A wind turbine multi-source data fusion and power prediction method

Abstract

The present invention discloses a method for multi-source data fusion and power prediction of a wind turbine. The method comprises: S1: preprocessing raw data from a wind turbine monitoring and data acquisition system, including feature selection, outlier removal, and data set generation and partitioning; S2: constructing a Transformer basis generation module and combining it with a multi-layer fully connected neural network to form a Transformer block; S3: constructing a power curve basis generation module and combining it with a multi-layer fully connected neural network to form a power curve block; S4: connecting several Transformer blocks using the double residual stacking principle to form a Transformer stack; S5: connecting several power curve blocks using the double residual stacking principle to form a power curve stack; S6: connecting the Transformer stack and the power curve stack in series, establishing a dynamic trainable weighting mechanism, fusing the outputs of each stack, and obtaining a multi-step wind power prediction result; and S7: determining the optimal configuration scheme for network hyperparameters based on the prediction results using a grid search method.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power prediction, and in particular relates to a method for multi-source data fusion and power prediction of a wind turbine generator set. Background Art

[0002] As wind power continues to grow in the global renewable energy landscape, accurate and physically interpretable wind power forecasting has become a critical support for the safe operation and economic dispatch of power grids. During wind turbine operation, the high-frequency fluctuations and intermittent nature of wind speed not only lead to strong nonlinearity in power output and significant threshold constraints (cut-in wind speed, rated wind speed, and cut-out wind speed), but also, because monitoring data often has a low sampling rate and high noise, the complexity and uncertainty of the input features are further increased. The demand for high accuracy is particularly urgent in short-term forecasts (ranging from minutes to hours), as they are directly related to reducing wind curtailment rates, optimizing power generation, and rationally formulating turbine maintenance plans. However, the cumulative error effect and insufficient model robustness in multi-step forecasts continue to hinder improvements in forecasting performance.

[0003] Existing wind power forecasting technologies can be roughly divided into two categories: statistical models based on random processes and data-driven models based on machine learning. Statistical methods such as autoregressive moving average and Gaussian process regression can, to a certain extent, capture the randomness and temporal correlation of wind speed and power. However, their strong reliance on stationarity and linearity assumptions makes it difficult to handle the highly non-stationary and multi-source coupling characteristics of wind power monitoring data. In contrast, traditional machine learning methods such as support vector machines, k-nearest neighbors, and shallow neural networks have made significant progress in short-term wind power forecasting through nonlinear mapping and multi-dimensional feature fusion. However, these models are still limited in deeply capturing long-term temporal dependencies and adaptively modeling large-scale heterogeneous data.

[0004] In recent years, deep learning methods have rapidly become a research hotspot for wind power forecasting, thanks to their end-to-end feature extraction and powerful nonlinear fitting capabilities. Recurrent and convolutional architectures such as long short-term memory (LSTM), gated recurrent units (GRU), and temporal convolutional networks (TCN) effectively capture short-term and local spatiotemporal features through gating mechanisms or dilated convolutions. Transformer-based models such as Informer and Autoformer further utilize global self-attention mechanisms to improve medium- and long-term forecasting accuracy. Furthermore, graph neural networks and their variants enable modeling of non-Euclidean spatial dependencies by constructing turbine or sensor topologies. However, most of these deep models adhere to a "black box" paradigm, ignoring the physical characteristics and operational constraints of wind turbines. This results in predictions that lack credibility and robustness under extreme operating conditions or topological changes.

[0005] To address the lack of interpretability of "black box" models, physics-informed machine learning methods are beginning to be applied to wind power forecasting. These include paradigms such as using conservation equations as loss regularization terms, physical pre-training based on simulation data, hybrid physics-deep learning models, and the design of physics-aware architectures. While these methods incorporate physical priors into the network to some extent, they are often hindered by complex high-order derivative calculations, large discrepancies between simulated and measured data, difficulty in adaptively adjusting integrated strategy weights, and the need for repetitive architecture design tailored to specific scenarios. These limitations hinder their widespread application in large-scale, multi-operating wind farms.

[0006] In summary, achieving flexible adaptation to wind power time series patterns while balancing high-precision predictions with physical interpretability is a core challenge that currently faces challenges in wind power forecasting technology. Addressing the limitations of existing Neural Basis Analysis for Extensions (NBEATSx) networks, such as static equal-weight aggregation and the lack of physical meaning in basis functions, introducing physical priors and assigning dynamic trainable weights to the model, in order to create a wind power forecasting architecture that is both interpretable and robust, is crucial for achieving engineering applications of wind power forecasting technology. Summary of the Invention

[0007] In response to the above-mentioned problems existing in the prior art, the present invention provides a method for multi-source data fusion and power prediction of wind turbines, which takes into account both high prediction accuracy and physical interpretability while achieving flexible adaptation to wind power timing patterns.

[0008] In order to achieve the above object, the present invention provides a method for multi-source data fusion and power prediction of a wind turbine generator system, comprising the following steps:

[0009] S1. Data preprocessing of the raw data of the wind turbine monitoring and data acquisition system, including feature selection, outlier removal, data set generation and partitioning;

[0010] S2. Construct a Transformer-based generation module and combine it with a multi-layer fully connected neural network to form a Transformer block:

[0011] S3, constructing a power curve base generation module and combining it with a multi-layer fully connected neural network to form a power curve block;

[0012] S4. Connect several Transformer blocks through the double residual stacking principle to form a Transformer stack:

[0013] S5, connecting the plurality of power curve blocks through a double residual stacking principle to form a power curve stack;

[0014] S6: Connect the Transformer stack and the power curve stack in series, establish a dynamic trainable weighting mechanism, fuse the outputs of each stack, and obtain a multi-step prediction result of wind power;

[0015] S7. Determine the optimal configuration of network hyperparameters using the grid search method based on the prediction results.

[0016] Furthermore, the Transformer base is combined with a multi-layer fully connected neural network to form a Transformer block: a multi-level "stack-block" architecture is adopted to achieve time series decomposition and prediction by integrating exogenous variables; the Transformer block is the basic structural unit of the multi-layer weighted connection neural network, all Transformer blocks follow the same logic to process input data, and multiple blocks are connected to form a stack structure through the double residual stacking principle; the Transformer block includes a dynamic residual part and a static exogenous variable part, and the dynamic residual part is expressed as:

[0017]

[0018] in, represents the stack index, Represents the block index, represents the batch size, is the length of the lookback window; the designed multi-layer fully connected neural network outputs two vectors: including the lookback value of the Transformer block and predicted values ,in, is the prediction window length.

[0019] Furthermore, the Transformer block contains a multi-layer fully connected neural network that is responsible for learning the basis expansion coefficients The specific process is as follows:

[0020]

[0021] Network Integration Backtracking Base and prediction base ; In the base layer, for each sample , using the backtracking basis and prediction base Perform explicit summation operations and expand the backtracking basis coefficients and the predicted base expansion coefficient Mapping to backtracking value and predicted values , the process is expressed as:

[0022]

[0023] .

[0024] Furthermore, the generation step of the Transformer base includes:

[0025] S201: Given an input tensor and Splicing along the time dimension ,in , is the lookback window, is the prediction window; following the standard Transformer architecture, the tensor is transposed to And perform linear projection to get :

[0026]

[0027] in, is the embedding layer size, scaling factor Used to stabilize the gradient magnitude; the time position information is integrated into the embedded features through the sine encoding function, the position encoding matrix Offline precomputation, where Encodes the length of the maximum position for the precomputed sequence and dynamically truncated to match the sequence length. , the final encoding is expressed as:

[0028]

[0029] S202: Apply the Transformer encoder layer for encoding. Each layer contains multi-head self-attention, residual connection and position-by-position feedforward network. The process is expressed as:

[0030]

[0031] in, is the lower triangular causal mask, defined as:

[0032]

[0033] S203: Output the Transformer through the output projection layer From the dimension Mapping to target dimension :

[0034]

[0035] By combining permutation and residual connection operations, the fusion is expressed as:

[0036] .

[0037] Furthermore, the power curve base generation module uses a smooth logistic growth model to capture the nonlinear characteristics of the wind power curve, and its mathematical expression is:

[0038]

[0039] in, Indicates wind speed The predicted power at is the maximum power output, is the steepness parameter that controls the rising rate of the curve, is the half-maximum power wind speed, and are the cut-in wind speed and rated wind speed respectively.

[0040] Furthermore, the processing method of the power curve base generation module includes:

[0041] S301: Parameter processing and dimension adjustment, setting and Represent the wind speed data in the lookback window and the forecast window respectively, and concatenate the two input components along the time dimension:

[0042]

[0043] in, ; Based on rated cut-out wind speed Formula The parameters in are adjusted to scale the horizontal and vertical coordinates of the smoothed logistic growth model to the range [0,1], where the horizontal coordinate represents wind speed and the vertical coordinate represents power:

[0044]

[0045] in, , , , , , , ;parameter 、 and the input sequence The dimensions of are adjusted as follows:

[0046]

[0047]

[0048]

[0049] S302: Construct the power curve basis function and solve the smoothed logistic growth model through element-level broadcasting in tensor calculations:

[0050]

[0051] The third dimension and the fourth dimension After merging and removing the singleton dimension, we get:

[0052]

[0053] in, ,and ; Calculation formula Nonlinear scaling term in And after expanding its dimension we get:

[0054]

[0055] The final basis function matrix is ​​obtained as follows:

[0056]

[0057] in, represents element-wise multiplication, and ; Finally, through tensor slicing Split into backtracking basis function and predictive basis function:

[0058]

[0059] S303: base expansion coefficient processing, and Apply the ReLU activation function to ensure non-negativity, and then normalize it through softmax:

[0060]

[0061] in, Represents normalization along the last dimension of the tensor; the result is then split into the back-basis expansion coefficients and the predicted base expansion coefficient .

[0062] Furthermore, the stack is obtained by: multiple blocks form a stack, The input of the stack is the backtrace value , the prediction output of the stack is aggregated from the prediction results of each block within it:

[0063] .

[0064] Furthermore, the multi-step prediction result of wind power is the weighted sum of all stack prediction results:

[0065]

[0066] in, is the trainable weight coefficient.

[0067] Furthermore, the number of the Transformer blocks and the power curve blocks are respectively obtained by grid search optimization using the data set.

[0068] Compared with the prior art, the advantages and positive effects of the present invention are embodied in the following aspects:

[0069] First, through dual residual connections and a multi-layer "stack-block" architecture, the present invention achieves efficient propagation of residuals and multi-component separation modeling of signals, organically combining the nonlinear fitting ability of deep learning with the interpretability of traditional decomposition methods, greatly improving the accuracy and stability of time series feature extraction; secondly, a trainable weight matrix is ​​introduced to replace the traditional fixed accumulation strategy, so that the model can dynamically and adaptively adjust the contribution of each stack output to the final prediction, enhancing the response capability to power changes under different wind speed modes; thirdly, the operating constraints such as the cut-in wind speed, rated wind speed, cut-out wind speed and rated power of the wind turbine are explicitly embedded in the smooth logistic growth basis function, and physical priors are injected through a dedicated power curve stack to ensure that the output strictly meets the physical consistency requirements and conforms to the nonlinear saturation characteristics of the wind turbine; finally, the Transformer encoder is used to generate basis vectors to effectively capture complex time dependencies. The above design not only overcomes the limitations of the traditional NBEATSx fixed stack and simple aggregation method, but also achieves deep synergy between physical consistency and interpretability, providing a more accurate, reliable and interpretable solution for wind power forecasting, and effectively bridging the gap between physical prior knowledge and data-driven models. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 This is a flow chart of a method for multi-source data fusion and power prediction of a wind turbine generator system according to an embodiment of the present invention;

[0071] Figure 2 This is a schematic diagram of the structure of a physical information adaptive weighted neural-based extended analysis network according to an embodiment of the present invention;

[0072] Figure 3 This is a comparison diagram of the smoothed logistic growth model curve shapes under different parameters in an embodiment of the present invention;

[0073] Figure 4 This is a schematic diagram of the overall framework of the short-term wind power prediction method according to an embodiment of the present invention;

[0074] Figure 5 Generate a schematic diagram of the module framework for the Transformer base;

[0075] Figure 6 Schematic diagram of the module framework for power curve base generation;

[0076] Figure 7 Box plot of the distribution of the prediction error of the Bi-LSTM network at each prediction step;

[0077] Figure 8 Box plot of the distribution of the Transformer network's prediction error at each prediction step;

[0078] Figure 9 Box plot of the distribution of the prediction error of the TCN network at each prediction step;

[0079] Figure 10 Box plot of the distribution of prediction error of the NBEATSx network at each prediction step;

[0080] Figure 11 This is a box plot of the distribution of the prediction error of the PIAW-NBEATSx network at each prediction step according to an embodiment of the present invention;

[0081] Figure 12 Global predictions generated for the PIAW-NBEATSx network Comparison chart with the measured power curve;

[0082] Figure 13 Transformer stack prediction for the PIAW-NBEATSx network Decomposed component graph;

[0083] Figure 14 Power curve stack prediction for the PIAW-NBEATSx network Decomposed component graph;

[0084] Figure 15 is the dynamic weight coefficient of the PIAW-NBEATSx network and Evolution graph during training;

[0085] Figure 16 Figure 2 shows the stable convergence process of training and validation loss of the PIAW-NBEATSx network.

[0086] Figure 17 This is the interval prediction result diagram of the PIAW-NBEATSx network at different confidence levels;

[0087] Figure 18 for Figure 17 A partial enlarged view of the middle A;

[0088] Figure 19 for Figure 17 A partial enlarged view of point B in the middle. DETAILED DESCRIPTION

[0089] The present invention is described in detail below by way of exemplary embodiments. However, it should be understood that the present invention may be beneficially incorporated into other embodiments without further description.

[0090] This paper provides a method for multi-source data fusion and power prediction for wind turbines based on a physical information adaptive weighted neural basis extended analysis network. This method, based on a short-term wind power prediction framework using a physical information adaptive weighted neural basis extended analysis (PIAW-NBEATSx) network, achieves a deep fusion of physical prior knowledge and data-driven models to predict wind turbine power. This model directly embeds wind turbine operating constraints and wind power curve characteristics into the architecture through a dedicated power curve stack, and leverages the Transformer stack to capture complex temporal dependencies. Furthermore, the model uses a dynamically trainable weighting mechanism instead of a fixed aggregation approach to adaptively balance the contributions of each stack. These designs not only ensure the physical consistency of the prediction results but also enhance the model's interpretability and robustness to non-physical outputs.

[0091] See also Figure 1 , based on data fusion physical information adaptive weight neural base extended analysis network to perform short-term wind power forecasting. The method includes:

[0092] S1. Data preprocessing is performed on the raw data of the wind turbine supervisory control and data acquisition system (SCADA), including feature selection, outlier removal, data set generation and partitioning.

[0093] The specific feature selection method is to calculate the correlation matrix using the Pearson correlation coefficient to achieve feature selection. For variable groups with near-perfect collinearity, only one representative variable is retained. Physical outliers are then removed, and the data is normalized to the range [0, 1]. This completes preliminary preprocessing and generates a dataset. The dataset is then divided into training, validation, and test sets according to a specific ratio. The specific ratio can be determined based on the amount of data and actual needs, and is not limited here.

[0094] S2. Construct a Transformer base generation module and combine it with a multi-layer fully connected neural network to form a Transformer block.

[0095] This method uses a multi-level "stack-block" architecture to achieve time series decomposition and prediction by integrating exogenous variables. The Transformer is a deep learning architecture. The Transformer block is the basic structural unit of a multi-layer weighted neural network. All Transformer blocks follow the same logic to process input data. Multiple Transformer blocks are connected to form a stack structure through the double residual stacking principle. For example, stack Middle Block The input of consists of two parts: dynamic residual part:

[0096]

[0097] in, represents the batch size, is the length of the lookback window; and the static exogenous variable part The Transformer block then outputs two vectors: the block’s lookback value and and predicted values ,in, is the prediction window length.

[0098] Each Transformer block contains a multi-layer fully connected neural network (FCNN) that is responsible for learning the basis expansion coefficients The specific process is as follows:

[0099]

[0100] Multi-layer fully connected neural network integrates backtracking basis and prediction base , back-tracking and prediction base It can be a predefined basis function or a basis vector learned from data. In the base layer, for each sample , using the backtracking basis and prediction base Perform explicit summation operations and convert the backtracking basis expansion coefficients in the basis expansion coefficients into and the predicted base expansion coefficient Mapping to backtracking value and predicted values The process can be expressed as:

[0101]

[0102]

[0103] The generation step of the Transformer base consists of the following three core stages:

[0104] S201: Embedding and Positional Encoding

[0105] Given an input tensor (lookback window) and (prediction window), first spliced ​​along the time dimension to form ,in Following the standard Transformer architecture, the tensor is transposed into And perform linear projection to get :

[0106]

[0107] in, is the embedding layer size, scaling factor Used to stabilize the gradient magnitude.

[0108] The temporal position information is incorporated into the embedded features through the sinusoidal encoding function. Position encoding matrix Offline precomputation ( is the precomputed maximum positional encoding length) and is dynamically truncated to match the sequence length The final encoding is:

[0109]

[0110] S202: Construct Transformer encoder layer for encoding

[0111] Applying Transformer encoder layers, each layer contains multi-head self-attention, residual connections, and position-wise feed-forward networks (FFNs), the process is expressed as:

[0112]

[0113] in, is the lower triangular causal mask, defined as:

[0114]

[0115] This mask is used to mask out future time steps to preserve causal relationships in the attention mechanism.

[0116] S203: Output projection and residual connection

[0117] Transformer output is passed through the output projection layer From the dimension Mapping to target dimension (For blocks that obtain basis vectors based on data learning models, there are ):

[0118]

[0119] By combining permutation and residual connection operations, the fusion is expressed as:

[0120]

[0121] S3. Construct a power curve base generation module and combine it with a multi-layer fully connected neural network to form a power curve block.

[0122] Smooth logistic growth models (SLGMs) are used to capture the nonlinear characteristics of the wind power curve (WPC), and their mathematical expression is:

[0123]

[0124] in, Indicates wind speed The predicted power at is the maximum power output (usually the rated power), is the steepness parameter that controls the rising rate of the curve, is the half-maximum power wind speed (i.e. the power output at this wind speed is ), and are the cut-in wind speed and rated wind speed respectively. By setting different parameter combinations, curves of different shapes can be obtained, such as Figure 3 shown.

[0125] The processing flow of the power curve base generation module includes the following steps:

[0126] S301: Parameter processing and dimension adjustment

[0127] set up and Represent the wind speed data within the lookback window and forecast window respectively.

[0128] First, concatenate the two input components along the time dimension:

[0129]

[0130] in, .

[0131] because To standardize the sequence after processing, it is necessary to base it on the rated cut-out wind speed The parameters in Equation (8) are adjusted to scale the horizontal and vertical coordinates of the smoothed logistic growth model to the range [0,1], where the horizontal coordinate represents wind speed and the vertical coordinate represents power:

[0132] (9)

[0133] in, , , , , , , .

[0134] To facilitate tensor broadcasting in deep learning frameworks, the parameter 、 and the input sequence The dimensions of are adjusted as follows:

[0135]

[0136]

[0137]

[0138] S302: Constructing power curve basis function

[0139] The smoothed logistic growth model is solved using element-wise broadcasting in tensor computations:

[0140] (10)

[0141] The third and fourth dimensions (corresponding to and ) and removing the singleton dimension, we get:

[0142] (11)

[0143] in, ,and .

[0144] Calculate the nonlinear scaling term in formula (9) And after expanding its dimension we get:

[0145] (12)

[0146] The final basis function matrix is ​​obtained as follows:

[0147] (13)

[0148] in, represents element-wise multiplication, and Specifically, The operation changes the second and third dimensions of the tensor from Replace with .

[0149] Finally, by slicing the tensor Split into backtracking basis function and predictive basis function:

[0150] (14)

[0151] S303: Basis expansion coefficient processing

[0152] like Figure 2 Figure 1 shows a multi-layer “stack-block” architecture derived from Neural Basis Analysis for Extension (NBEATSx). The network consists of two stacks: a Transformer stack (blue blocks), whose basis vectors are generated by the Transformer Basis Generation Module; and a power curve stack (green blocks), whose predefined basis functions are generated by the Power Curve Basis Generation Module. Each block contains a multi-layer fully connected neural network (FCNN) to learn the basis extension coefficients. Global prediction value Through adaptive weight Build and achieve adaptive fusion of all stack prediction results.

[0153] For the power curve block, since the physical meaning of the basis expansion coefficient output by the fully connected neural network is the weight of the smooth logistic growth model SLGMs, it is necessary to first and Apply the ReLU activation function to ensure non-negativity, and then normalize it through softmax:

[0154] (15)

[0155] in, Represents normalization along the last dimension of the tensor. The result is then split into the back-basis expansion coefficients and the predicted base expansion coefficient .

[0156] S4. Connect several Transformer blocks through the double residual stacking principle to form a Transformer stack.

[0157] The number of Transformer blocks is obtained by grid search using the dataset. Multiple blocks form a stack. The input of the stack is the backtrace value , the prediction output of the stack is aggregated from the prediction results of each block within it:

[0158] (16)

[0159] S5. Connect several power curve blocks through the double residual stacking principle to form a power curve stack.

[0160] The number of power curve blocks is obtained by using a data set through grid search optimization.

[0161] S6. Connect the Transformer stack and the power curve stack in series, establish a dynamic trainable weighting mechanism, fuse the outputs of each stack, and obtain the multi-step prediction results of wind power.

[0162] The global prediction (i.e., model output) is the weighted sum of all stack predictions:

[0163] (17)

[0164] in, is the trainable weight coefficient.

[0165] S7. Use the grid search method to determine the optimal configuration scheme of the network hyperparameters based on the prediction results.

[0166] After obtaining the optimal configuration scheme, the wind power is predicted using the physical information adaptive weighted neural extended basis analysis network of the optimal configuration scheme.

[0167] The method will be further explained below with reference to historical data from a monitoring and data acquisition system for a wind farm in China throughout 2023.

[0168] 1) Obtaining datasets and performance metrics

[0169] The performance of the prediction model is verified using historical data from the Supervisory Control and Data Acquisition (SCADA) system of a wind farm in China for the entire year 2023. The data contains 37 variables with a sampling frequency of 15 minutes. Figure 4-6 As shown, the Transformer basis vector and power curve basis function generation module includes the cut-in wind speed , rated wind speed , Cut-out wind speed and rated power The key parameters including are explicitly embedded into the power curve basis function via a smooth logistic growth model.

[0170] The data set construction process is as follows: First, the correlation matrix is ​​calculated through the Pearson correlation coefficient to achieve feature selection. For the variable group that is approximately completely collinear, only one representative variable is retained. This process finally screened out 13 variables as static exogenous variables X, including wind speed, grid phase voltage, power factor, blade motor current value, blade angle, impeller speed, gearbox oil temperature, generator winding temperature, generator drive end / non-drive end bearing temperature, ambient temperature, converter machine side / grid side module temperature. The 13 variables are the selected features. Then, the physical outliers are removed, and the data is normalized to the range of [0,1] to complete the preliminary preprocessing and obtain the data set. Lookback window length Set to the prediction window length 5 times of that.

[0171] To evaluate the accuracy of the prediction model, the performance indicators used include mean absolute percentage error (MAPE), symmetric mean absolute percentage error (SMAPE), normalized mean absolute error (NMAE), normalized root mean square error (NRMSE) and coefficient of determination ( ).

[0172] 2) Experimental setup and proposed prediction model configuration

[0173] In the experiment, the dataset was split into training, validation, and test sets in a 6:2:2 ratio. The Physical Information Adaptive Weighted Neural Extended Basis Analysis (PIAW-NBEATSx) network, constructed in this paper, was trained on the training set using the Adaptive Moment Estimation (Adam) optimizer. The mean squared error (MSE) was used as the loss function. Hyperparameters were tuned using three grid search methods: architectural parameters (number of blocks per stack), , the number of hidden layers in each block , the number of hidden units in each layer ), regularization parameter (dropout rate , L1 regularization coefficient , L2 regularization coefficient ), learning rate scheduling (initial learning rate , attenuation factor , minimum loss reduction threshold ). The parameter update is guided by the model performance on the validation set. The final optimized network configuration is (This dataset uses network optimization to find three Transformer blocks and three power curve blocks respectively.) 、 、 、 、 、 、 、 , with details shown in Table 1. After training, the model generates five-step predictions on the test set, which are then concatenated to form a complete prediction sequence. All experiments were implemented in Python 3.9 and run on an NVIDIA RTX A4000 GPU.

[0174] Table 1 Network structure of Physical Information Adaptive Weight Neural Extended Basis Analysis (PIAW-NBEATSx)

[0175]

[0176] 3) Comparison of deterministic prediction results

[0177] Four baseline forecasting models were established, including the original Neural Basis Analysis for Extensions (NBEATSx) model and three sequence-to-sequence bidirectional long short-term memory (Bi-LSTM), temporal convolutional network (TCN), and Transformer variants. All models have the same input and output dimensions, with 14 dimensions for the input and 5 steps for the output, and all incorporate a residual shortcut from the original input to the output. The Neural Basis Analysis for Extensions (NBEATSx) model uses the architecture shown in Table 1, but utilizes classic trend and seasonal stacking.

[0178] Table 2 Hyperparameter configuration of the baseline model

[0179]

[0180] in, Represents the hidden layer size. The hidden representation dimension of Bi-LSTM is , including the forward dimension and the backward dimension.

[0181] All model hyperparameters were optimized using grid search across hidden layer size, number of layers, dropout rate, learning rate, and batch size, maintaining a consistent search space and using the same loss function and early stopping strategy. The main hyperparameters are summarized in Table 2.

[0182] Table 3 Performance comparison of prediction models

[0183]

[0184] Table 3 compares the performance of the physics-based adaptive weighted neural extended basis analysis (PIAW-NBEATSx) network with a baseline model across five metrics. Notably, due to the high volatility of real-world data and the presence of a large number of near-zero wind power values ​​(corresponding to wind turbine shutdown states), the baseline model exhibits extremely high mean absolute percentage error (MAPE). The baseline model's limited ability to capture complex temporal dependencies and multi-scale patterns leads to significant accumulated prediction errors for low-value samples. In contrast, the neural extended basis analysis (NBEATSx), with its interpretable stacking architecture and signal decomposition mechanism, mitigates this issue to some extent.

[0185] The proposed PIAW-NBEATSx showed significant advantages in all evaluation indicators, achieving not only the lowest error rate but also the highest determination coefficient. This outstanding performance stems from the innovative construction of a power curve stack, which explicitly embeds operational parameter constraints into the basis functions, ensuring that the prediction results conform to the nonlinear saturation characteristics of WPC. Furthermore, the introduction of a trainable weight mechanism and a Transformer stack further improves the accuracy and stability of predictions.

[0186] Overall, these comparative results demonstrate the necessity of incorporating prior knowledge into wind power forecasting. The proposed method not only improves the interpretability and reliability of the forecast results, but also demonstrates excellent forecast accuracy and robustness when processing real-world datasets containing a large number of near-zero samples.

[0187] Figure 7-11 Boxplots visualize the distribution of forecast errors for all models at each forecast step. As the forecast horizon increases, the interquartile range widens and the whiskers lengthen, reflecting increasing uncertainty in long-term forecasts. In contrast, the boxplot for the PIAW-NBEATSx model is significantly more compact, with a concentrated distribution of outliers. The median and mean (represented by the solid and dashed lines, respectively) are almost zero, indicating minimal bias and high stability. The baseline models, particularly the LSTM and TCN networks, have significantly wider boxplots and more dispersed outlier distributions, reflecting their poor stability and susceptibility to extreme errors.

[0188] 4) Discussion of the proposed method

[0189] i) Interpretability Analysis

[0190] Figure 12 Comparison of global predictions generated by the PIAW-NBEATSx network and the measured power curve; Figure 13 and Figure 14 Demonstrating Transformer stack predictions and power curve stack prediction Two decomposition components. Their dynamic weight coefficients and The evolution during training is as follows Figure 15 As shown: the weight starts from 1 and converges to and , indicating that the Transformer stack contributes significantly more to global prediction than the power curve stack. This suggests that temporal dependency modeling plays a dominant role in prediction, while the stack guided by physical prior information serves as a supplementary regularizer. Figure 16 It further demonstrates stable convergence of training and validation losses with no signs of overfitting.

[0191] ii) Sensitivity analysis

[0192] To evaluate the contributions of the three core innovations, this study conducted ablation experiments, focusing on dynamic trainable weights, power curve stacks, and Transformer stacks.

[0193] Table 4. Prediction performance evaluation of Nbeatsx variant models by replacing the Transformer stack in PIAW-NBEATSx with other types of stacks

[0194]

[0195] Table 5. Prediction performance evaluation of variant models by replacing the power curve stack in PIAW-NBEATSx with other types of stacks Nbeatsx

[0196]

[0197] The mean squared error (MSE) was used as the loss function for both training and validation. PI-NBEATSx represents the simplified network obtained by removing the dynamic weighting mechanism from the proposed method. Table 3 shows that after removing the dynamic weighting mechanism, the MAPE increased by 22.7% (from 9.21% to 11.30%), while R 2 It also decreases (from 0.9874 to 0.9846). This shows that the mechanism plays a key role in adaptively balancing the contributions of multiple stacks.

[0198] Table 4 lists the forecasting performance of several NBEATSx variants. Each variant decomposes the time series into different components through specialized stacks: the trend stack uses polynomial basis functions to model long-term trends; the seasonality stack uses sine and cosine basis functions to capture periodic oscillations; the identity stack omits basis function generation and directly outputs the fully connected network. (as defined in formula (2)) as the block prediction result In addition, architectures such as WaveNet, TCN, and LSTM can be used as alternative temporal dependency learners, which function similarly to the Transformer encoder to extract features from the input sequence and use the network output as the basis vector. Combined with the results in Table 3, it can be seen that the performance of the PIAW-NBEATSx architecture is significantly better than all variants, which verifies the Figure 2 The necessity of a dual stack design is shown.

[0199] When replacing the Transformer stack (Table 4), model performance degrades significantly: the traditional stack (trend / seasonality) experiences a MAPE spike of 144.6%–534.7% due to the limitations of predefined basis functions. While the general time series model (WaveNet / TCN / LSTM) partially mitigates this issue (MAPE ranges from 80.6%–294%), its R² value (≤ 0.978) remains lower than that of the proposed model. This demonstrates the irreplaceable nature of the Transformer as a core component of adaptive time series modeling—its multi-head attention mechanism effectively captures complex dependencies and avoids the inductive bias of traditional methods that assume trends or periodicity.

[0200] Experiments replacing the power curve stack (Table 5) reveal its domain-specific value. While general-purpose time series models such as WaveNet achieve a MAPE (10.1522%) closest to PIAW-NBEATSx, their NMSE (0.1603%) remains higher than the original architecture (0.1463%). Traditional stacks exhibit extreme errors due to physical property mismatches. These results demonstrate that the power curve stack, by embedding physical constraints derived from WPC, addresses the generalization limitations of purely data-driven models in specific scenarios. Its role as a carrier of domain knowledge is particularly essential when addressing nonlinear saturation effects.

[0201] The innovative synergy of the two types of stacks is the key to performance improvement: the Transformer stack acts as a general time series engine to extract complex features, while the power curve stack injects domain prior knowledge (such as wind power physical constraints) to ensure physical consistency. Ablation studies show that removing either stack leads to a significant performance degradation: replacing the Transformer stack leads to overall performance degradation, while replacing the power curve stack leads to prediction collapse in specific scenarios. In contrast, traditional NBEATSx variants rely on universal basis functions and have difficulty in representing the complex domain-specific structures in industrial time series. This verifies the design advantage of PIAW-NBEATSx - it not only overcomes the limitations of the fixed stack in traditional NBEATSx, but also enhances the engineering applicability of the framework through configurable domain-specific blocks. The proposed method provides more accurate, reliable and interpretable predictions, providing a new approach for the research and application of complex wind power time series analysis.

[0202] iii) Uncertainty analysis

[0203] We use Pinball loss to evaluate the probability prediction ability of the PIAW-NBEATSx network, which is defined as:

[0204]

[0205] This loss function quantifies the difference between the predicted quantile and the actual value. Deviation from the level.

[0206] like Figure 17-19 As shown in the figure, the model achieves 50%, 70%, and 90% empirical coverage, with average interval widths of 59.90 kW, 93.82 kW, and 199.03 kW, respectively. The relatively narrow intervals (especially at the 50% level) indicate that the model has the ability to accurately quantify uncertainty.

[0207] It is worth noting that Figure 18 The results show that PIAW-NBEATSx maintains high accuracy and low uncertainty in the low power range (≤ 750 kW). This local accuracy explains its MAPE advantage over the baseline model and all NBEATSx variants in Table 4. The large number of near-zero values ​​in the wind power dataset significantly increases the MAPE, while PIAW-NBEATSx demonstrates its excellent robustness in power forecasting with the lowest MAPE of 9.21%.

[0208] Of course, the above description is only a preferred embodiment of the present invention, and the present invention is not limited to the above-mentioned embodiments. It should be noted that all equivalent substitutions and obvious deformation forms made by any technician familiar with this field under the guidance of this specification fall within the substantive scope of this specification and should be protected by the present invention.

Claims

1. A method for multi-source data fusion and power prediction of wind turbines, characterized in that: The steps include: S1. Data preprocessing of the raw data of the wind turbine monitoring and data acquisition system, including feature selection, outlier removal, data set generation and partitioning; S2. Construct a Transformer-based generation module and combine it with a multi-layer fully connected neural network to form a Transformer block: S3. Construct a power curve base generation module and combine it with a multi-layer fully connected neural network to form a power curve block. The power curve base generation module uses a smooth logistic growth model to capture the nonlinear characteristics of the wind power curve. Its mathematical expression is: in, Indicates wind speed The predicted power at is the maximum power output, is the steepness parameter that controls the rising rate of the curve, is the half-maximum power wind speed, and are the cut-in wind speed and rated wind speed respectively; The processing method of the power curve base generation module includes: S301: Parameter processing and dimension adjustment, setting and Represent the wind speed data in the lookback window and the forecast window respectively, and concatenate the two input components along the time dimension: in, ; Based on rated cut-out wind speed Formula Adjust the parameters in to scale the horizontal and vertical coordinates of the smoothed logistic growth model to the range [0,1], where the horizontal coordinate represents wind speed and the vertical coordinate represents power: in, , , , , , , ;parameter 、 and the input sequence The dimensions of are adjusted as follows: S302: Construct the power curve basis function and solve the smoothed logistic growth model through element-level broadcasting in tensor calculations: The third dimension and the fourth dimension After merging and removing the singleton dimension, we get: in, ,and ; Calculation formula Nonlinear scaling term in And after expanding its dimension we get: The final basis function matrix is ​​obtained as follows: in, represents element-wise multiplication, and ; Finally, through tensor slicing Split into backtracking basis function and predictive basis function: S303: base expansion coefficient processing, and Apply the ReLU activation function to ensure non-negativity, and then normalize it through softmax: in, Represents normalization along the last dimension of the tensor; the result is then split into the back-basis expansion coefficients and the predicted base expansion coefficient ; S4. Connect several Transformer blocks through the double residual stacking principle to form a Transformer stack: S5, connecting the plurality of power curve blocks through a double residual stacking principle to form a power curve stack; S6: Connect the Transformer stack and the power curve stack in series, establish a dynamic trainable weighting mechanism, fuse the outputs of each stack, and obtain a multi-step prediction result of wind power; S7. Determine the optimal configuration of network hyperparameters using the grid search method based on the prediction results.

2. The method for multi-source data fusion and power prediction of a wind turbine generator system according to claim 1, characterized in that: The Transformer base is combined with a multi-layer fully connected neural network to form a Transformer block. A multi-layer "stack-block" architecture is used to integrate exogenous variables to achieve time series decomposition and prediction. The Transformer block is the basic structural unit of a multi-layer weighted neural network. All Transformer blocks follow the same logic to process input data. Multiple blocks are connected to form a stack structure using the double residual stacking principle. The Transformer block includes a dynamic residual part and a static exogenous variable part. The dynamic residual part is expressed as: in, represents the stack index, Represents the block index, represents the batch size, is the length of the lookback window; the designed multi-layer fully connected neural network outputs two vectors: including the lookback value of the Transformer block and predicted values ,in, is the prediction window length.

3. A wind turbine multi-source data fusion and power prediction method according to claim 2, characterized in that The Transformer block contains a multi-layer fully connected neural network that is responsible for learning the basis expansion coefficients The specific process is as follows: Network Integration Backtracking Base and prediction base ; In the base layer, for each sample , using the backtracking basis and prediction base Perform explicit summation operations and expand the backtracking basis coefficients and the predicted base expansion coefficient Mapping to backtracking value and predicted values , the process is expressed as: 。 4. A wind turbine multi-source data fusion and power prediction method according to claim 3, characterized in that The steps for generating the Transformer base include: S201: Given an input tensor and Splicing along the time dimension ,in , is the lookback window, is the prediction window; following the standard Transformer architecture, the tensor is transposed to And perform linear projection to get : in, is the embedding layer size, scaling factor Used to stabilize the gradient magnitude; the time position information is integrated into the embedded features through the sine encoding function, the position encoding matrix Offline precomputation, where Encodes the length of the maximum position for the precomputed sequence and dynamically truncated to match the sequence length. , the final encoding is expressed as: S202: Apply the Transformer encoder layer for encoding. Each layer contains multi-head self-attention, residual connection and position-by-position feedforward network. The process is expressed as: in, is the lower triangular causal mask, defined as: S203: Output Transformer through the output projection layer From the dimension Mapping to target dimension : By combining permutation and residual connection operations, the fusion is expressed as: 。 5. A wind turbine multi-source data fusion and power prediction method according to claim 1, characterized in that The method of obtaining a stack is: multiple blocks form a stack, the first The input of the stack is the backtrace value , the prediction output of the stack is aggregated from the prediction results of each block within it: 。 6. A wind turbine multi-source data fusion and power prediction method according to claim 1, characterized in that The multi-step wind power prediction result is the weighted sum of all stack prediction results: in, is the trainable weight coefficient.

7. A wind turbine multi-source data fusion and power prediction method according to claim 1, characterized in that The number of the Transformer blocks and the power curve blocks are respectively obtained by grid search optimization using the data set.

Images