CNN-BiGRU and adaptive loss-based wind power prediction method

Abstract

A wind power prediction method based on CNN-BiGRU and adaptive loss comprises the steps that original time sequence data output by a wind power plant SCADA system is acquired, and the original time sequence data at least comprises a wind speed sequence and an active power sequence; the method comprises the following steps: carrying out anomaly identification and grading correction on original time sequence data based on fan physical characteristics, fitting a rated wind speed and an optimal wind energy utilization coefficient by utilizing a wind speed-power theoretical curve, and according to a difference value between actual power and theoretical power, a wind speed threshold value and a power zero value state; according to the method, through an anomaly recognition and grading correction mechanism fusing a wind turbine aerodynamic principle, an optimal wind energy utilization coefficient and a rated wind speed are automatically fitted by utilizing a wind speed and power theoretical curve, and a differential repair strategy is adopted according to an anomaly duration: a wind energy conversion efficiency weighted interpolation is adopted for short-time anomaly, and local dynamic characteristics are reserved; if long-time abnormity occurs, the training samples are eliminated, and pollution to the training samples is avoided.

Description

Technical Field

[0001] This invention belongs to the field of wind power prediction technology, specifically a wind power prediction method based on CNN-BiGRU and adaptive loss. Background Technology

[0002] Wind power forecasting is a crucial foundation for grid dispatching and wind farm operation management. However, wind power data is highly random, volatile, and intermittent, and the raw SCADA (Supervisory Control and Data Acquisition) data often contains a large amount of "dirty data" generated due to sensor failures, power curtailment (wind curtailment), and maintenance shutdowns.

[0003] Traditional forecasting methods often struggle to simultaneously handle data non-stationarity and outliers caused by equipment operating conditions, resulting in insufficient forecast accuracy, especially when wind speed fluctuates drastically or when there are data gaps, leading to larger forecast errors. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, this invention provides a wind power prediction method based on CNN-BiGRU and adaptive loss, so as to at least partially solve the above-mentioned technical problems.

[0005] The technical solution adopted in this invention is as follows: This invention proposes a wind power prediction method based on CNN-BiGRU and adaptive loss, comprising the following steps: S1: Obtain the raw time series data output by the wind farm SCADA system, wherein the raw time series data includes at least the wind speed sequence and the active power sequence; S2: Based on the physical characteristics of the wind turbine, the original time series data is subjected to anomaly identification and hierarchical correction. Specifically, the rated wind speed and the optimal wind energy utilization coefficient are fitted using the wind speed-power theoretical curve. Based on the difference between the actual power and the theoretical power, the wind speed threshold, and the zero power state, power restriction conditions, abnormal shutdown conditions, and sensor anomalies are identified respectively. For short-term anomaly segments with fewer than a preset threshold consecutive anomalies, the wind energy conversion efficiency of the preceding and following effective time periods is used for filling. For long-term anomaly segments with more than the preset threshold consecutive anomalies, they are directly removed. S3: For the continuous time series processed by S2, calculate the first difference of its timestamps. If the interval between adjacent timestamps exceeds the preset time breakpoint threshold, then divide the time series into multiple non-overlapping continuous segments. S4: Perform variational mode decomposition (VMD) independently on each of the continuous segments to obtain K intrinsic mode function (IMF) components, where K is an integer between 5 and 8, and the VMD penalty factor α is set to 2000. S5: Collect all IMF components obtained by VMD decomposition of all segments in the training set, construct a globally unified normalizer, and apply the normalizer to the IMF components corresponding to the validation set and test set to achieve scale consistency across datasets and prevent information leakage. S6: For each IMF component, an independent deep learning prediction sub-model is constructed. The sub-model sequentially includes: a one-dimensional convolutional layer Conv1D, a bidirectional gated recurrent unit BiGRU layer, a multi-head self-attention mechanism module, and a residual connection structure. S7: The combined loss function is used to train each sub-model. The combined loss function consists of a mean squared error (MSE) term and a weighted absolute error term. The weighted absolute error term introduces a low-power interval enhancement coefficient and an asymmetric error penalty weight. S8: After inverse normalization, the prediction results corresponding to each IMF component are linearly superimposed to output the final wind power prediction value.

[0006] In one embodiment of the present invention, in step S3, the time breakpoint threshold is set to 30 minutes, that is, when the difference between the timestamps of any two adjacent valid data points is greater than 1800 seconds, it is determined that there is a time breakpoint, and the original sequence is divided into multiple continuous segments based on this. The timestamps within each segment are continuous and without missing values.

[0007] In one embodiment of the present invention, in step S4, when performing VMD decomposition on each continuous segment, the modality number K=5, the penalty factor α=2000 is preset, and the initial value of the center frequency is fixed as K values ​​with spacing distributed in the interval [0,0.5], so as to ensure that each IMF component has clear separation in the frequency domain.

[0008] In one embodiment of the present invention, in step S5, the fitting parameters of the global normalizer are calculated only based on all IMF components of the training set, and the same scaling transformation is performed on the corresponding IMF components of the validation set and the test set during the inference phase, thereby ensuring that the data distribution is consistent between the training and inference phases and that there is no future information leakage.

[0009] In one embodiment of the present invention, in step S6, the one-dimensional convolutional layer Conv1D has a kernel size of 3, a filter count of 64, and an activation function of ReLU; the BiGRU layer contains two stacked layers with 128 hidden units per layer; the multi-head self-attention mechanism contains four attention heads, each with a dimension of 32; and the residual connection adds the Conv1D output to the BiGRU+Attention output and then feeds it into the fully connected layer.

[0010] In one embodiment of the present invention, the method is deployed on a computer device equipped with a GPU computing unit. The device is configured with a graphics processor with a video memory capacity of not less than 4GB, and uses the Pandas library to perform vectorization and cleaning operations on millions of SCADA data. At the same time, it supports batch training of deep learning models and adaptive video memory growth strategies.

[0011] In one embodiment of the present invention, the method adopts a modular storage structure, in which the global normalizer parameters generated in the data preprocessing stage and the deep learning sub-model weights corresponding to each IMF component are saved as independent files. The normalizer file is used for data standardization before inference, and the model weight file is used to load the corresponding sub-model for prediction. The two are isolated from each other in physical storage path.

[0012] In one embodiment of the present invention, the method is applicable to short-term prediction scenarios of the aggregated power of a single wind turbine or multiple turbines. The input data time resolution is 15 minutes, and the prediction duration is from 4 to 72 hours in the future. The VMD decomposition, IMF modeling, and loss function together constitute an end-to-end trainable prediction pipeline, in which each technical module is executed in series in sequence, and the output of the previous module is used as the input of the next module, forming a deterministic mapping path from the original SCADA data to the final power prediction value.

[0013] The beneficial effects of the technical solution of this invention are as follows: This invention improves the physical rationality and usability of input data by integrating anomaly identification and hierarchical correction mechanisms based on the aerodynamic principles of wind turbines. It automatically fits the optimal wind energy utilization coefficient and rated wind speed using wind speed and power theoretical curves, and uses this as a benchmark to distinguish three typical abnormal states. Differentiated repair strategies are adopted according to the duration of the anomaly: short-term anomalies are treated with wind energy conversion efficiency weighted interpolation to preserve local dynamic characteristics; long-term anomalies are decisively removed to avoid contaminating the training samples.

[0014] This invention introduces a 30-minute time breakpoint threshold for structured segmentation, dividing the original long sequence into multiple internally continuous and mechanism-consistent operational segments. The threshold setting is based on empirical analysis of SCADA sampling frequency and wind turbine operation event duration, effectively distinguishing between transient jitter and substantive shutdown. Furthermore, each segment undergoes independent VMD decomposition with fixed parameters (K=5, α=2000, initial center frequency spacing distribution), ensuring not only clear separation of each IMF component in the frequency domain (high-frequency turbulence, mid-frequency diurnal variation, low-frequency baseline) but also maintaining the consistency of the physical meaning of similar IMFs across different segments.

[0015] This invention achieves dual optimization through a global normalization strategy and a dedicated sub-model architecture. The normalizer fits parameters based solely on the training set IMF and reuses them for the validation / test set during inference, preventing future information leakage and ensuring the stability of cross-time predictions. Meanwhile, configuring an independent CNN-BiGRU-Attention sub-model for each IMF enables "frequency-specific teaching": Conv1D (3 kernels, 64 filters) efficiently extracts nonlinear responses within a local three-step window; a two-layer BiGRU (128 units per layer) models long-range bidirectional temporal dependencies; four-head self-attention (32 dimensions per head) dynamically focuses on key historical moments; and residual connections fuse low-level details with high-level semantics to prevent information loss.

[0016] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0017] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a schematic diagram illustrating the data cleaning effect proposed in an embodiment of the present invention; Figure 2 This is a project flowchart proposed in an embodiment of the present invention; Figure 3 This is a diagram showing the IMF component prediction results proposed in an embodiment of the present invention; Figure 4 This is a prediction comparison chart proposed in an embodiment of the present invention; Figure 5 This is a scatter plot of the predicted values ​​proposed in an embodiment of the present invention. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0019] The following describes an embodiment of the present invention, based on CNN-BiGRU and adaptive loss, with reference to the accompanying drawings.

[0020] like Figures 1 to 5 As shown, this embodiment of the invention provides a wind power prediction method based on CNN-BiGRU and adaptive loss, including the following method steps: S1: Obtain the raw time series data output by the wind farm SCADA system. The raw time series data shall include at least the wind speed series and the active power series. S2: Based on the physical characteristics of the wind turbine, anomaly identification and hierarchical correction are performed on the original time series data. Specifically, the rated wind speed and the optimal wind energy utilization coefficient are fitted using the wind speed-power theoretical curve. Based on the difference between the actual power and the theoretical power, the wind speed threshold, and the zero power state, power restriction conditions, abnormal shutdown conditions, and sensor anomalies are identified respectively. For short-term anomaly segments with fewer than a preset threshold of consecutive anomalies, the wind energy conversion efficiency of the preceding and following effective time periods is used for filling. For long-term anomaly segments with more than a preset threshold of consecutive anomalies, they are directly removed. S3: For the continuous time series processed by S2, calculate the first difference of its timestamps. If the interval between adjacent timestamps exceeds the preset time breakpoint threshold, then divide the time series into multiple non-overlapping continuous segments. S4: Perform variational mode decomposition (VMD) independently on each continuous segment to obtain K intrinsic mode function (IMF) components, where K is an integer between 5 and 8, and the VMD penalty factor α is set to 2000. S5: Collect all IMF components obtained by VMD decomposition of all segments in the training set, construct a globally unified normalizer, and apply the normalizer to the IMF components corresponding to the validation set and test set to achieve scale consistency across datasets and prevent information leakage. S6: For each IMF component, an independent deep learning prediction sub-model is constructed. The sub-model includes, in sequence: a one-dimensional convolutional layer Conv1D, a bidirectional gated recurrent unit BiGRU layer, a multi-head self-attention mechanism module, and a residual connection structure. S7: The combined loss function is used to train each sub-model. The combined loss function consists of the mean squared error (MSE) term and the weighted absolute error term. The weighted absolute error term introduces the low power interval enhancement coefficient and the asymmetric error penalty weight. S8: After inverse normalization, the prediction results corresponding to each IMF component are linearly superimposed to output the final wind power prediction value.

[0021] In practical applications, this invention involves acquiring raw time-series data output from a wind farm monitoring and data acquisition (SCADA) system. This data is typically recorded at fixed time intervals (e.g., 15 minutes) and includes at least two core variables: wind speed and active power. However, due to communication interruptions, sensor failures, human-caused power rationing, or wind turbine protective shutdowns, the raw data often contains numerous outliers, missing segments, and non-physical state points. Directly inputting this "dirty data" into a deep learning model not only introduces noise interference but also misleads the model into learning incorrect mapping relationships. Therefore, the system first performs data cleaning based on the physical characteristics of the wind turbine. Specifically, the system utilizes the basic physical principles of wind energy conversion—that is, within the cut-in wind speed to rated wind speed range, the wind turbine output power should theoretically be proportional to the cube of the wind speed—to fit a theoretical wind speed and power curve that closely matches historical effective operating points. Subsequently, the system compares the deviation between actual power and theoretical power point by point, and, in conjunction with whether the wind speed is within a reasonable range and whether the power is zero but the wind speed is not zero, identifies three typical anomalies: first, power curtailment conditions (actual power is significantly lower than the theoretical value but the wind turbine is still running); second, abnormal shutdown (wind speed is higher than the cut-in wind speed but the power is zero); and third, sensor drift or jumps (power sudden changes that do not conform to physical laws). For the identified abnormal segments, the system further determines their duration: if the number of consecutive abnormal points is less than a preset threshold (e.g., 3 time steps), it is considered a short-term disturbance, and the wind energy conversion efficiency calculated from the effective time periods before and after is used for weighted interpolation to fill in the gaps, preserving local dynamic information; if it exceeds the threshold, it is determined to be an unreliable long period and is directly removed to avoid contaminating the training samples.

[0022] After cleaning, the system performs structured analysis on the continuity of the time series. Since wind power data often has long gaps due to communication interruptions, forcibly splicing discontinuous segments can lead to misjudgments of temporal dependencies in the model. Therefore, the system calculates the first-order difference between the timestamps of adjacent valid data points. Once the interval exceeds a preset time breakpoint threshold (e.g., 30 minutes), the sequence is cut off at that point, generating multiple non-overlapping, internally time-continuous segments. Each segment represents a complete, uninterrupted wind turbine operation record, possessing stable dynamic characteristics, making it more suitable for subsequent modeling.

[0023] Next, the system independently performs Variational Mode Decomposition (VMD) on each continuous segment. VMD is an adaptive signal decomposition method that can decompose complex non-stationary wind power signals into several Intrinsic Mode Function (IMF) components with different center frequencies and bandwidths. In this scheme, the number of modes K is limited to between 5 and 8 (preferably 5), and the penalty factor α is fixed at 2000 to balance the fidelity of the decomposition with the mode separation. Through VMD, the original signal is decoupled into high-frequency fluctuation components (reflecting rapid changes in turbulence and gusts), mid-frequency trend components (corresponding to daily power fluctuations), and low-frequency baseline components (reflecting seasonal or long-term attenuation trends).

[0024] To ensure data distribution consistency and prevent information leakage during model training and inference, the system employs a global normalization strategy. It collects all IMF components obtained from VMD decomposition of all segments in the training set, uniformly calculates their mean and standard deviation (or minimum / maximum), and constructs a global StandardScaler (or MinMaxScaler) normalizer. The normalizer only fits parameters on the training set and is then applied to the corresponding IMF components on the validation and test sets. This ensures that the model is trained and evaluated on the same scale across all datasets, avoids information leakage during the preprocessing stage, and enhances the model's generalization ability.

[0025] During the modeling phase, the system constructs an independent deep learning prediction sub-model for each IMF component. Each sub-model employs a carefully designed hybrid neural network architecture: First, a one-dimensional convolutional layer (Conv1D) is used to extract local temporal patterns. The kernel size is 3, and the number of filters is 64, which can effectively capture the short-term nonlinear response between wind speed and power. Then, two stacked bidirectional gated recurrent units (BiGRU) are connected, with 128 hidden units per layer. The BiGRU structure can simultaneously model the contextual dependencies of the past and future, making it particularly suitable for processing wind power data with bidirectional temporal correlations. To further enhance the model's ability to focus on key time steps, the system introduces a multi-head self-attention mechanism (4 heads, each with a dimension of 32) after the BiGRU output, dynamically allocating the contribution weights of different historical moments to the current prediction. Finally, the original features of Conv1D are added to the high-level representation of BiGRU+Attention through residual connections, alleviating the gradient vanishing problem in deep networks and preserving low-level details.

[0026] During training, the system abandons the traditional single loss function and instead adopts a combined adaptive loss function, where the loss is a linear combination of the mean squared error (MSE) and weighted absolute error (MAE) terms. The weighted MAE term specifically introduces an enhancement coefficient for low-power regions: when the actual power is below 20% of the rated power, the error penalty weight is significantly increased; while in the high-power region, a lower weight is used. This asymmetry stems from the actual distribution characteristics of wind power; although the low-power region has small numerical values, it occurs frequently, and traditional models often systematically underestimate this region. By strengthening the error penalty in the low-power region, the model is forced to fit the complex dynamics of the region (such as the startup process and light wind response), thereby comprehensively improving the overall prediction accuracy, especially enhancing the prediction performance during low-output periods, which is of most concern in day-ahead scheduling.

[0027] Finally, during the inference phase, the system denormalizes the sub-model prediction results corresponding to each IMF component to restore them to the original power dimensions, and then performs linear superposition to obtain the final wind power prediction value. Since each IMF component is accurately predicted by a dedicated model and its physical meaning is clear (high frequency, medium frequency, low frequency), the superposition result is not only highly accurate but also has good interpretability.

[0028] In one specific implementation, in step S3, the time breakpoint threshold is set to 30 minutes. That is, when the difference between the timestamps of any two adjacent valid data points is greater than 1800 seconds, a time breakpoint is determined to exist, and the original sequence is divided into multiple continuous segments based on this. The timestamps within each segment are continuous and without missing values. In step S4, when performing VMD decomposition on each continuous segment, the modality number K=5, the penalty factor α=2000 is preset, and the initial value of the center frequency is fixed as K values ​​with spacing distributed in the interval [0,0.5] to ensure that each IMF component has clear separation in the frequency domain.

[0029] In practical applications, the system first performs time continuity analysis on the continuous and valid data after physical rule cleaning. By calculating the timestamp difference between adjacent valid data points, if an interval exceeding 1800 seconds (i.e., 30 minutes) is found, it is determined that there is a substantial operational interruption or communication loss, constituting a "time breakpoint". The threshold of 30 minutes is an empirical value derived from the typical sampling frequency of wind farm SCADA systems (usually 5–15 minutes) and the actual duration of wind turbine start-up / power curtailment events. Missing data shorter than 30 minutes is merely instantaneous communication jitter, which can be repaired by interpolation; while gaps exceeding 30 minutes highly likely correspond to actual shutdowns, maintenance, or long-term power curtailment states, where the data before and after no longer have a continuous dependency relationship in terms of physical mechanisms and operating modes. Therefore, the system uses this as a boundary to cut the original long sequence into several segments with strictly continuous internal timestamps, no missing data, and representing a complete operating cycle. Each segment is essentially an independent and self-consistent dynamic process record, with a relatively stable mapping relationship between wind speed and power within it, avoiding false time-series correlations caused by splicing across interruptions.

[0030] Based on this, the system independently performs variational mode decomposition (VMD) on each such continuous segment. VMD, as an advanced adaptive signal decomposition method, can decompose complex non-stationary wind power signals into a set of intrinsic mode functions (IMFs) with compact spectral support. Each set of IMFs corresponds to the oscillation component of a specific frequency band in the signal. To ensure the stability, repeatability, and clear separation of modes in the frequency domain of the decomposition results, the number of modes K is fixed at 5. Too few modes (e.g., K=3) cannot adequately decouple high-frequency turbulence disturbances from mid-frequency diurnal variation trends, while too many modes (e.g., K=10) can easily lead to mode aliasing or introduce redundant noise components. The penalty factor α is set to 2000, which is large enough to suppress mode bandwidth overlap without excessive smoothing and loss of key transient features. The system explicitly sets the initial center frequencies of the K modes to 6 points (e.g., 0.04, 0.12, 0.20, 0.28, 0.36, 0.44) spaced within the normalized frequency range [0, 0.5], thereby guiding the optimization algorithm to quickly converge to a solution with clear physical meaning from a reasonable initial guess, effectively avoiding the problems of disordered mode order or spectral leakage caused by random initialization.

[0031] Furthermore, time segmentation ensures that each decomposed sub-signal is physically continuous and mechanistically consistent, avoiding the mixing of different operating states (such as normal power generation, power curtailment, and shutdown recovery) for decomposition. This prevents VMD from mistakenly treating state transitions as high-frequency oscillations and contaminating the modes. On the other hand, fixed-parameter VMD decomposition ensures that the IMF components generated by all sub-segments remain highly consistent in quantity, frequency band, and physical meaning. For example, the first IMF always corresponds to high-frequency fluctuations on the second to minute scale (caused by gusts and mechanical vibrations), the third and fourth IMFs characterize hourly diurnal variation trends, and the last one or two IMFs reflect slow baseline drift on a multi-day scale.

[0032] In one specific implementation, in step S5, the global normalizer fitting parameters are calculated based solely on all IMF components of the training set, and the corresponding IMF components of the validation set and test set are subjected to the same scaling transformation during the inference phase, thereby ensuring that the data distribution is consistent between the training and inference phases and that there is no leakage of future information; in step S6, the one-dimensional convolutional layer Conv1D has a kernel size of 3, a filter count of 64, and an activation function of ReLU; the BiGRU layer contains two stacked layers with 128 hidden units per layer; the multi-head self-attention mechanism contains four attention heads, each with a dimension of 32; the residual connection adds the Conv1D output to the BiGRU+Attention output and then feeds it into the fully connected layer.

[0033] In practical applications of this invention, after completing the variational mode decomposition (VMD) of all continuous segments, the system enters the standardization phase. At this point, the IMF components of each order obtained from the decomposition of all training set segments are uniformly collected and used as the basis for fitting the global normalizer. The system calculates the mean and standard deviation (or minimum and maximum values) of each IMF channel based solely on this batch of training data, and constructs a fixed scaling mapping rule accordingly. This rule will be reused in subsequent validation and testing phases. That is, the corresponding order IMF components in the validation and testing sets, regardless of their original values, are standardized using the same set of parameters determined during the training phase. The model cannot "peek" into the statistical characteristics of the validation or testing data during training, and will not experience scale shifts due to unknown distributions during deployment. Especially for non-stationary signals such as wind power with seasonal and interannual fluctuations, the strategy effectively maintains the stability of input features in cross-period prediction tasks, ensuring that the model's generalization ability does not degrade due to dataset switching.

[0034] Based on this data representation, the system constructs a dedicated deep prediction sub-model for each IMF component. Each sub-model takes a standardized single-channel IMF sequence as input and first feeds it into a one-dimensional convolutional layer (Conv1D). The convolutional layer uses a time window of size 3 to slide and scan the input sequence, combined with 64 learnable filters, which can efficiently capture nonlinear wind speed and power response patterns within a local three-step time window, such as the instantaneous power surge of a wind turbine under a sudden gust of wind or the small oscillations caused by turbulence. The introduction of the ReLU activation function gives the model nonlinear expressive power while avoiding the gradient vanishing problem. The high-dimensional feature map output by the convolution is then passed to two stacked bidirectional gated recurrent units (BiGRU). The key advantage of the BiGRU structure lies in its bidirectional temporal modeling capability: the forward GRU captures the evolutionary trend from the past to the present, while the backward GRU uses future context information to infer the current state. The combination of the two enables the model to understand the relative position and role of power changes at any given time in the entire running segment. The configuration of 128 hidden units per layer ensures sufficient memory capacity while taking into account computational efficiency, which is sufficient to model long-range dependencies that last for hours or even tens of hours.

[0035] To further enhance the model's ability to dynamically focus on key historical moments, the system embeds a multi-head self-attention mechanism on top of the BiGRU output. This mechanism comprises four parallel attention heads, each projecting the input into a 32-dimensional query, key, and value space, independently calculating attention weights, and then concatenating and fusing them. For example, one head focuses on rapid changes in the last hour, another on periodic patterns from the same period yesterday, and a third identifies typical precursor patterns to power rationing events. By weighted aggregation of these multi-perspectives, the model can adaptively highlight the most valuable historical information for current predictions while suppressing irrelevant noise.

[0036] Specifically, the system introduces a residual connection structure in the feature fusion stage: the local feature representation output from the initial Conv1D layer is element-wise added to the high-level semantic representation processed by BiGRU and self-attention, and then fed into the final fully connected output layer. On the one hand, the residual path preserves the local details extracted by the original convolution, preventing the deep network from losing high-frequency transient information during the abstraction process; on the other hand, it alleviates the gradient decay problem during backpropagation, enabling the entire deep architecture of more than 10 layers to still be trained stably. Finally, the fully connected layer maps the fused features to power prediction values ​​for one or more steps.

[0037] In one specific implementation, the method is deployed on a computer device equipped with a GPU computing unit. The device is configured with a graphics processor with a video memory capacity of not less than 4GB and uses the Pandas library to perform vectorization and cleaning operations on millions of SCADA data. It also supports batch training of deep learning models and adaptive video memory growth strategies. The method adopts a modular storage structure, saving the global normalizer parameters generated in the data preprocessing stage and the deep learning sub-model weights corresponding to each IMF component as separate files. The normalizer file is used for data standardization before inference, and the model weight file is used to load the corresponding sub-model for prediction. The two are isolated from each other in physical storage path.

[0038] In a specific application of this invention, the system first utilizes the Pandas library to perform vectorization cleaning on the raw time series data from the wind farm SCADA system. Pandas' high-performance DataFrame structure and broadcasting mechanism enable parallel execution of anomaly identification based on physical rules (such as wind speed and power deviation judgment, zero-power state detection, and power curtailment marking) using columnar operations, avoiding the performance bottlenecks of traditional loop traversal. Even when faced with historical data containing millions of records covering hundreds of wind turbines spanning several years, the vectorization cleaning process can be completed within seconds or minutes, improving data preparation efficiency.

[0039] To support the training and inference of subsequent deep learning models, the system is deployed on dedicated servers or edge computing nodes equipped with GPU computing units. The GPU memory capacity is no less than 4GB, a minimum requirement determined after comprehensive calculation of the number of model parameters, batch size, and input sequence length. Each IMF sub-model contains convolutional layers, a two-layer BiGRU, and a multi-head attention module. The GPU memory required for a single forward propagation increases linearly with the batch size. Insufficient GPU memory will force the use of extremely small batches or even single-sample training, severely slowing down convergence and affecting gradient stability. Therefore, 4GB of GPU memory ensures smooth model training at reasonable batch sizes (e.g., 32 or 64). Simultaneously, the system enables an adaptive GPU memory growth strategy (e.g., memory_growth=True in TensorFlow), meaning that GPU memory is dynamically allocated on demand rather than occupying all available space at once.

[0040] After model training is complete, the system employs a strict modular storage structure to persistently store components. Specifically, the global normalizer (such as a StandardScaler object) generated during the data preprocessing stage is serialized into an independent configuration file (e.g., .pkl or .json format), which fully records the mean, standard deviation, and scaling parameters of each IMF channel. The deep learning sub-model weights corresponding to each IMF component (including Conv1D filters, BiGRU gating parameters, and attention projection matrices) are saved as separate model files (e.g., .h5 or .onnx format). The two types of files are explicitly isolated in the physical storage path. For example, the normalizer is uniformly stored in the / preprocessing / scalers / directory, and the model weights are stored in subdirectories such as / models / imf_1 / , / models / imf_2 / , etc. This decoupling offers several technical advantages: First, during the inference phase, the system can load the corresponding IMF normalizer file to standardize the input data before loading the matching model weight file for prediction, resulting in a clear process with no cross-dependencies. Second, when a component model needs retraining (e.g., due to distribution drift caused by new data), only the corresponding model file needs to be replaced, without needing to regenerate or deploy other IMF models or normalizers, reducing update costs. Third, the structure inherently supports model version management and A / B testing, allowing operations personnel to easily switch between different versions of normalization strategies or network architectures for performance comparison. Finally, since the normalization parameters do not contain trainable variables, their file size is extremely small, facilitating rapid transmission and caching, while the model file can be loaded using GPU acceleration, and the separation of the two avoids I / O blocking during large file reads.

[0041] In one specific implementation, the method is applicable to short-term prediction scenarios of the combined power of a single wind turbine or multiple turbines. The input data has a time resolution of 15 minutes, and the prediction duration is from 4 to 72 hours in the future. VMD decomposition, IMF modeling, and loss function together constitute an end-to-end trainable prediction pipeline, in which each technical module is executed in sequence, and the output of the previous module serves as the input of the next module, forming a deterministic mapping path from the original SCADA data to the final power prediction value.

[0042] In a specific application of this invention, the system first receives the raw time series stream from the SCADA platform. The data continuously records key variables such as wind speed and active power at 15-minute intervals. Regardless of whether the target is single-unit or aggregated power across the entire field, the system employs a unified processing logic: first, it cleanses and removes anomalies through physical rules-driven data cleaning; then, based on a 30-minute time breakpoint threshold, it divides the long sequence into multiple internally continuous operating segments. Each segment is then fed into the Variational Mode Decomposition (VMD) module, where it is decoupled into six intrinsic mode functions (IMFs) with clearly defined frequency bands under fixed parameters (K=5, α=2000). This decomposition is not the end of preprocessing, but rather the starting point of the entire prediction pipeline, transforming the original non-stationary, multi-scale power signal into a set of sub-signals with a clear structure and relatively simple dynamic characteristics, providing a "learnable" input basis for subsequent modeling.

[0043] Next, each IMF component is independently fed into its dedicated deep neural network sub-model. Although the sub-models share the same structure (all consisting of Conv1D, BiGRU, multi-head attention, and residual connections), their parameters are trained completely independently, each focusing on capturing the temporal evolution patterns within its corresponding frequency band: the high-frequency IMF sub-model learns turbulent disturbance responses at the second to minute level, the mid-frequency model captures daily periodic rises and falls, and the low-frequency model models baseline drift or seasonal decay at multi-day scales. All sub-models share the same set of normalized parameters globally fitted from the training set, ensuring consistent input scales and no leakage of future information. During the training phase, each sub-model is optimized synchronously, with its loss function employing a combination of MSE and weighted MAE, particularly strengthening the error penalty for low-power areas (such as light winds at night or the initial stage of power curtailment recovery), thereby guiding the model to improve prediction accuracy during the low-output periods most sensitive to the power grid.

[0044] After training is complete, the system enters inference mode. At this point, the new 15-minute resolution SCADA data stream undergoes the same cleaning, segmentation, and VMD processes to generate IMF sequences to be predicted. Each IMF sequence is first standardized using a pre-saved normalizer file, then the corresponding sub-model weight file is loaded for forward computation, outputting the modal power prediction values ​​for the next 4 to 72 hours (i.e., 16 to 288 15-minute steps). All IMF predictions are linearly superimposed after inverse normalization to form the final wind power prediction curve. Due to the completeness and orthogonality (approximately) of VMD decomposition, the superposition result theoretically reconstructs the predicted values ​​of the original signal rigorously, ensuring physical consistency.

[0045] Furthermore, when the forecast target is ultra-short-term (e.g., 4 hours), the system can focus on detailed modeling of high-frequency IMFs; when the target is medium-term (e.g., 72 hours), it relies more on the trend capture capability of medium- and low-frequency IMFs. The 15-minute input resolution and multi-step output enable the system to meet the 15-minute rolling update requirements of AGC (Automatic Generation Control) and support the preparation of day-ahead scheduling plans. Whether it's single-unit performance evaluation or overall power output declaration, the pipeline can operate efficiently within a unified framework, truly achieving the engineering goal of "one system, multiple scenarios for reuse."

[0046] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0047] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A wind power prediction method based on CNN-BiGRU and adaptive loss, characterized in that, The following steps are included: S1: Obtain the raw time series data output by the wind farm SCADA system, wherein the raw time series data includes at least the wind speed sequence and the active power sequence; S2: Based on the physical characteristics of the wind turbine, the original time series data is subjected to anomaly identification and hierarchical correction. Specifically, the rated wind speed and the optimal wind energy utilization coefficient are fitted using the wind speed-power theoretical curve. Based on the difference between the actual power and the theoretical power, the wind speed threshold, and the zero power state, power restriction conditions, abnormal shutdown conditions, and sensor anomalies are identified respectively. For short-term anomaly segments with fewer than a preset threshold consecutive anomalies, the wind energy conversion efficiency of the preceding and following effective time periods is used for filling. For long-term anomaly segments with more than the preset threshold consecutive anomalies, they are directly removed. S3: For the continuous time series processed by S2, calculate the first difference of its timestamps. If the interval between adjacent timestamps exceeds the preset time breakpoint threshold, then divide the time series into multiple non-overlapping continuous segments. S4: Perform variational mode decomposition (VMD) independently on each of the continuous segments to obtain K intrinsic mode function (IMF) components, where K is an integer between 5 and 8, and the VMD penalty factor α is set to 2000. S5: Collect all IMF components obtained by VMD decomposition of all segments in the training set, construct a globally unified normalizer, and apply the normalizer to the IMF components corresponding to the validation set and test set to achieve scale consistency across datasets and prevent information leakage. S6: For each IMF component, an independent deep learning prediction sub-model is constructed. The sub-model sequentially includes: a one-dimensional convolutional layer Conv1D, a bidirectional gated recurrent unit BiGRU layer, a multi-head self-attention mechanism module, and a residual connection structure. S7: The combined loss function is used to train each sub-model. The combined loss function consists of a mean squared error (MSE) term and a weighted absolute error term. The weighted absolute error term introduces a low-power interval enhancement coefficient and an asymmetric error penalty weight. S8: After inverse normalization, the prediction results corresponding to each IMF component are linearly superimposed to output the final wind power prediction value.

2. The wind power prediction method based on CNN-BiGRU and adaptive loss according to claim 1, characterized in that, In step S3, the time breakpoint threshold is set to 30 minutes. That is, when the difference between the timestamps of any two adjacent valid data points is greater than 1800 seconds, a time breakpoint is determined to exist, and the original sequence is divided into multiple continuous segments based on this threshold. The timestamps within each segment are continuous and without any missing ones.

3. The wind power prediction method based on CNN-BiGRU and adaptive loss according to claim 1, characterized in that, In step S4, when performing VMD decomposition on each continuous segment, the number of modes K=5, the penalty factor α=2000 is preset, and the initial value of the center frequency is fixed as K values ​​with spacing distributed in the interval [0,0.5], to ensure that each IMF component has clear separation in the frequency domain.

4. The wind power prediction method based on CNN-BiGRU and adaptive loss according to claim 1, characterized in that, In step S5, the fitting parameters of the global normalizer are calculated based only on all IMF components of the training set, and the same scaling transformation is performed on the corresponding IMF components of the validation set and the test set during the inference phase, thereby ensuring that the data distribution is consistent between the training and inference phases and that there is no leakage of future information.

5. The wind power prediction method based on CNN-BiGRU and adaptive loss according to claim 1, characterized in that, In step S6, the one-dimensional convolutional layer Conv1D has a kernel size of 3, a filter count of 64, and an activation function of ReLU; the BiGRU layer contains two stacked layers with 128 hidden units per layer; the multi-head self-attention mechanism contains four attention heads, each with a dimension of 32; and the residual connection adds the Conv1D output to the BiGRU+Attention output and then feeds it into the fully connected layer.

6. The wind power prediction method based on CNN-BiGRU and adaptive loss according to claim 1, characterized in that, The method is deployed on a computer device equipped with a GPU computing unit. The device is equipped with a graphics processor with a video memory capacity of not less than 4GB and uses the Pandas library to perform vectorization and cleaning operations on millions of SCADA data. It also supports batch training of deep learning models and adaptive video memory growth strategies.

7. The wind power prediction method based on CNN-BiGRU and adaptive loss according to claim 1, characterized in that, The method adopts a modular storage structure, which saves the global normalizer parameters generated in the data preprocessing stage and the deep learning sub-model weights corresponding to each IMF component as separate files. The normalizer file is used for data standardization before inference, and the model weight file is used to load the corresponding sub-model for prediction. The two are isolated from each other in physical storage path.

8. The wind power prediction method based on CNN-BiGRU and adaptive loss according to claim 1, characterized in that, The method is applicable to short-term prediction scenarios of power aggregation of a single wind turbine or multiple turbines. The input data time resolution is 15 minutes, and the prediction duration is from 4 to 72 hours in the future. The VMD decomposition, IMF modeling and loss function together constitute an end-to-end trainable prediction pipeline, in which each technical module is executed in series in sequence, and the output of the previous module is used as the input of the next module, forming a deterministic mapping path from the original SCADA data to the final power prediction value.

Images