating status, its time series exhibits significant non-stationarity, random fluctuations, strong noise interference, and multi-timescale coupling characteristics, making it a critical and technically challenging problem in the power system field.

[0003] Current wind power forecasting technologies still face several technical bottlenecks. Multiple components in non-stationary signals are difficult to effectively isolate, and continuous-time structure and cross-period correlation information cannot be fully preserved. In extracting complex nonlinear features, deep fusion of local and global information is often difficult, and key structural parameters lack a joint optimization mechanism. These shortcomings, combined, severely limit the prediction accuracy of existing models. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a wind power prediction method that combines MCG-STE, Time-TK and FNO joint optimization. This method integrates signal stabilization processing, time correlation modeling, nonlinear characterization enhancement, global interactive fusion and joint optimization of key parameters, effectively improving prediction accuracy.

[0005] The objective of this invention is achieved through the following technical solution: a wind power prediction method combining MCG-STE, Time-TK, and FNO joint optimization, comprising the following steps:

[0006] Step S1: Acquire the original multi-source time-series data of the wind farm, and perform unified resampling and time alignment on the original multi-source time-series data from different sources and with different sampling frequencies; extract the historical power sequence corresponding to the wind power channel from the original multi-source time-series data after unified resampling and time alignment, and use the VMD algorithm to adaptively decompose the historical wind power sequence into several intrinsic mode functions with finite bandwidth, and then use the VMD decomposition to obtain the first... The eigenmode functions are denoted as ,in, , The number of intrinsic mode functions obtained from VMD decomposition is denoted as ; and other original multi-source temporal features that did not enter the VMD mode decomposition branch are retained for joint input feature construction in step S4.

[0007] Step S2: Obtain the eigenmode functions output in step S1. Subsequently, a complexity quantification calculation is performed on the discrete mode sequences corresponding to each intrinsic mode function to construct an entropy weight vector that reflects the prediction difficulty of each intrinsic mode function. ;

[0008] Step S3: Generate the discrete mode sequence corresponding to each eigenmode function and the corresponding entropy weight vector. The input is fed into a spatiotemporal autoencoder network for feature encoding;

[0009] Step S4: Perform multi-source heterogeneous data fusion and reconstruction to obtain a standardized input matrix, and perform multi-offset time embedding and offset sub-sequence construction. Input the generated multiple offset sub-sequences into the MI-KAN module to perform nonlinear feature mapping on each offset branch.

[0010] Step S5: After obtaining the high-order representation of each offset subsequence, perform multi-head self-attention operation on each offset branch to learn the correlation between different time positions under the same offset perspective;

[0011] Step S6: Joint optimization of key parameters of MOTE and MI-KAN based on FNO;

[0012] Step S7: Solve for the optimal parameters output. The process is re-injected into the MOTE and MI-KAN modules to re-execute the multi-offset subsequence construction, nonlinear higher-order representation extraction, MOTI modeling, and global fusion, resulting in a unified fused feature representation with optimal parameter configuration. .

[0013] The beneficial effects of this invention are: (1) The integrated prediction architecture improves the overall accuracy and robustness of wind power prediction:

[0014] In existing wind power prediction methods, signal enhancement, temporal embedding, feature representation, interactive modeling, and parameter optimization are mostly designed in a decentralized manner, with insufficient connection between modules, making it difficult to form a unified closed-loop prediction link. This results in difficulties in co-modeling local features and global temporal relationships. This application constructs an integrated prediction architecture composed of MCG-STE, MOTE, MI-KAN, MOTI, global fusion, and FNO parameter joint optimization. This architecture enables multi-source data processing, signal enhancement, multi-offset interaction, nonlinear representation, and parameter optimization to be executed collaboratively within the same process, thereby enhancing the overall ability to characterize short-term fluctuations, long-term trends, and complex operating conditions of wind power, and improving prediction accuracy and robustness.

[0015] (2) The MCG-STE preprocessing enhancement mechanism improves the input representation quality and prediction stability:

[0016] Existing methods typically model the raw wind power sequence directly, which struggles to effectively separate non-stationary components, identify modal complexity differences, and implement differentiated enhancements. This results in insufficient modelability of the input signal and volatile prediction results under complex operating conditions. This application, based on the MCG-STE module, constructs a unified preprocessing enhancement stage that integrates VMD modal decomposition, RCMFDE complexity quantification, and STAE spatiotemporal coding. This stage decomposes, quantifies, and enhances the raw wind power sequence and injects complexity information into the spatiotemporal feature extraction process, improving the quality of the input representation from the source and enhancing prediction stability under complex operating conditions.

[0017] (3) The Time-TK module improves wind power prediction performance:

[0018] The Time-TK module consists of Multi-Offset Temporal Embedding (MOTE), Nonlinear Mapping Representation (MI-KAN), and Multi-Offset Temporal Interaction (MOTI). MOTE divides the input sequence along the time dimension into multiple offset sub-sequences reflecting different time spans, enhancing the preservation of the continuous temporal structure and cross-period correlation information of the original sequence. MI-KAN extracts high-order dynamic features from each offset sub-sequence through independent nonlinear mapping, improving the model's representation of complex nonlinear coupling relationships, strong fluctuations, and abrupt changes. MOTI performs multi-head attention interaction and aggregation on the features of each offset branch and maps them to a unified hidden space, achieving the synergistic fusion of local dynamic information and global trend information. Through the synergistic effect of these modules, the model's comprehensive modeling ability for short-term fluctuations, periodic evolution, and long-term trends in wind power can be effectively improved, thereby increasing the accuracy of wind power prediction.

[0019] (4) The FNO joint optimization mechanism improves parameter matching ability and model adaptation ability:

[0020] Existing prediction models often rely on empirically set temporal embedding parameters and nonlinear representation parameters, making it difficult to achieve coordinated matching for different wind power scenarios and easily leading to model performance fluctuations. This application introduces FNO (Flexible Noise Reduction) into the model to jointly optimize MOTE and MI-KAN parameters, and then injects the optimal parameter solution back into the subsequent modeling process, forming a closed-loop optimization mechanism of "joint search - fitness evaluation - parameter backinjection". This design improves the matching ability of key structural parameters and the model's adaptive ability, thereby enhancing the matching between model structural parameters and input features. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the principle of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0023] This invention integrates a Mode-Complexity-Guided Spatiotemporal Enhancement (MCG-STE) module, which is composed of Variational Mode Decomposition (VMD), Refined Composite Multiscale Fractional Dispersion Entropy (RCMFDE), and Spatiotemporal Attention Autoencoder (STAE). It further combines a Multi-Offset Temporal Interaction Framework (Combining Transformer and Kolmogorov-Arnold Networks for Time Series Forecasting, Time-TK) with a Farthest Better or Nearest Worse Optimizer (FNO) algorithm to achieve wind power prediction. The method is deployed in a wind farm power prediction master station, a wind farm edge computing gateway, a new energy centralized control platform or a cloud-edge collaborative prediction server, and the execution entities are industrial control servers, prediction computing nodes, wind farm edge intelligent terminals or cloud computing platforms.

[0024] During operation, the edge acquisition terminal receives heterogeneous time-series data related to the wind farm in real time through the SCADA system, power prediction substation, meteorological monitoring interface, numerical weather prediction interface, and station operation monitoring interface. The central prediction server performs unified resampling and time alignment on the data, extracts historical wind power sequences, and sequentially performs power sequence decomposition, complexity quantification, and feature enhancement. The enhanced features are then fused, normalized, and time-series reconstructed with other original multi-source features to finally complete model inference and output wind power sequences for multiple future prediction times.

[0025] The algorithm structure effectively extracts complex dynamic laws in long-term, multi-variable, and highly fluctuating wind power scenarios by performing a stationarization decomposition of the wind power signal, explicitly quantizing different modal complexities, and combining multi-offset subsequence modeling, nonlinear function mapping, cross-offset time series interaction, and joint optimization of key parameters.

[0026] like Figure 1 As shown, a wind power prediction method combining MCG-STE, Time-TK, and FNO optimization includes the following steps:

[0027] Step S1: Acquire the original multi-source time series data and complete unified resampling and time alignment, extract the historical wind power sequence from it and perform VMD mode decomposition;

[0028] Raw multi-source time-series data is received from the SCADA system, meteorological monitoring interface, numerical weather prediction interface, and station operation monitoring interface. This raw multi-source time-series data includes historical wind power, wind speed, wind direction, air density, ambient temperature, turbine operating status, and numerical weather prediction features. To address differences in sampling frequency and timestamps between different data sources, resampling is performed according to a preset unified sampling period, and time alignment is completed based on the timestamps. Historical wind power sequences are extracted from the raw multi-source time-series data after unified resampling and time alignment as VMD input. The remaining wind speed, wind direction, air density, ambient temperature, turbine operating status, and numerical weather prediction features are not involved in VMD decomposition but are retained as raw multi-source time-series features in step S4 for constructing a joint input feature vector.

[0029] When solving for each modal component, iterative optimization is performed using the alternating direction multiplier method. The core is the first... The modal update formula and the center frequency update formula are as follows:

[0030]

[0031]

[0032] in, For the first The modal component in the th ... Frequency domain expression under the next iteration;

[0033] This is the center frequency corresponding to this mode;

[0034] This is the Fourier transform of the original wind power signal;

[0035] Indicates the current mode The sum of all other modal components;

[0036] For the frequency domain form of Lagrange multipliers;

[0037] It is a secondary penalty factor.

[0038] The above alternating iteration is repeated for the frequency domain expression of each intrinsic mode function and its center frequency, and the change between two adjacent iterations is used to determine whether the preset convergence condition is met. When the convergence condition is met, the iteration is stopped, and the converged... Frequency domain modes Performing the inverse Fourier transform yields the... Individual eigenmode functions After the above processing, the original wind power sequence is decomposed into several eigenmode functions that tend to be stationary. Together, they constitute the VMD decomposition result of step S1 and serve as the input for complex quantification in step S2. VMD effectively suppresses the frequency aliasing phenomenon in the original non-stationary power signal, and removes the strong randomness and non-stationary features within the wind power sequence, thus providing a high signal-to-noise ratio stationary data foundation for subsequent feature encoding and attention allocation.

[0039] Step S2: Obtain the eigenmode functions output in step S1. Each intrinsic mode function is discretized into a discrete mode sequence according to a uniform sampling interval. Furthermore, complex metric calculations are performed on each discrete mode sequence to construct an entropy weight vector that reflects the prediction difficulty of each intrinsic mode function. A multi-scale coarse-grained strategy is employed to process each discrete mode sequence, and the RCMFDE is calculated. The specific process is as follows:

[0040] (1) The first step obtained in step S1 Individual eigenmode functions Discretize according to a uniform sampling interval to obtain a length of Discrete mode sequence and in scale factor The discrete mode sequence is then subjected to composite multi-scale coarsening processing to obtain coarse-grained time subsequences. Its composite multi-scale coarsening formula is:

[0041]

[0042] (2) Calculate the time series after each transformation fine state probability :

[0043]

[0044] (3) Obtain RCMFDE under different scaling factors based on fractional calculus:

[0045]

[0046] In the formula; This is the coarsened time subsequence;

[0047] The scaling factor is a positive integer subject to the following constraint: ;

[0048] Let be the interval probability of cosine similarity;

[0049] For multi-scale time series at the same scale factor The relative frequency below;

[0050] Represents the interval number;

[0051] The fractional order limit is represented as ;

[0052] Indicates the embedding dimension;

[0053] These are the elemental characteristic parameters of the state probability;

[0054] Represents the gamma function;

[0055] This represents the double gamma function.

[0056] For each intrinsic mode function, the divergence entropy value is calculated for the discrete mode sequence, normalized, and then concatenated to form an entropy weight vector. The entropy weight vector is then input into the STAE module.

[0057] Step S3: Generate the discrete mode sequence corresponding to each eigenmode function. and the corresponding entropy weight vector The input is encoded using a spatiotemporal autoencoder network;

[0058] Discrete mode sequences corresponding to each eigenmode function and the corresponding entropy weight vector After acquisition, the features are input into a spatiotemporal autoencoder network for feature encoding. This process utilizes a cascaded structure of a one-dimensional convolutional network (1D-CNN) and a bidirectional long short-term memory network to sequentially extract spatial local features and bidirectional temporal dependency features. The entropy weight vector is then obtained. The corresponding number in the middle scalar weight elements of a discrete modal sequence This is replicated and extended in the time dimension to construct a sequence similar to the discrete mode sequence. One-dimensional weight vectors of the same length are used, and the one-dimensional weight vectors are matched with the corresponding discrete mode sequences. Perform element-wise multiplication and fusion to generate weighted feature input terms. and The weighted feature input terms are then input into the network and adjusted through a dual-path system of high-entropy enhancement and low-entropy protection. The formulas for the encoder's spatial feature extraction and bidirectional temporal feature extraction are as follows:

[0059]

[0060]

[0061]

[0062]

[0063] In the formula, and For the first A discrete mode sequence With corresponding scalar weight elements The weighted feature input term is formed after element-wise multiplication and fusion;

[0064] This is a dot product operation;

[0065] For a one-dimensional convolutional network The output features of each convolutional kernel;

[0066] and These are the kernel weights and bias terms, respectively.

[0067] For activation functions;

[0068] The hidden state output of the feedforward long short-term memory network;

[0069] This is the hidden state output of the backward Long Short-Term Memory network;

[0070] These represent the input gate, output gate, forget gate, and cell state within the LSTM network, respectively.

[0071] for Network bias;

[0072] The final spatiotemporal feature output representation obtained after bidirectional stitching (i.e., the reconstructed time) data);

[0073] For all discrete mode sequences corresponding to intrinsic mode functions, the above encoding process is performed respectively to obtain the discrete mode sequence at time t. Bidirectional spatiotemporal feature output representation The outputs of each discrete mode sequence are spliced ​​and fused along the feature dimension to form a time-sequence sequence. Unified Enhanced Feature Representation and the This serves as an enhanced source of information for constructing the joint input feature vector in subsequent S4.

[0074] Step S4: Perform multi-source heterogeneous data fusion and reconstruction to obtain a standardized input matrix, and perform multi-offset time embedding and offset sub-sequence construction. Input the generated multiple offset sub-sequences into the MI-KAN module to perform nonlinear feature mapping on each offset branch.

[0075] Based on the original multi-source time series data that has undergone unified resampling and time alignment in step S1 and has not yet entered VMD mode decomposition, extract the time... The corresponding original multi-source feature vector ,in, The characteristics of S3 output consist of wind speed, wind direction, air density, ambient temperature, unit operating status, and numerical weather forecast features that do not enter the VMD mode decomposition branch; This is an enhanced representation of the historical wind power channel obtained after VMD mode decomposition, RCMFDE complex quantization, and STAE spatiotemporal coding. and By concatenating and fusing features along their respective dimensions, a joint input feature vector is constructed.

[0076]

[0077] In the formula, This represents a vector concatenation operation performed along the feature dimension; Indicates time The total dimension of the fused joint input feature vector is denoted as . .

[0078] After acquiring data from multiple consecutive time points, according to the historical window length Stack the input sequences along the time dimension to form a multivariate time series matrix:

[0079]

[0080] In the formula, This represents the historical sequence matrix of the model input; Indicates the length of the history window.

[0081] For the input matrix Normalization is performed to obtain the standardized input matrix:

[0082]

[0083] In the formula, This represents the normalized input matrix; This represents the normalization operator; This represents the normalized single-step feature vector.

[0084] To avoid mapping each time step to an independent token in isolation, the normalized input matrix is... Perform multi-offset splitting to generate multiple sets of offset subsequences:

[0085]

[0086] In the formula, Indicates the preset offset number. Indicates the first The offset subsequence. Each offset subsequence is defined as:

[0087]

[0088]

[0089] In the formula, Indicates the offset interval. Indicates the starting index of the offset position. Indicates the first The offset subsequence within the historical window length The maximum number of offset steps achievable under constraints as well as Both represent the standardized input matrix The normalized single-step feature vector at the corresponding time position. For those exceeding the historical window length. Elements that are not included in the corresponding offset subsequence are no longer included. In this way, multiple observation perspectives with different time spans are formed within the same historical window, thus simultaneously representing both short-term fluctuation information and long-term trend information.

[0090] The generated multiple offset sub-sequences are input into the MI-KAN module, and nonlinear feature mapping is performed on each offset branch. The inter-layer mapping relationship can be expressed as:

[0091]

[0092] In the formula, Indicates the first Layer The output of each neuron;

[0093] Indicates the first Layer The output of each neuron;

[0094] Indicates from the first Layer The first neuron to the second Layer A learnable unary mapping function between neurons;

[0095] This represents the summation and aggregation of the outputs of all relevant neurons in the previous layer. When hour, This represents any offset subsequence generated in step S404. The first normalized single-step feature vector corresponding to the time position in the middle Each characteristic component; when hour, This indicates the network layer above the MI-KAN module. Scalar output of a neuron.

[0096] In this embodiment, the unary mapping function adopts the form of a radial basis function, and the input variable of the current unary mapping function is denoted as... When the unary mapping function is located in the first layer of the MI-KAN module, For any offset subsequence generated in step S404 A feature component of the normalized single-step feature vector corresponding to the time position; when the unary mapping function is located in a subsequent network layer of the MI-KAN module. This represents the scalar output of the neuron in the previous network layer. The radial basis function is expressed as:

[0097]

[0098] In the formula, Indicates the output of the radial basis function;

[0099] Indicates input variables The distance between the radial base center;

[0100] The smoothing parameter represents the radial basis function;

[0101] This represents the natural exponential function.

[0102] Based on the radial basis function, the nonlinear mapping output of the current input variable is expressed as:

[0103]

[0104] In the formula, Indicates input variables The nonlinear mapping output obtained after processing by a radial basis function network; This represents the scalar input of the current unary mapping function of the MI-KAN module, when the unary mapping function is located in the first layer. The feature component originating from the corresponding normalized single-step feature vector in the offset subsequence generated in step S404, when the unary mapping function is located in a subsequent network layer... This is the scalar output of the neuron in the previous network layer; Represents radial basis functions; Indicates the first Learnable weights corresponding to radial basis functions; Indicates the first One radial base center; Indicates input variables With the The distance between the centers of the radial bases; Indicates the number of radial base centers.

[0105] After processing by the MI-KAN module, the high-order representations of each offset subsequence are obtained:

[0106]

[0107] In the formula, This represents the high-order feature representation extracted by the MI-KAN module;

[0108] The representation function of the MI-KAN module;

[0109] This represents multiple offset subsequences of the input.

[0110] Step S5: After obtaining the high-order representations of each offset subsequence, perform multi-head self-attention operations on each offset branch to learn the correlation between different time positions from the same offset perspective:

[0111]

[0112] In the formula, Indicates the first Feature representation of each offset branch after self-attention enhancement; Indicates the first A high-order representation of each offset subsequence; This represents a multi-head self-attention operation; the three parameters in the formula correspond to the query, key, and value in the attention mechanism, respectively.

[0113] The interaction results of multiple offset branches are concatenated along the feature dimension and mapped to the desired result using linear projection. The same hidden space yields multi-offset interaction features:

[0114]

[0115] In the formula, This represents a linear projection operator used to map the concatenated attention interaction results of multiple offset branches to a coordinate system. The same hidden dimension is used to meet the dimension alignment requirements in the subsequent global fusion process.

[0116] Based on this, the multi-offset interactive features after linear projection As a query term, the standardized input matrix As key and value terms, global multi-head attention fusion is performed, and combined with residual connections to obtain a unified temporal representation:

[0117]

[0118] In the formula, This represents the unified temporal feature representation after global fusion;

[0119] , , These represent the query item, key item, and value item, respectively.

[0120] This indicates a multi-head self-attention operation.

[0121] Through the above processing, we can simultaneously utilize the contextual information on the original time axis and the local dynamic features after multi-offset segmentation, thereby enhancing the ability to model long-term dependencies, periodic fluctuations, and mutation behaviors.

[0122] Step S6: Joint optimization of key parameters of MOTE and MI-KAN based on FNO;

[0123] FNO is a swarm intelligence optimization algorithm driven by the collaborative factors of candidate solution quality and location distance. It continuously searches for optimal parameter regions within the parameter space, jointly optimizes key parameters in the MOTE and MI-KAN modules, and then injects the optimal parameter solutions back into the prediction model, thereby improving the subsequent modeling performance of MOTI and global fusion.

[0124] (1) Initialization phase

[0125] Initialize the population, and uniformly denote the candidate parameter positions of FNO as... Each candidate parameter position corresponds to a set of parameters to be optimized, including the offset number parameter and offset interval parameter in the construction of the offset subsequence, and the radial basis function smoothing parameter and the number of radial basis centers in MI-KAN. The positions of the candidate parameters are represented as follows:

[0126]

[0127] In the formula, Represents the first in the search space One candidate parameter position;

[0128] Represents the decision variables in the corresponding dimension;

[0129] Indicates the dimension of the parameter space;

[0130] Multiple candidate parameter positions together constitute a population, and the population size is denoted as . .

[0131] Among them, the offset parameter Offset interval parameter and the number of radial base centers The smoothing parameters of the radial basis function are obtained using integer encoding. Continuous real number encoding is employed; integer parameters are first rounded, and then boundary pruning is performed based on preset upper and lower bounds for each parameter. Furthermore, structural feasibility checks are performed on the candidate parameter positions to ensure that the offset subsequence index does not exceed the historical window length. Furthermore, the number of radial basis centers is not greater than the effective sample length of the corresponding offset subsequence; for candidate parameter positions that do not meet the above constraints, correction or discard processing is performed to ensure that the parameter solutions injected back into the MOTE module and MI-KAN module meet the feasibility constraints.

[0132] (2) Determining the farthest optimal location

[0133] For any candidate position in the population, FNO searches for the farthest and highest-quality candidate position based on the distance between the objective function value and the candidate position. This farthest optimal position is expressed as:

[0134]

[0135] In the formula, Indicates candidate position The corresponding farthest optimal position;

[0136] and Representing candidate positions and The objective function value;

[0137] This represents the maximum distance between candidate positions. This stage is used to enhance the global exploration capability during the parameter search process.

[0138] (3) The stage of determining the most recent inferior candidate positions

[0139] Based on determining the farthest optimal position, FNO further filters out the candidate positions closest to the current position from those that satisfy the objective function value worse than the current position, and uses these as the nearest inferior candidate positions. The expression for this is:

[0140]

[0141] In the formula, Indicates candidate position The corresponding nearest poor candidate position;

[0142] This represents the minimum distance between candidate positions.

[0143] This stage is used to further identify candidate parameter locations that are close to the current location and whose objective function values ​​are inferior to the current location within the neighborhood of the current location. This provides a local avoidance reference for the subsequent location update process and together with the farthest optimal location, it forms the basis for the coordinated advancement of global exploration and local search.

[0144] (4) Dynamic Focusing Stage

[0145] To broaden the search scope in the early stages and enhance local development capabilities in the later stages, FNO introduces a dynamic focusing strategy, Depth-First Search (DFS), which uses control parameters that gradually change with each iteration. Adjusting the search area, its expression is:

[0146]

[0147]

[0148] In the formula, Indicates the number of times the function has been evaluated;

[0149] This indicates the maximum number of function evaluations.

[0150] This represents a random number that follows a uniform distribution within the interval [0,1].

[0151] As iterations progress... By gradually reducing the size of the search area, the search region shrinks around the far-optimal position, thereby enhancing the ability to perform fine-grained optimization in the later stages.

[0152] (5) Candidate parameter update stage

[0153] The FNO position update includes two processes: the nearest inferior position jump update and the farthest optimal position update. The temporary position of the candidate position is calculated using the following expression:

[0154]

[0155]

[0156] When the probability condition is met, perform a jump to the nearest inferior position:

[0157]

[0158] Furthermore, FNO performs far-optimal exploration updates using the farthest optimal location and a dynamic focusing strategy, expressed as:

[0159]

[0160]

[0161]

[0162] In the formula, Indicates the switching factor;

[0163] Represents a random number within the interval [0,1].

[0164] Indicates the step size coefficient;

[0165] This indicates the position of the updated candidate parameter.

[0166] Through the above update mechanism, candidate parameters can prioritize skipping inferior candidate positions that are closer to the current position in the neighborhood, and perform further exploration in combination with the farthest optimal position, thereby achieving the coordinated advancement of local search and global search.

[0167] (6) Fitness evaluation and memory update stage

[0168] For the updated new position, boundary clipping is performed based on the upper and lower bounds of the parameters. During the parameter optimization phase, input data and corresponding label data are constructed based on known labeled samples, and the current candidate parameters are solved. The process of re-injecting the MOTE and MI-KAN modules, constructing multiple offset subsequences, extracting nonlinear higher-order representations, modeling MOTI, and global fusion is re-executed to obtain the unified temporal representation corresponding to the candidate parameter solution. Based on this, by invoking the linear mapping layer and the inverse normalization operator, the wind power prediction sequence of the candidate parameter solution in the prediction time domain corresponding to the labeled sample is derived forward. Furthermore, the predicted sequence is compared with the actual wind power sequence in the labeled sample labels. The fitness function value of the current candidate parameter solution is calculated by comparing the results. In this embodiment, the fitness function is defined as the root mean square error, i.e.:

[0169]

[0170] In the formula, Indicates the position of candidate parameters The fitness function value;

[0171] Indicates in the parameter The first one obtained Predict wind power output at any given time;

[0172] A label indicating the actual wind power output at the corresponding moment;

[0173] This indicates the length of the prediction time domain.

[0174] If the fitness function value corresponding to the new position is better than that of the original candidate position, then the new position is accepted; otherwise, the new position is rejected and the original position is retained, while the optimal parameter solution under the current iteration is updated.

[0175] (7) Iteration Termination Phase

[0176] Repeat the above process of determining the farthest optimal position, determining the nearest inferior candidate position, dynamic focusing, updating candidate parameters, and evaluating fitness until the maximum number of function evaluations is reached or the convergence condition is met, and output the optimal parameter solution. The optimal parameter solution The MOTE and MI-KAN modules are used to inject data back into the wind power prediction model for subsequent prediction calls.

[0177] Step S7: Solve for the optimal parameters output. The process is re-injected into the MOTE and MI-KAN modules to re-execute the multi-offset subsequence construction, nonlinear higher-order representation extraction, MOTI modeling, and global fusion, resulting in a unified fused feature representation with optimal parameter configuration. .

[0178] Solve the optimal parameters from the output. The process is re-injected into the MOTE and MI-KAN modules to re-execute the multi-offset subsequence construction, nonlinear higher-order representation extraction, MOTI modeling, and global fusion, resulting in a unified fused feature representation with optimal parameter configuration. By inputting the unified fusion feature representation into the linear mapping layer, the wind power prediction sequence in the future prediction time domain is directly obtained:

[0179]

[0180] In the formula, Represents a linear mapping layer;

[0181] Indicates the prediction step size;

[0182] This indicates that in the optimal parameter solution A unified temporal feature representation obtained through global fusion under constraints;

[0183] Indicates from time At the time The predicted wind power sequence.

[0184] Finally, inverse normalization is performed on the wind power prediction sequence to restore the prediction results to their true physical dimensions:

[0185]

[0186] In the formula, Indicates the inverse normalization operator;

[0187] This represents the final wind power prediction result after restoring it to its true physical dimensions.

[0188] Taking a wind farm as an example, the system operates in 15-minute intervals. Edge terminals continuously collect historical data from the past 24 hours as an input window, with a window length of 96. The input feature dimension is set to 12, including historical wind power, nacelle wind speed, wind direction, ambient temperature, air density, pitch angle, rotational speed, NWP predicted wind speed, NWP predicted wind direction, power curtailment markers, turbine availability status, and bus voltage. After unified resampling and time alignment of the above multi-source data, the historical wind power channel is extracted and VMD mode decomposition is performed, decomposing the original non-stationary power sequence into several intrinsic mode functions (EMFs). Then, the discrete mode sequences corresponding to each EMF are subjected to RCMFDE complexity quantification to construct entropy weight vectors. These discrete mode sequences and their entropy weights are input into the STAE module to obtain enhanced spatiotemporal features adjusted for prior complexity. These enhanced spatiotemporal features are then concatenated and fused with other original multi-source features not included in the VMD mode decomposition branch to construct a joint input sequence, which is then normalized. Based on the optimal offset number of 4 obtained through joint optimization using FNO, four sets of offset sub-sequences are formed. Each offset sub-sequence is input into the MI-KAN module to extract nonlinear high-order dynamic features, and then multi-head attention is used to complete intra-offset interaction and cross-offset global fusion. Furthermore, the optimal parameter solution, which includes the optimal offset number, optimal offset interval, optimal radial basis function smoothing parameter, and optimal number of radial basis centers, is back-injected into the model. The construction of multiple offset sub-sequences, nonlinear high-order feature extraction, MOTI modeling, and global fusion are then re-executed, finally outputting the wind power prediction results for the next 4 hours, with a prediction step size set to 16.

[0189] The foregoing description illustrates and describes a preferred embodiment of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A wind power prediction method combining MCG-STE, Time-TK, and FNO joint optimization, characterized in that: Includes the following steps: Step S1: Obtain the original multi-source time series data of the wind farm, and perform unified resampling and time alignment on the original multi-source time series data from different sources and with different sampling frequencies; Historical power sequences corresponding to wind power channels are extracted from the original multi-source time-series data after unified resampling and time alignment. The VMD algorithm is then used to adaptively decompose these historical wind power sequences into several intrinsic mode functions (EMFs) with finite bandwidth. The VMD decomposition yields the first... The eigenmode functions are denoted as ,in, , The number of intrinsic mode functions obtained from VMD decomposition; and other original multi-source temporal features that did not enter the VMD mode decomposition branch are retained; Step S2: Obtain the eigenmode functions output in step S1. Complex metric calculations are performed on the discrete mode sequences corresponding to each intrinsic mode function to construct an entropy weight vector that reflects the prediction difficulty of each intrinsic mode function. ; Step S3: Generate the discrete mode sequence corresponding to each eigenmode function and the corresponding entropy weight vector. The input is processed using a spatiotemporal autoencoder network for feature encoding. Step S4: Perform multi-source heterogeneous data fusion and reconstruction to obtain a standardized input matrix, and perform multi-offset time embedding and offset sub-sequence construction. Input the generated multiple offset sub-sequences into the MI-KAN module to perform nonlinear feature mapping on each offset branch. Step S5: After obtaining the high-order representation of each offset subsequence, perform multi-head self-attention operation on each offset branch to learn the correlation between different time positions under the same offset perspective; Step S6: Joint optimization of key parameters of MOTE and MI-KAN based on FNO; Step S7: Solve for the optimal parameters output. The process is re-injected into the MOTE and MI-KAN modules to re-execute the multi-offset subsequence construction, nonlinear higher-order representation extraction, MOTI modeling, and global fusion, resulting in a unified fused feature representation with optimal parameter configuration. .

2. The wind power prediction method combining MCG-STE, Time-TK, and FNO for joint optimization as described in claim 1, characterized in that: In step S1, the VMD algorithm uses the alternating direction multiplier method to iteratively solve the frequency domain expression of each eigenmode function and its center frequency: In the During the iteration, with the remaining eigenmode components, Lagrange multipliers, and quadratic penalty factor fixed, the ... Frequency domain expression of each eigenmode function And calculate the corresponding center frequency based on the updated frequency domain expression. Its update formula is as follows: ； ； in, For the first The modal component in the ... Frequency domain expression under the next iteration; This is the center frequency corresponding to this mode; The Fourier transform of the original wind power signal; Indicates the current mode The sum of all other modal components; For the frequency domain form of Lagrange multipliers; It is a secondary penalty factor; Repeat the above alternating iteration on the frequency domain expressions and center frequencies of each intrinsic mode function until the changes between the frequency domain expressions obtained in two adjacent iterations satisfy the preset convergence condition; for the converged... Frequency domain modes Performing the inverse Fourier transform yields the... Individual eigenmode functions ; Eigenmode functions Together they constitute the VMD decomposition result of step S1 and serve as the input for complex metric quantification in step S2.

3. The wind power prediction method combining MCG-STE, Time-TK, and FNO for joint optimization as described in claim 1, characterized in that: Step S2 includes: S201. The result obtained in step S1 is... Individual eigenmode functions Discretize according to a uniform sampling interval to obtain a length of Discrete mode sequence and in scale factor The discrete mode sequence is then subjected to composite multi-scale coarsening processing to obtain coarse-grained time subsequences. Its composite multi-scale coarsening formula is: ； In the formula, This indicates that the first step S1 obtained through VMD decomposition is... One intrinsic mode function; Indicates the first After discretizing the intrinsic mode functions according to a uniform sampling interval, the th intrinsic mode function is... The values ​​at each discrete-time sampling point; Indicates the eigenmode function index; Indicates the index of discrete-time sampling points; Indicates the length of the discrete mode sequence; Indicates the scale factor; Indicates the position index of the coarse-grained sequence; Represents scale factor The coarse-grained time series values ​​obtained below; S202. Calculate the time series for each transformation. fine state probability : ； S203. Obtaining RCMFDE under different scaling factors based on fractional calculus: ； In the formula, This is the coarsened time subsequence; The scaling factor is a positive integer subject to the following constraint: ; Let be the interval probability of cosine similarity; For multi-scale time series at the same scale factor The relative frequency below; Represents the interval number; The fractional order limit is represented as ; Indicates the embedding dimension; These are the elemental characteristic parameters of the state probability; Represents the gamma function; Represents a bi-gamma function; RCMFDE refers to a sophisticated composite multi-scale fractional divergence entropy complex quantification. S204. Calculate the divergence entropy value for the discrete mode sequence corresponding to each eigenmode function, and then concatenate them after normalization to form... The vector is injected into STAE, where STAE refers to the Spatiotemporal Attention Autoencoder Network.

4. The wind power prediction method combining MCG-STE, Time-TK, and FNO for joint optimization as described in claim 1, characterized in that: Step S3 includes: By utilizing a cascaded structure of a one-dimensional convolutional network and a bidirectional long short-term memory network, spatial local features and bidirectional temporal dependency features are extracted sequentially: Obtain the entropy weight vector The corresponding number in the middle scalar weight elements of a discrete modal sequence This is replicated and extended in the time dimension to construct a sequence similar to the discrete mode sequence. One-dimensional weight vectors of the same length are used, and the one-dimensional weight vectors are matched with the corresponding discrete mode sequences. Perform element-wise multiplication and fusion to generate weighted feature input terms. and The weighted feature input terms are then input into the network and adjusted through a dual path of high-entropy enhancement and low-entropy protection. The formulas for spatial feature extraction and bidirectional temporal feature extraction of the encoder are as follows: ； ； ； ； In the formula, and For the first A discrete mode sequence With corresponding scalar weight elements The weighted feature input term is formed after element-wise multiplication and fusion; This is a dot product operation; For a one-dimensional convolutional network The output features of each convolutional kernel; and These are the kernel weights and bias terms, respectively. For activation functions; The hidden state output of the feedforward long short-term memory network; This is the hidden state output of the backward Long Short-Term Memory network; These represent the input gate, output gate, forget gate, and cell state within the LSTM network, respectively. for Network bias; This represents the final spatiotemporal feature output after bidirectional splicing. For the discrete mode sequence corresponding to all intrinsic mode functions, the above encoding process is performed respectively to obtain the time of each mode. Bidirectional spatiotemporal feature output representation The outputs of each modality are spliced ​​and fused along the feature dimension to form a time-series graph. Unified Enhanced Feature Representation .

5. The wind power prediction method combining MCG-STE, Time-TK, and FNO for joint optimization as described in claim 1, characterized in that: Step S4 includes: S401: Based on the original multi-source time series data that has undergone unified resampling and time alignment in step S1 and has not yet entered VMD mode decomposition, extract the time... The corresponding original multi-source feature vector ,in, It consists of wind speed, wind direction, air density, ambient temperature, unit operating status, and numerical weather forecast characteristics that do not enter the VMD mode decomposition branch; Features output in step S3 The enhanced representation of the historical wind power channel is obtained after VMD mode decomposition, RCMFDE complex quantization and STAE spatiotemporal coding. Will and By concatenating and fusing features along their respective dimensions, a joint input feature vector is constructed. ； In the formula, This represents a vector concatenation operation performed along the feature dimension; Indicates time The total dimension of the fused joint input feature vector is denoted as . ; Step S402: After acquiring data from multiple consecutive time points, sort the data according to the historical window length. Stack the input sequences along the time dimension to form a multivariate time series matrix: ； In the formula, This represents the historical sequence matrix of the model input; Indicates the length of the history window; Step S403: For the input matrix Normalization is performed to obtain the standardized input matrix: ； In the formula, This represents the normalized input matrix; This represents the normalization operator; This represents the normalized single-step feature vector; Step S404: Standardize the input matrix Perform multi-offset splitting to generate multiple sets of offset subsequences: ； In the formula, Indicates the preset offset number. Indicates the first The offset subsequence, the th offset subsequence Each offset subsequence is defined as: ； ； In the formula, Indicates the offset interval. Indicates the starting index of the offset position. Indicates the first The offset subsequence within the historical window length The maximum number of offset steps achievable under constraints as well as Both represent the standardized input matrix The normalized single-step feature vector at the corresponding time position; for those exceeding the historical window length Elements that are not included in the corresponding offset subsequence; Multiple observation perspectives with different time spans are formed within the same historical window, thus simultaneously representing short-term fluctuation information and long-term trend information; Step S405: Input the generated multiple offset sub-sequences into the MI-KAN module and perform nonlinear feature mapping on each offset branch: The inter-layer mapping relationship is represented as follows: ； In the formula, Indicates the first Layer The output of each neuron; Indicates the first Layer The output of each neuron; Indicates from the first Layer The first neuron to the second Layer A learnable unary mapping function between neurons; This indicates summing and aggregating the outputs of all relevant neurons in the previous layer; when hour, This represents any offset subsequence generated in step S404. The first normalized single-step feature vector corresponding to the time position in the middle Each characteristic component; when hour, This indicates the network layer above the MI-KAN module. scalar output of a neuron; The unary mapping function adopts the form of a radial basis function, and the input variable of the current unary mapping function is denoted as... When the unary mapping function is located in the first layer of the MI-KAN module, For any offset subsequence generated in step S404 A feature component of the normalized single-step feature vector corresponding to the time position; when the unary mapping function is located in a subsequent network layer of the MI-KAN module. This represents the scalar output of the neuron in the previous network layer. The radial basis function is expressed as: ； In the formula, Indicates the output of the radial basis function; Indicates input variables The distance between the corresponding radial base center; The smoothing parameter represents the radial basis function; Represents the natural exponential function; Based on the radial basis function, the nonlinear mapping output of the current input variable is expressed as: ； In the formula, Indicates input variables The nonlinear mapping output obtained after processing by a radial basis function network; This represents the scalar input of the current unary mapping function of the MI-KAN module, when the unary mapping function is located in the first layer. The feature component originating from the corresponding normalized single-step feature vector in the offset subsequence generated in step S404, when the unary mapping function is located in a subsequent network layer... This is the scalar output of the neuron in the previous network layer; Represents radial basis functions; Indicates the first Learnable weights corresponding to radial basis functions; Indicates the first One radial base center; Indicates input variables With the The distance between the centers of the radial bases; Indicates the number of radial base centers; After processing by the MI-KAN module, the high-order representations of each offset subsequence are obtained: ； In the formula, This represents the high-order feature representation extracted by the MI-KAN module; The representation function of the MI-KAN module; This represents multiple offset subsequences of the input.

6. The wind power prediction method combining MCG-STE, Time-TK, and FNO for joint optimization as described in claim 1, characterized in that: In step S5, the correlation between different time positions under the same offset viewpoint is learned, represented as follows: ； In the formula, Indicates the first Feature representation of each offset branch after self-attention enhancement; Indicates the first A high-order representation of each offset subsequence; This represents a multi-head self-attention operation; the three parameters in the formula correspond to the query term, key term, and value term in the attention mechanism, respectively. The interaction results of multiple offset branches are concatenated along the feature dimension and mapped to the desired result using linear projection. The same hidden space yields multi-offset interaction features: ； In the formula, This represents a linear projection operator used to map the concatenated attention interaction results of multiple offset branches to a coordinate system. The same hidden dimension is used to meet the dimension alignment requirements in the subsequent global fusion process; Multi-off interaction features after linear projection As a query term, the standardized input matrix As key and value terms, global multi-head attention fusion is performed, and combined with residual connections to obtain a unified temporal representation: ； In the formula, This represents the unified temporal feature representation after global fusion; , , These represent the query item, key item, and value item, respectively. This indicates a multi-head self-attention operation.

7. The wind power prediction method combining MCG-STE, Time-TK, and FNO for joint optimization as described in claim 1, characterized in that: Step S6 includes: (1) Initialization phase Initialize the population, and uniformly denote the candidate parameter positions of FNO as... Each candidate parameter position corresponds to a set of parameters to be optimized, including the offset number parameter and offset interval parameter in the construction of the offset subsequence, as well as the radial basis function smoothing parameter and the number of radial basis centers in MI-KAN; the first... The positions of the candidate parameters are represented as follows: ； In the formula, Represents the first in the search space One candidate parameter position; Represents the decision variables in the corresponding dimension; The parameter space dimension is represented; multiple candidate parameter positions together constitute the population, and the population size is denoted as . ; Among them, the offset parameter Offset interval parameter and the number of radial base centers The smoothing parameters of the radial basis function are obtained using integer encoding. Continuous real number encoding is adopted; integer parameters are first rounded, and then boundary clipping is performed according to the preset upper and lower bounds of each parameter. (2) Determining the farthest optimal location For any candidate position in the population, FNO searches for the farthest and highest-quality candidate position based on the distance between the objective function value and the candidate position. This farthest optimal position is expressed as: ； In the formula, Indicates candidate position The corresponding farthest optimal position; and Representing candidate positions and The objective function value; This represents the maximum distance between candidate positions. This stage is used to enhance the global exploration capability during the parameter search process; (3) The stage of determining the most recent inferior candidate positions Based on determining the farthest optimal position, FNO further filters out the candidate positions closest to the current position from those that satisfy the objective function value worse than the current position, and uses these as the nearest inferior candidate positions. The expression for this is: ； In the formula, Indicates candidate position The corresponding nearest poor candidate position; This represents the minimum distance between candidate positions; (4) Dynamic Focusing Stage To broaden the search scope in the early stages and enhance local development capabilities in the later stages, FNO introduces a dynamic focusing strategy, Depth-First Search (DFS), which uses control parameters that gradually change with each iteration. Adjusting the search area, its expression is: ； ； In the formula, Indicates the number of times the function has been evaluated; This indicates the maximum number of function evaluations. This represents a random number that follows a uniform distribution within the interval [0,1]; as the iteration progresses, By gradually reducing the size of the search area, the search region gradually shrinks around the far-optimal position, thereby enhancing the ability for fine-grained optimization in the later stages. (5) Candidate parameter update stage The FNO position update includes two processes: the nearest inferior position jump update and the farthest optimal position update. The temporary position of the candidate position is calculated, and its expression is: ； ； When the probability condition is met, perform a jump to the nearest inferior position: ； FNO performs farthest-optimal exploration and updates using the farthest-optimal location and a dynamic focusing strategy. Its expression is: ； ； ； In the formula, Indicates the switching factor; Represents a random number within the interval [0,1]. Indicates the step size coefficient; Indicates the position of the updated candidate parameter; (6) Fitness evaluation and memory update stage For the updated new position, perform boundary clipping based on the upper and lower bounds of the parameters; During the parameter optimization phase, input data and corresponding label data are constructed based on known labeled samples, and the current candidate parameters are solved. The process of re-injecting the MOTE and MI-KAN modules, constructing multiple offset subsequences, extracting nonlinear higher-order representations, modeling MOTI, and global fusion is re-executed to obtain the unified temporal representation corresponding to the candidate parameter solution. Based on this, by invoking the linear mapping layer and the inverse normalization operator, the wind power prediction sequence of the candidate parameter solution in the prediction time domain corresponding to the labeled sample is derived forward. The predicted sequence is compared with the actual wind power sequence in the labeled sample labels. Compare the results and calculate the fitness function value of the current candidate parameter solution; (7) Iteration Termination Phase Repeat the above process of determining the farthest optimal position, determining the nearest inferior candidate position, dynamic focusing, updating candidate parameters, and evaluating fitness until the maximum number of function evaluations is reached or the convergence condition is met, and output the optimal parameter solution. The optimal parameter solution The MOTE and MI-KAN modules are used to inject data back into the wind power prediction model for subsequent prediction calls.

8. The wind power prediction method combining MCG-STE, Time-TK, and FNO for joint optimization as described in claim 1, characterized in that: Step S7 includes: By inputting the unified fusion feature representation into the linear mapping layer, the wind power prediction sequence in the future prediction time domain is directly obtained: ； In the formula, Represents a linear mapping layer; Indicates the prediction step size; This indicates that in the optimal parameter solution A unified temporal feature representation obtained through global fusion under constraints; Indicates from time At the time The predicted wind power sequence; Finally, inverse normalization is performed on the wind power prediction sequence to restore the prediction results to their true physical dimensions: ； In the formula, Indicates the inverse normalization operator; This represents the final wind power prediction result after restoring it to its true physical dimensions.

Images