Wind turbine generator load real-time prediction method and system based on PLC

Abstract

The embodiment of the invention provides a wind turbine generator load real-time prediction method and system based on a PLC. According to the wind turbine generator load real-time prediction method and system based on the PLC, original and mixed SCADA data are converted into high-quality feature vectors through a feature engineering module. On the basis, a decoupling architecture of a basic model and a residual model is adopted: the basic model subjected to quantitative compression efficiently runs on a PLC (Programmable Logic Controller) and provides load reference prediction; and an extremely lightweight residual error model is specially used for learning and compensating the prediction error of the residual error model, so that the precision is improved with small calculation overhead. And finally, in order to cope with model performance drift caused by working condition changes, a set of low-cost drift monitoring and self-adaptive updating closed loop is introduced, errors are monitored and predicted in real time, and only the lightweight residual error model is subjected to rapid online fine adjustment, so that high precision and accuracy of the model in long-term operation are ensured.

Description

TECHNICAL FIELD

[0001] The embodiment of the application relates to the technical field of intelligent regulation, and particularly relates to a wind turbine generator set load real-time prediction method and system based on a PLC. BACKGROUND

[0002] Wind energy, as a clean and renewable energy source, plays an increasingly important role in the transformation of global energy structure. Wind turbine generators are the core equipment for converting wind energy. They are operated in complex and changeable natural environments for a long time, and key load-bearing components such as blades, gearboxes and towers continuously bear random dynamic loads caused by turbulence, wind shear and the like, which directly relates to the fatigue life and operation reliability of the generators. Therefore, real-time and accurate prediction of the key loads of the wind turbine generators is of great significance for realizing intelligent load control of the generators, optimizing operation and maintenance strategies, prolonging service life and reducing the cost of electricity. Deploying such a prediction capability on a main controller (PLC) of the wind turbine generator is the most direct and efficient path to realize seamless docking of the prediction results with the control system of the generator and achieve a closed loop of active load reduction control.

[0003] However, in the prior art, most of the load prediction methods based on supervisory control and data acquisition (SCADA) system data rely on complex deep learning models (such as recurrent neural networks, long short-term memory networks, etc.) to pursue high prediction accuracy. These models are usually trained and reasoned on cloud servers or high-performance workstations, and their huge calculation amount and memory occupation form a sharp contradiction with the limited CPU performance and storage resources of the PLC. Directly transplanting such complex models to the PLC will result in too high reasoning delay, which cannot meet the real-time requirement, and even cannot run at all. Therefore, the existing load prediction models based on SCADA data cannot be deployed and run on the resource-limited PLC in real time, accurately,

[0004] Therefore, it is expected to provide a real-time and accurate wind turbine generator load real-time prediction scheme.

[0005] Therefore, it is expected to provide a real-time and accurate wind turbine generator load real-time prediction scheme. SUMMARY

[0006] The embodiment of the application aims to at least solve one of the technical problems in the prior art, and provides a wind turbine generator load real-time prediction method and system based on a PLC.

[0007] In a first aspect, the embodiment of the application provides a wind turbine generator load real-time prediction method based on a PLC, comprising:

[0008] performing feature engineering on the obtained original SCADA data to obtain a SCADA feature vector;

[0009] performing a base model based load benchmarking prediction on the SCADA feature vector to obtain a base load prediction value;

[0010] performing a residual model based adaptive residual prediction on the SCADA feature vector to obtain a prediction error correction value;

[0011] adding the prediction error correction value to the base load prediction value to obtain a final load prediction value;

[0012] performing drift monitoring and residual model adaptive updating based on a prediction error between the final load prediction value and a true load value to obtain an updated residual model.

[0013] In some embodiments, performing feature engineering on the obtained raw SCADA data to obtain the SCADA feature vector comprises:

[0014] performing a sliding time window based data alignment and null value detection on the raw SCADA data to obtain an aligned data record;

[0015] performing a history valid value based data cleaning and filling on the aligned data record to obtain a clean data record;

[0016] performing a pre-stored parameter based numerical normalization and vector assembly on the clean data record to obtain the SCADA feature vector.

[0017] In some embodiments, the base model is a quantized GRU model and the residual model is a small linear regression model.

[0018] In some embodiments, performing a residual model based adaptive residual prediction on the SCADA feature vector to obtain a prediction error correction value comprises:

[0019] performing feature refining on the SCADA feature vector to obtain an enhanced SCADA feature vector;

[0020] inputting the enhanced SCADA feature vector into the residual model to obtain the prediction error correction value.

[0021] In some embodiments, performing feature refining on the SCADA feature vector to obtain an enhanced SCADA feature vector comprises:

[0022] performing structured feature masking on the SCADA feature vector to obtain a pressure time-frequency feature masking optimized vector;

[0023] calculating a feature optimization gain of the pressure time-frequency feature masking optimized vector with respect to the SCADA feature vector;

[0024] In response to the feature optimization gain satisfying a preset condition, the pressure time-frequency feature mask optimization vector is set as the enhanced SCADA feature vector.

[0025] In some embodiments, drift monitoring and residual error model adaptive updating are performed based on a prediction error between the final load prediction value and the actual load value to obtain an updated residual error model, including:

[0026] A difference between the final load prediction value and the actual load value is calculated to obtain a prediction error;

[0027] The prediction error is stored in an error history buffer, and the error history buffer includes N latest prediction errors;

[0028] Statistical analysis is performed on the error history buffer to obtain a time sequence of error moving average values;

[0029] In response to an absolute value of the error moving average value exceeding a preset threshold value for M consecutive times, it is determined that the basic model has a significant drift and an updating process of the residual error model is triggered;

[0030] A recent feature-target pair is extracted, and parameters of the residual error model are updated based on the recent feature-target pair through a stochastic gradient descent.

[0031] In some embodiments, the feature in the feature-target pair is a SCADA feature vector, and the target is a prediction error.

[0032] In a second aspect, an embodiment of the present application provides a wind turbine load real-time prediction system based on a PLC, including:

[0033] A feature vector acquisition module is configured to perform feature engineering on acquired original SCADA data to obtain a SCADA feature vector;

[0034] A load benchmark prediction module is configured to perform a load benchmark prediction based on a basic model on the SCADA feature vector to obtain a basic load prediction value;

[0035] An adaptive residual prediction module is configured to perform an adaptive residual prediction based on a residual error model on the SCADA feature vector to obtain a prediction error correction value;

[0036] A final load prediction module is configured to add the prediction error correction value and the basic load prediction value to obtain a final load prediction value;

[0037] A residual error model generation module is configured to perform drift monitoring and residual error model adaptive updating based on a prediction error between the final load prediction value and the actual load value to obtain an updated residual error model.

[0038] In some embodiments, the feature vector acquisition module is further configured to:

[0039] Data alignment and null value detection based on a sliding time window are performed on the raw SCADA data to obtain aligned data records;

[0040] The aligned data records are cleaned and filled based on historical valid values ​​to obtain clean data records;

[0041] The clean data records are subjected to numerical normalization and vector assembly based on pre-stored parameters to obtain the SCADA feature vector.

[0042] In some embodiments, the residual model generation module is further configured to:

[0043] The difference between the final predicted load value and the actual load value is calculated to obtain the prediction error;

[0044] The prediction error is stored in an error history buffer, which includes the most recent N prediction errors.

[0045] Statistical analysis was performed on the error history buffer to obtain the time series of the error moving average;

[0046] If the absolute value of the moving average error exceeds a preset threshold M times consecutively, it is determined that the base model has drifted significantly and the update process of the residual model is triggered.

[0047] Extract recent feature-target pairs and update the parameters of the residual model based on the recent feature-target pairs using stochastic gradient descent.

[0048] Compared with existing technologies, this invention provides a PLC-based real-time load prediction method and system for wind turbines. Through a pre-processed feature engineering module, it transforms raw, mixed SCADA data into high-quality feature vectors, ensuring prediction reliability from the source. Building upon this, it innovatively employs a decoupled architecture between the base model and the residual model: a quantized and compressed base model runs efficiently on the PLC, providing a baseline load prediction; an extremely lightweight residual model specifically learns and compensates for the former's prediction errors, achieving improved accuracy with minimal computational overhead. Finally, to address model performance drift caused by changes in operating conditions, a low-cost drift monitoring and adaptive update closed loop is introduced. By monitoring prediction errors in real time, only the lightweight residual model is rapidly fine-tuned online, ensuring high precision and accuracy of the model during long-term operation. Attached Figure Description

[0049] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0050] Figure 1 A flowchart of a PLC-based real-time load prediction method for wind turbines according to an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of the data flow in the PLC-based real-time load prediction method for wind turbines according to an embodiment of the present invention.

[0052] Figure 3 This is a flowchart illustrating the method for real-time load prediction of wind turbine units based on PLC according to an embodiment of the present invention, which performs adaptive residual prediction of SCADA feature vectors based on a residual model to obtain prediction error correction values.

[0053] Figure 4 This is a flowchart illustrating the process of performing feature refinement on SCADA feature vectors to obtain enhanced SCADA feature vectors in a PLC-based real-time wind turbine load prediction method according to an embodiment of the present invention.

[0054] Figure 5 This is a block diagram of a PLC-based real-time load prediction system for wind turbines according to an embodiment of the present invention. Detailed Implementation

[0055] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0056] Unless otherwise specifically stated, the technical or scientific terms used in the embodiments of this invention should be understood in their ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms "comprising" or "including," as used in the embodiments of this invention, do not limit the shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof mentioned, nor do they exclude the appearance or addition of one or more other different shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof, or the inclusion of these.

[0057] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale, and techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail; however, where appropriate, the illustrated techniques, methods, and apparatus should be considered part of the specification. In all the examples shown and discussed herein, any other specific example may have different values. It should be noted that similar symbols and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0058] In the description of the embodiments of the present invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In the embodiments of the present invention, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of the present invention, as well as the features of different embodiments or examples.

[0059] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0060] In existing technologies, load prediction methods based on wind turbine monitoring and data acquisition (SCADA) system data mostly rely on complex deep learning models (such as recurrent neural networks and long short-term memory networks) to pursue high prediction accuracy. These models are trained and inferred on cloud servers or high-performance workstations, and their huge computational and memory consumption creates a sharp contradiction with the limited CPU performance and storage resources of PLCs. Directly porting such complex models to PLCs will result in excessively high inference latency, failing to meet real-time requirements, or even making them impossible to run. Therefore, existing load prediction models based on SCADA data are difficult to deploy and run in real-time, accurately, and adaptively on resource-constrained PLCs, which constitutes a major bottleneck in current technological development.

[0061] To address the aforementioned technical problems, this invention proposes a PLC-based real-time load prediction scheme for wind turbines. This solves the technical challenge of achieving real-time, accurate, and adaptive operation of load prediction models on resource-constrained PLCs, a problem inherent in existing technologies. Specifically, this scheme proposes a comprehensive solution integrating feature engineering, a hybrid model, and adaptive updates through a novel prediction method. More specifically, deep preprocessing of the raw SCADA data, including sliding time window alignment, data cleaning and filling, and normalization, transforms the mixed raw data stream into high-quality, standardized SCADA feature vectors, laying a solid foundation for subsequent accurate prediction. Subsequently, these feature vectors are fed into an innovative hybrid prediction architecture: first, a quantized and compressed GRU base model rapidly calculates the feature vectors to obtain the base load prediction value, solving the core computational efficiency problem; simultaneously, an extremely lightweight small linear regression model, acting as a residual model, processes the same feature vectors, aiming to accurately predict the potential errors of the base model. By adding the base load prediction value to the prediction error correction value, a final load prediction value that combines high efficiency and high accuracy can be obtained. More importantly, to address the performance drift that inevitably occurs during long-term model operation, a low-cost adaptive update mechanism is designed. This mechanism monitors the error between the final prediction value and the actual load in real time. Once statistical analysis reveals that the error consistently exceeds a threshold, model drift is determined, and an update process is automatically triggered. This update process cleverly bypasses retraining of the complex base model. Instead, it utilizes recently accumulated feature-error data pairs and employs efficient algorithms such as stochastic gradient descent to rapidly fine-tune the parameters of only the lightweight residual model. This allows the entire prediction system to quickly adapt to new operating conditions, ensuring continuous prediction accuracy.

[0062] The present invention proposes a real-time load prediction method for wind turbine units based on PLC. Figure 1 This is a flowchart of a PLC-based real-time load prediction method for wind turbines according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the data flow in a PLC-based real-time load prediction method for wind turbines according to an embodiment of the present invention. Figure 1 and Figure 2As shown, the real-time load prediction method for wind turbines based on PLC according to an embodiment of the present invention includes the following steps: S100, performing feature engineering on the acquired raw SCADA data to obtain SCADA feature vectors; S200, performing load baseline prediction based on a base model on the SCADA feature vectors to obtain base load prediction values; S300, performing adaptive residual prediction based on a residual model on the SCADA feature vectors to obtain prediction error correction values; S400, adding the prediction error correction values ​​to the base load prediction values ​​to obtain the final load prediction values; S500, performing drift monitoring and adaptive updating of the residual model based on the prediction error between the final load prediction values ​​and the actual load values ​​to obtain the updated residual model.

[0063] Specifically, in step S100, feature engineering is performed on the acquired raw SCADA data to obtain SCADA feature vectors. It should be understood that in actual operation, the physical quantity data (such as wind speed, rotational speed, and power) collected by the SCADA system of wind turbines often suffers from inconsistent sampling frequencies and difficulty in strictly aligning timestamps. Furthermore, sensor failures or communication interruptions inevitably introduce null values ​​and abnormal noise. If these raw, mixed, and defective data streams are directly used for model training and prediction, they will severely erode the accuracy and stability of the model. Therefore, in the technical solution of this invention, feature engineering is performed on the acquired raw SCADA data. This is achieved through data alignment based on sliding time windows, data cleaning and filling based on historical valid values, and numerical normalization and vector assembly based on pre-stored parameters. This systematically transforms the raw data into SCADA feature vectors with a unified format, stable values, and complete information. This eliminates the interference of noise and uncertainty on subsequent model predictions from the data source, providing a high-quality and highly reliable input for the subsequent basic and residual models.

[0064] More specifically, in this embodiment of the invention, feature engineering is performed on the acquired raw SCADA data to obtain SCADA feature vectors, including: performing data alignment and null value detection on the raw SCADA data based on a sliding time window to obtain aligned data records; performing data cleaning and filling on the aligned data records based on historical valid values ​​to obtain clean data records; and performing numerical normalization and vector assembly on the clean data records based on pre-stored parameters to obtain the SCADA feature vectors.

[0065] Specifically, the raw SCADA data undergoes data alignment and null value detection based on a sliding time window to obtain aligned data records. It should be understood that in the actual operation of wind turbine units, their SCADA systems integrate data streams from different sensors (such as anemometers, tachometers, power meters, temperature sensors, etc.). These sensors have inherent differences in sampling frequencies, and coupled with uncertainties such as network transmission delays and packet loss, the acquired raw data is discrete and asynchronous in the time dimension. Directly using this inconsistent timestamp data as model input is tantamount to describing the instantaneous state of the unit with a distorted snapshot of time, inevitably leading to the model learning incorrect causal relationships. Therefore, in the technical solution of this invention, the raw SCADA data undergoes data alignment and null value detection based on a sliding time window to forcibly normalize discrete data points from different sources and frequencies onto a unified and continuous time reference, and in this process, accurately identify data missing points caused by various reasons. This generates a series of aligned data records with strict consistency in the time dimension, providing a synchronous, complete, and structured data foundation for subsequent data cleaning and feature construction.

[0066] More specifically, in a concrete example of the present invention, the system first sets a sliding time window with a predetermined duration (e.g., 10 seconds) and a fixed step size (e.g., 1 second). This time window slides continuously forward along the time axis at the set step size. At each step position, the system captures and collects all raw SCADA data points whose timestamps fall within the current time window interval. For multiple data points of the same physical quantity (e.g., wind speed) existing within a single time window, the system employs a preset aggregation strategy, such as calculating their arithmetic mean or median, to generate a single value that represents the central trend of the physical quantity within that time window. If a specific physical quantity has no data record within the current time window, the field corresponding to that physical quantity will be explicitly marked as null in the final aligned data record generated for that time window. The values ​​or null value markings obtained from the aggregation calculation of all physical quantities together constitute a multi-dimensional but time-synchronized aligned data record associated with the end timestamp of the current time window. This process is repeated cyclically as the time window moves forward, thereby continuously generating a time-aligned stream of data records that can be used by subsequent processing steps.

[0067] Specifically, the aligned data records are cleaned and filled with data based on historical valid values ​​to obtain clean data records. It should be understood that although time-aligned data records solve the timing synchronization problem, null values ​​(i.e., missing data) caused by sensor momentary failures, communication packet loss, etc., still exist. Simultaneously, the sensors themselves may produce outliers deviating from the normal physical range due to interference or malfunction. These data flaws are like breakpoints and noise in the information chain; if left untreated, they will directly lead to errors in subsequent mathematical model calculations or cause the model to learn incorrect operating condition-load mapping relationships, thereby severely reducing prediction accuracy. Therefore, in the technical solution of this invention, the aligned data records are further cleaned and filled with data based on historical valid values. This utilizes data confirmed as valid and conforming to physical laws over a past period to interpolate null values ​​in the current record and identify and correct potential outliers. This generates a clean data record that is numerically complete and logically self-consistent, ensuring that the data stream input to the subsequent prediction model is continuous and of high quality.

[0068] More specifically, in a concrete example of the present invention, the system examines each physical quantity field in each input aligned data record. For detected null fields, the system accesses a buffer storing the N most recent valid historical values ​​of that physical quantity. By performing a preset interpolation algorithm, such as linear interpolation or filling based on historical mean, on the data in the buffer, a reasonable replacement value is calculated and filled into the null position. After completing the null filling, the system enters the data cleaning stage. It iterates through each value in the record again and compares it with the preset valid physical range of that physical quantity (e.g., generator speed range of 0-2000 rpm). Any value outside this range is identified as an outlier. For identified outliers, the system uses a similar method to handling nulls, i.e., it calculates a correction value based on the data in its historical valid value buffer to replace the outlier data. Through this series of rigorous filling and cleaning operations, the original aligned data record is transformed into a clean data record without nulls or obvious outliers, providing a reliable data source for subsequent feature vector assembly.

[0069] Specifically, the clean data records are subjected to numerical normalization and vector assembly based on pre-stored parameters to obtain the SCADA feature vector. It should be understood that in the clean data records after cleaning and filling, each physical quantity (such as wind speed in m / s, rotational speed in rpm, and temperature in °C) still retains its original physical units and dimensions, resulting in significant differences in their numerical scales. This characteristic of inconsistent dimensions and vastly different numerical ranges will cause features with larger values ​​to disproportionately dominate the model's weight updates during subsequent training and inference of gradient descent-based optimization models (such as GRU models), thereby affecting the model's convergence speed and final prediction accuracy. Therefore, in the technical solution of this invention, the clean data records are further subjected to numerical normalization and vector assembly based on pre-stored parameters to eliminate the influence of different dimensions between features, mapping all feature values ​​to a unified, smaller numerical range, and constructing a standardized mathematical structure that meets the model input requirements. This ensures that the model treats all features equally when processing data, accelerates model convergence and improves its generalization ability, and ultimately generates a formatted SCADA feature vector that can be directly used by subsequent prediction models.

[0070] More specifically, in a concrete example of the present invention, firstly, for each physical quantity in the clean data record, the system reads the pre-stored normalized parameters for that physical quantity from the PLC memory, namely, its maximum and minimum values ​​statistically obtained from the historical training dataset. Then, the system applies a minimum-maximum normalization method, scaling the current value of the physical quantity to the [0,1] interval using the formula: Normalized value = (Current value - Minimum value) / (Maximum value - Minimum value). This normalization operation is applied sequentially to all physical quantities in the clean data record. After all values ​​have been normalized, the system arranges these normalized values ​​in a predefined, fixed order, forming a one-dimensional, fixed-length numerical array. For example, the first element of this array is fixed as the normalized wind speed, the second as the normalized generator speed, and so on. This final array, containing all the ordered normalized values, is the SCADA feature vector mentioned above, which can be directly used as input to the subsequent base model and residual model.

[0071] Specifically, in step S200, load baseline prediction based on a base model is performed on the SCADA feature vector to obtain a base load prediction value. It is worth noting that the base model is a quantized GRU model. It should be understood that the load response of a wind turbine is a complex, nonlinear time-series process, deeply dynamically coupled with multiple operating parameters such as wind speed, pitch angle, and rotational speed. A simple linear model alone cannot accurately capture this inherent pattern. Therefore, in the technical solution of this invention, load baseline prediction based on a base model is further performed on the SCADA feature vector after feature engineering. This base model uses a gated recurrent unit network pre-trained on a high-performance server and quantized and compressed to efficiently parse the complex time-series dependencies and nonlinear characteristics contained in the SCADA feature vector under the limited computing resources of the PLC. This allows for the rapid generation of a base load prediction value that reflects the main load change trends, providing a high-quality prediction benchmark for subsequent accuracy correction using a residual model.

[0072] More specifically, in a concrete example of this invention, firstly, the network structure and weight parameters of a quantized GRU base model are pre-stored in the PLC's program memory. When the upstream feature engineering module generates a new SCADA feature vector, this vector is fed as input data into the input layer of this GRU model. The model then performs a forward propagation calculation: the input vector passes through the GRU layers sequentially. At each time step, the model's update and reset gates determine how to update and transmit timing information based on the current input and the hidden state of the previous time step, thereby effectively capturing the dynamic characteristics of load changes. After passing through all GRU layers, the final hidden state is fed into a fully connected output layer, which maps the high-dimensional feature information into a single scalar output. This output value is within a normalized numerical range (e.g., [0,1]). Finally, the system calls the pre-stored inverse normalization parameters for this load physical quantity, i.e., its original maximum and minimum values, to perform an inverse transformation on the scalar output, restoring it to a load unit value with actual physical meaning. This final value is the aforementioned base load prediction value.

[0073] Specifically, in step S300, adaptive residual prediction based on a residual model is performed on the SCADA feature vector to obtain a prediction error correction value. It is worth noting that the residual model is a small linear regression model. It should be understood that, due to the quantization and compression of the basic model to accommodate the limited computing resources of the PLC, while capturing the main load trends, it inevitably introduces systematic prediction biases. These biases limit the upper limit of the final prediction accuracy, and the basic model itself has a complex structure, making it difficult to adjust in real time on the PLC to eliminate these biases. Therefore, in the technical solution of this invention, adaptive residual prediction based on a residual model is further performed on the SCADA feature vector. First, the original SCADA feature vector is screened and enhanced using feature refinement technology to extract the core feature subset most relevant to the prediction error, forming an enhanced SCADA feature vector. Then, this enhanced vector is input into an extremely lightweight small linear regression model, i.e., the residual model, to specifically learn and predict the prediction error of the basic model under the current operating conditions in real time. In this way, an accurate prediction error correction value can be generated. When this correction value is added to the basic load prediction value, the inherent deviation of the basic model can be effectively compensated without increasing the PLC calculation load, thereby improving the accuracy and reliability of the final load prediction value.

[0074] Figure 3 This is a flowchart illustrating the adaptive residual prediction of SCADA feature vectors based on a residual model to obtain prediction error correction values, according to an embodiment of the PLC-based real-time load prediction method for wind turbines based on an embodiment of the present invention. Figure 3 As shown, step S300 includes: S310, performing feature refinement on the SCADA feature vector to obtain an enhanced SCADA feature vector; S320, inputting the enhanced SCADA feature vector into the residual model to obtain the prediction error correction value.

[0075] Specifically, in step S310, feature refinement is performed on the SCADA feature vector to obtain an enhanced SCADA feature vector. It should be understood that although the SCADA feature vector fed into the residual model has undergone standardization, as a general description of the overall operating state of the wind turbine, it contains redundant or even misleading information for the specific sub-task of predicting the error of the basic model. These irrelevant features can interfere with the learning process of the lightweight residual model, limiting its upper limit of correction capability. Therefore, in the technical solution of this invention, feature refinement is further performed on the SCADA feature vector to obtain an enhanced SCADA feature vector. The traditional feature refinement idea is internalized into a dynamic feature purification module. This module performs structured perturbation on the original feature vector (such as generating pressure time-frequency feature masking optimization vectors) and calculates in real time the substantial gain of this refinement operation on improving the accuracy of the downstream residual prediction task. Then, based on whether the gain meets preset conditions, an intelligent decision is made: whether to adopt the refined feature vector or retain the original vector. This ensures that the final input to the residual model is an enhanced SCADA feature vector that has been adaptively refined, has higher information density, and is more correlated with the prediction error. This allows the simple residual model to capture and correct the prediction bias of the base model more efficiently and accurately.

[0076] Figure 4 This is a flowchart illustrating the process of performing feature refinement on SCADA feature vectors to obtain enhanced SCADA feature vectors in a PLC-based real-time wind turbine load prediction method according to an embodiment of the present invention. Figure 4 As shown, step S310 includes: S311, performing structured feature masking on the SCADA feature vector to obtain a pressure time-frequency feature masking optimization vector; S312, calculating the feature optimization gain of the pressure time-frequency feature masking optimization vector relative to the SCADA feature vector; S313, in response to the feature optimization gain satisfying a preset condition, setting the pressure time-frequency feature masking optimization vector as the enhanced SCADA feature vector.

[0077] Specifically, in step S311, structured feature masking is applied to the SCADA feature vector to obtain an optimized pressure time-frequency feature masking vector, expressed by the formula:

[0078] v ab =v⊙m+P(v⊙(1-m))

[0079] P(x)=W2σ(W1x+b)

[0080] Where v is the SCADA feature vector, m∈{0,1} is the learnable sparse mask weight, ⊙ is the positional dot product, P(x) is the deep feature extractor based on convolutional neural network, x is used to indicate v⊙(1-m), σ is the sigmoid function, W1 and W2 are the first and second trainable weight matrices respectively, b is the trainable bias vector, v ab Optimize the vector for masking the time-frequency features of pressure.

[0081] It is understandable that, since the initial SCADA feature vector contains complete pressure time-frequency spectrum information, and this spectrum data is high-dimensional and complex, it inevitably contains some noise regions or redundant frequency bands that are not helpful or even harmful to the downstream residual prediction task. If this complete spectrum information is input into the residual model without discrimination, these irrelevant features will interfere with the model's learning of the core error patterns, thereby limiting its correction accuracy. Therefore, in the technical solution of this invention, the SCADA feature vector is further subjected to structured feature masking to obtain a pressure time-frequency feature masking optimization vector. This is done by simulating visual masking to systematically and locally remove information from a small area of ​​the pressure time-frequency spectrum, thereby generating a hypothetical new feature vector that has eliminated specific potential noise sources. In this way, the macroscopic problem of feature optimization can be decomposed into a specific proposition about whether removing specific small feature regions is beneficial and can be quantitatively evaluated, creating the necessary conditions for subsequent experimental comparison to accurately identify and eliminate interference information.

[0082] Specifically, in step S312, the feature optimization gain of the pressure time-frequency feature masking optimization vector relative to the SCADA feature vector is calculated, expressed by the formula:

[0083] G(v,v ab =JS(Softmax(v ab )‖Softmax(c))

[0084] Where Softmax is the Softmax function, JS(·‖·) represents the Jensen-Shannon divergence, and g(v,v) ab ) represents the SCADA feature optimization gain, and c represents the SCADA feature vector.

[0085] It is understandable that the exploratory operation of merely generating the pressure time-frequency feature masking optimization vector cannot reveal its true utility for the downstream residual prediction task. The system cannot determine whether this refining operation removes interference noise that causes errors or removes key information beneficial to prediction, and any subsequent feature selection will lack objective basis. Therefore, in the technical solution of this invention, the feature optimization gain of the pressure time-frequency feature masking optimization vector relative to the SCADA feature vector is further calculated to obtain the SCADA feature optimization gain. Specifically, these two feature vectors are input into the downstream residual model respectively, and the performance difference between the pressure time-frequency spectrum visual features and SCADA features in driving the residual model to predict the true error is directly compared, thereby accurately quantifying the utility of the refining operation. In this way, a clear, data-driven decision basis can be provided for the subsequent condition selection steps, transforming the fuzzy judgment about the merits of features into a measurable numerical index, and providing a quantitative basis for realizing adaptive refining of feature vectors.

[0086] Specifically, in step S313, in response to the feature optimization gain satisfying a preset condition, the pressure time-frequency feature masking optimization vector is set as the enhanced SCADA feature vector, expressed by the formula:

[0087]

[0088] Where τ is the preset threshold, γ is the trainable weight coefficient, and v en This refers to the enhanced SCADA feature vector.

[0089] It should be understood that the feature optimization gain calculated in the aforementioned steps is merely a quantitative indicator. Without a clear decision-making mechanism to apply this indicator, the system will be unable to determine whether the refinement operation on the pressure time-frequency spectrum removes interference noise or key information, thus lacking a closed-loop guarantee link to ensure positive returns in the entire feature purification process. Therefore, in the technical solution of this invention, in response to the feature optimization gain meeting a preset condition, the pressure time-frequency feature masking optimization vector is set as the enhanced SCADA feature vector, thereby constructing an intelligent gating selection mechanism. This mechanism ensures that the optimization is adopted only when a specific refinement operation on the visual features of the pressure time-frequency spectrum is proven to bring substantial performance improvement. In this way, it can be ensured that each enhancement of the final generated enhanced SCADA feature vector is based on evidence and beneficial to the downstream residual prediction task, thereby avoiding the risk of losing key pressure time-frequency spectrum visual information due to blindly modifying features, realizing online adaptive optimization of feature representation, and providing the highest quality input for the residual model.

[0090] Specifically, in step S320, the enhanced SCADA feature vector is input into the residual model to obtain the prediction error correction value. It should be understood that the enhanced SCADA feature vector obtained after feature refinement is merely an optimized, high-information-density data carrier. Without an efficient and specialized model to analyze it and establish a mapping relationship with the prediction error of the basic model, the advantages brought by feature refinement cannot be translated into actual accuracy improvement. Therefore, in the technical solution of this invention, the enhanced SCADA feature vector is further input into the residual model to utilize this extremely lightweight linear regression model to perform rapid mathematical transformations and fitting on these core features highly correlated with the error. In this way, the systematic deviation of the basic model under the current specific operating conditions can be accurately calculated with extremely low computational resource consumption, thereby generating a numerical prediction error correction value that can be directly used for correction.

[0091] More specifically, in a concrete example of this invention, a very simple linear regression residual model is first pre-deployed in the PLC. Its core contains only a weight vector and a bias term parameter, both of which are stored in the PLC's memory. When the upstream feature refining module generates an enhanced SCADA feature vector, this vector is immediately passed as input to the residual model. The system then performs a forward propagation calculation, which essentially involves a vector dot product operation between the enhanced SCADA feature vector and the weight vector of the residual model, followed by adding the result to the bias term parameter. Due to the simplicity of the model structure, this calculation process consumes almost no CPU time. The final scalar result obtained from this calculation is the prediction error correction value, which numerically represents the residual model's estimate of the current prediction error of the base model. This correction value is then output and used to fuse with the base load prediction value to complete the final accuracy correction.

[0092] Specifically, in step S400, the prediction error correction value is added to the basic load prediction value to obtain the final load prediction value. It should be understood that in the previous steps, the system has generated two independent values ​​in parallel: one is the basic load prediction value generated by the basic model, which reflects the main load change trend but has systematic deviations; the other is the prediction error correction value generated by the residual model, specifically used to quantify this systematic deviation. These two values ​​are functionally complementary, but without effective fusion, a high-precision final prediction result that combines the advantages of both cannot be formed. Therefore, in the technical solution of this invention, the prediction error correction value is further added to the basic load prediction value to directly and arithmetically integrate the macroscopic prediction of the main trend with the precise compensation for microscopic deviations. In this way, the systematic errors of the basic model can be canceled out through an extremely simple calculation operation, thereby synthesizing a high-fidelity final load prediction value that is numerically closer to the actual physical load.

[0093] More specifically, in a concrete example of this invention, firstly, within a specific program scan cycle of the PLC, the system simultaneously reads two key floating-point variables from its internal data register. The first variable is the basic load prediction value calculated and output by the preceding basic model module, and the second variable is the prediction error correction value calculated and output by the preceding residual model module. Subsequently, the PLC's central processing unit invokes a basic arithmetic addition instruction. This instruction uses the two read values ​​as operands to perform an addition operation. After the operation is completed, the result, i.e., the sum of the two values, is immediately written to a pre-allocated independent data register for storing the final result. This new value stored in the register is defined as the final load prediction value output by this technical solution and can be directly called by the wind turbine's control system or monitoring system.

[0094] Specifically, in step S500, drift monitoring and adaptive updating of the residual model are performed based on the prediction error between the final predicted load value and the actual load value to obtain the updated residual model. It should be understood that the operating conditions of wind turbines, the external environment, and the state of the equipment itself will slowly and continuously change over time. This phenomenon is called concept drift, which causes the predictive performance of the pre-trained fixed model (including the basic model and the initial residual model) to gradually decay over time because the historical data distribution learned by the model no longer matches the current actual data distribution. Therefore, in the technical solution of this invention, a closed-loop feedback mechanism for drift monitoring and adaptive updating of the residual model based on the prediction error between the final predicted load value and the actual load value is further implemented. This continuously monitors the actual performance of the model, and when performance degradation is detected, the latest operating data is used to retrain the simple and easily updated residual model online. This ensures that the entire prediction system has the adaptive capability to cope with concept drift, dynamically correcting the prediction logic without manual intervention or retraining of complex basic models, thereby guaranteeing the long-term effectiveness and high accuracy of the prediction scheme throughout the entire life cycle of the wind turbine.

[0095] More specifically, in this embodiment of the invention, drift monitoring and adaptive updating of the residual model based on the prediction error between the final predicted load value and the actual load value to obtain an updated residual model includes: calculating the difference between the final predicted load value and the actual load value to obtain the prediction error; storing the prediction error in an error history buffer, the error history buffer including the most recent N prediction errors; performing statistical analysis on the error history buffer to obtain a time series of the error moving average; in response to the absolute value of the error moving average exceeding a preset threshold M times consecutively, determining that the base model has significantly drifted and triggering the residual model update process; extracting recent feature-target pairs, and updating the parameters of the residual model based on the recent feature-target pairs using stochastic gradient descent.

[0096] Specifically, the difference between the final predicted load value and the actual load value is calculated to obtain the prediction error. It should be understood that the final predicted load value is merely an estimate of the future state by the model, and its accuracy is unknown. Without an objective benchmark for comparison, it is impossible to evaluate the model's true performance, let alone determine whether the model's performance has degraded or drifted due to changes in operating conditions. Therefore, in the technical solution of this invention, the difference between the final predicted load value and the actual load value is further calculated to transform the abstract performance representation predicted by the model into a precise and measurable value, namely, the prediction error. This provides the most direct and crucial feedback signal and quantitative basis for subsequent drift monitoring and adaptive updates of the residual model, enabling the entire prediction system to form a closed loop with self-evaluation capabilities.

[0097] More specifically, in a concrete example of the present invention, firstly, within a subsequent control cycle, the system simultaneously retrieves two values ​​from its historical data buffer, based on a timestamp alignment principle. One is the previously stored final load prediction value for a specific historical moment; the other is the actual load value, collected and confirmed by the load sensor, that corresponds exactly to that historical moment. Subsequently, the PLC's central processing unit executes a subtraction instruction, subtracting the final load prediction value from the actual load value, and stores the calculated difference, i.e., the prediction error, as a new floating-point variable in a specially allocated error sequence storage area. This prediction error value is then read by the subsequent drift monitoring module to evaluate the long-term performance stability of the model.

[0098] Specifically, the prediction error is stored in an error history buffer, which includes the most recent N prediction errors. It should be understood that a single instantaneous prediction error value is easily affected by random disturbances and measurement noise, and lacks statistical representativeness. Directly judging whether the model performance has undergone systematic drift based on this isolated value is highly prone to misjudgment. Therefore, in the technical solution of this invention, the prediction error is further stored in an error history buffer to construct a fixed-length time series sample that reflects the recent performance of the model. This provides a smoothed, effectively noise-filtered, and statistically more robust performance evaluation basis for the subsequent drift monitoring module, making the judgment of the model performance degradation trend more accurate and reliable.

[0099] More specifically, in a concrete example of the present invention, the error history buffer is first implemented using a first-in-first-out circular queue structure. This queue is pre-allocated with N consecutive storage units in the PLC's data storage area and includes a pointer indicating the current write position. When a new prediction error value is calculated in the previous step, this value is written to the storage unit currently pointed to by the pointer. After the write operation is complete, the pointer automatically moves forward one position. When the pointer reaches the end of the queue (i.e., the Nth unit), it automatically wraps back to the beginning of the queue (i.e., the first unit) during the next write operation, thus overwriting the earliest historical error value stored in the buffer. Through this circular write mechanism, the error history buffer dynamically maintains the system's most recent N prediction error records at any given time, forming a sliding observation window.

[0100] Specifically, statistical analysis is performed on the error history buffer to obtain the time series of the moving average of errors. It should be understood that the original error sequence directly stored in the error history buffer contains a large amount of random noise and instantaneous fluctuations. These high-frequency disturbances can mask the true trend of systematic and gradual changes in model performance. Directly judging drift based on this original sequence would lead to an extremely high false alarm rate. Therefore, in the technical solution of this invention, statistical analysis is further performed on the error history buffer to smooth the original error sequence and filter out high-frequency noise and instantaneous disturbances by calculating the error mean within a sliding window. This generates a stationary index called the moving average of errors, which more accurately reflects the long-term trend of model performance, providing a more accurate decision-making basis for subsequent drift detection that is less susceptible to random factors.

[0101] More specifically, in a concrete example of the present invention, the process is first configured to be triggered after each new prediction error is stored in the error history buffer. Once triggered, the system reads all N error values ​​stored in the current error history buffer. Subsequently, the PLC's central processing unit performs a loop summation operation, accumulating these N error values ​​to obtain a sum. Next, a division operation is performed, dividing the sum of errors by a constant N, thereby calculating the current moving average error value. Finally, this calculated moving average error value is appended as a new data point to a time-series data structure specifically used to store this indicator, for use by the subsequent drift threshold judgment module.

[0102] Specifically, in response to the absolute value of the moving average error exceeding a preset threshold M times consecutively, a significant drift in the base model is determined, triggering the update process of the residual model. It should be understood that a single instance of the moving average error exceeding the threshold is merely a short-term, drastic fluctuation or a statistically random event, rather than a true systemic decline in model performance. Triggering a computationally expensive model update process based solely on a single instance of exceeding the threshold would lead to unnecessary resource waste, causing the model to overfit to transient noise and compromising its stability. Therefore, in the technical solution of this invention, a mechanism is further implemented where drift is determined and an update is triggered only when the absolute value of the moving average error exceeds the preset threshold M times consecutively. This constructs a precise decision-making logic with a time dimension, ensuring that adaptive correction is only initiated when model performance exhibits persistent and deterministic degradation. This improves the accuracy and reliability of drift detection, effectively filters out transient interference, and guarantees the necessity and effectiveness of model update operations, thereby achieving a balance between adaptability and stability.

[0103] More specifically, in a concrete example of the present invention, the system first presets two key parameters in the non-volatile memory area of ​​the PLC: a floating-point threshold as a drift judgment benchmark, and an integer constant M for confirming persistence. Simultaneously, the system maintains an integer counter with an initial value of zero. In each control cycle, when a new error moving average is calculated, the system first takes its absolute value and compares it with the preset floating-point threshold. If the absolute value is greater than the threshold, the value of the internal counter is incremented; otherwise, if it is less than or equal to the threshold, the counter is immediately cleared. After each operation on the counter, the system immediately checks whether the current value of the counter has reached the integer constant M. Once the counter value is detected to be equal to M, the system determines that a significant drift has occurred and immediately executes a preset trigger operation, such as setting a specific internal relay or calling a subroutine specifically for initiating online updates of the residual model. After the trigger operation is completed, the counter is cleared again to prepare for the next drift detection cycle.

[0104] Specifically, recent feature-target pairs are extracted, and the parameters of the residual model are updated using stochastic gradient descent based on these pairs. It is worth noting that the features in the feature-target pairs are SCADA feature vectors, and the target is the prediction error. It should be understood that drift determination merely confirms that the existing residual model has failed. Without a concrete mechanism for relearning using the latest data, the system cannot correct this failure, and the entire adaptive update loop cannot be completed. Therefore, in the technical solution of this invention, recent feature-target pairs are further extracted, and the parameters of the residual model are updated using stochastic gradient descent based on these pairs. This utilizes data that best reflects the current operating conditions to efficiently fine-tune or refit this simple, computationally inexpensive linear model online. This allows the residual model to quickly learn and adapt to the changed data distribution, and its internal parameters are recalibrated to reflect the new mapping relationship between features and errors, thereby restoring the accuracy of the entire prediction system without affecting the complex underlying model.

[0105] More specifically, in a concrete example of the invention, firstly, once the update process is triggered, the most recent batch of feature-target data pairs is extracted from its internal historical data buffer. In this scenario, each data pair contains an enhanced SCADA feature vector (as a feature) at a historical moment and its corresponding prediction error (as a target). Subsequently, an iterative optimization loop based on the stochastic gradient descent algorithm is initiated. In each step of this loop, one or a small set of samples is randomly selected from the extracted batch of data pairs. For the selected sample, a forward propagation is performed, i.e., the feature vector of the sample is input into the current residual model to obtain a temporary error prediction value. Next, the loss (e.g., mean squared error) between this temporary error prediction value and the true error target of the sample is calculated. Then, backpropagation is performed based on this loss value to calculate the gradient of the loss function with respect to each parameter (i.e., weights and biases) of the residual model. Finally, based on the calculated gradient and a preset learning rate, a small adjustment is made to each parameter of the model to reduce the loss. This iterative cycle will continue until all extracted data pairs have been traversed or the preset number of iterations has been reached. At this point, the parameters of the residual model will be updated to the new state that best fits the recent data and fixed in the PLC for later use.

[0106] In summary, the PLC-based real-time load prediction method for wind turbines according to embodiments of the present invention has been clarified. It transforms raw, mixed SCADA data into high-quality feature vectors through a pre-processed feature engineering module, ensuring the reliability of the prediction from the source. Based on this, an innovative decoupled architecture of the base model and residual model is adopted: a quantized and compressed base model runs efficiently on the PLC, providing a load baseline prediction; an extremely lightweight residual model specifically learns and compensates for the prediction error of the base model, achieving improved accuracy with minimal computational overhead. Finally, to address model performance drift caused by changes in operating conditions, a low-cost drift monitoring and adaptive update closed loop is introduced. By monitoring prediction errors in real time, only the lightweight residual model is rapidly fine-tuned online, thereby ensuring high precision and accuracy of the model during long-term operation.

[0107] Furthermore, a PLC-based real-time load prediction system for wind turbine units is also provided.

[0108] Figure 5 This is a block diagram of a PLC-based real-time load prediction system for wind turbines according to an embodiment of the present invention. Figure 5As shown, the PLC-based real-time load prediction system 500 for wind turbines according to an embodiment of the present invention includes: a feature vector acquisition module 510, used to perform feature engineering on the acquired raw SCADA data to obtain SCADA feature vectors; a load baseline prediction module 520, used to perform load baseline prediction based on a base model on the SCADA feature vectors to obtain a base load prediction value; an adaptive residual prediction module 530, used to perform adaptive residual prediction based on a residual model on the SCADA feature vectors to obtain a prediction error correction value; a final load prediction module 540, used to add the prediction error correction value and the base load prediction value to obtain a final load prediction value; and a residual model generation module 550, used to perform drift monitoring and adaptive updating of the residual model based on the prediction error between the final load prediction value and the actual load value to obtain an updated residual model.

[0109] Furthermore, the feature vector acquisition module 510 is specifically used for: performing data alignment and null value detection on the original SCADA data based on a sliding time window to obtain aligned data records; performing data cleaning and filling on the aligned data records based on historical valid values ​​to obtain clean data records; and performing numerical normalization and vector assembly on the clean data records based on pre-stored parameters to obtain the SCADA feature vector.

[0110] Furthermore, the residual model generation module 550 is specifically used for: calculating the difference between the final predicted load value and the actual load value to obtain the prediction error; storing the prediction error in an error history buffer, the error history buffer including the most recent N prediction errors; performing statistical analysis on the error history buffer to obtain the time series of the error moving average; in response to the absolute value of the error moving average exceeding a preset threshold M times consecutively, determining that the base model has significantly drifted and triggering the residual model update process; extracting recent feature-target pairs, and updating the parameters of the residual model based on the recent feature-target pairs using stochastic gradient descent.

[0111] As described above, the PLC-based real-time wind turbine load prediction system 500 according to embodiments of the present invention can be implemented in various wireless terminals, such as servers with PLC-based real-time wind turbine load prediction algorithms. In one possible implementation, the PLC-based real-time wind turbine load prediction system 500 according to embodiments of the present invention can be integrated into the wireless terminal as a software module and / or hardware module. For example, the PLC-based real-time wind turbine load prediction system 500 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the PLC-based real-time wind turbine load prediction system 500 can also be one of many hardware modules of the wireless terminal.

[0112] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A PLC-based real-time load prediction method for wind turbine units, characterized in that, include: Feature engineering is performed on the acquired raw SCADA data to obtain SCADA feature vectors; The load baseline prediction based on the base model is performed on the SCADA feature vector to obtain the base load prediction value; Adaptive residual prediction based on a residual model is performed on the SCADA feature vectors to obtain the prediction error correction value; The final load prediction value is obtained by adding the prediction error correction value to the basic load prediction value. Drift monitoring and adaptive updating of the residual model are performed based on the prediction error between the final predicted load value and the actual load value to obtain the updated residual model.

2. The real-time load prediction method for wind turbine units based on PLC according to claim 1, characterized in that, Feature engineering is performed on the acquired raw SCADA data to obtain SCADA feature vectors, including: Data alignment and null value detection based on a sliding time window are performed on the raw SCADA data to obtain aligned data records; The aligned data records are cleaned and filled based on historical valid values ​​to obtain clean data records; The clean data records are subjected to numerical normalization and vector assembly based on pre-stored parameters to obtain the SCADA feature vector.

3. The real-time load prediction method for wind turbine units based on PLC according to claim 1, characterized in that, The base model is a quantized GRU model, and the residual model is a small linear regression model.

4. The real-time load prediction method for wind turbine units based on PLC according to claim 2, characterized in that, Adaptive residual prediction based on a residual model is performed on the SCADA feature vectors to obtain prediction error correction values, including: Perform feature refinement on the SCADA feature vectors to obtain enhanced SCADA feature vectors; The enhanced SCADA feature vector is input into the residual model to obtain the prediction error correction value.

5. The real-time load prediction method for wind turbine units based on PLC according to claim 4, characterized in that, Perform feature refinement on the SCADA feature vectors to obtain enhanced SCADA feature vectors, including: The SCADA feature vector is subjected to structured feature masking to obtain an optimized pressure time-frequency feature masking vector; Calculate the feature optimization gain of the pressure time-frequency feature masking optimization vector relative to the SCADA feature vector; In response to the feature optimization gain satisfying a preset condition, the pressure time-frequency feature masking optimization vector is set as the enhanced SCADA feature vector.

6. The real-time load prediction method for wind turbine units based on PLC according to claim 1, characterized in that, Based on the prediction error between the final predicted load value and the actual load value, drift monitoring and adaptive updating of the residual model are performed to obtain the updated residual model, including: The difference between the final predicted load value and the actual load value is calculated to obtain the prediction error; The prediction error is stored in an error history buffer, which includes the most recent N prediction errors. Statistical analysis was performed on the error history buffer to obtain the time series of the error moving average; If the absolute value of the moving average error exceeds a preset threshold M times consecutively, it is determined that the base model has drifted significantly and the update process of the residual model is triggered. Extract recent feature-target pairs and update the parameters of the residual model based on the recent feature-target pairs using stochastic gradient descent.

7. The real-time load prediction method for wind turbine units based on PLC according to claim 6, characterized in that, The feature in the feature-target pair is the SCADA feature vector, and the target is the prediction error.

8. A PLC-based real-time load prediction system for wind turbine generators, characterized in that, include: The feature vector acquisition module is used to perform feature engineering on the acquired raw SCADA data to obtain SCADA feature vectors. The load baseline prediction module is used to perform load baseline prediction on the SCADA feature vector based on the basic model to obtain the basic load prediction value. The adaptive residual prediction module is used to perform adaptive residual prediction based on the residual model on the SCADA feature vector to obtain the prediction error correction value. The final load prediction module is used to add the prediction error correction value to the basic load prediction value to obtain the final load prediction value. The residual model generation module is used to monitor drift and adaptively update the residual model based on the prediction error between the final load prediction value and the actual load value to obtain the updated residual model.

9. The PLC-based real-time load prediction system for wind turbine units according to claim 8, characterized in that, The feature vector acquisition module is further used for: Data alignment and null value detection based on a sliding time window are performed on the raw SCADA data to obtain aligned data records; The aligned data records are cleaned and filled based on historical valid values ​​to obtain clean data records; The clean data records are subjected to numerical normalization and vector assembly based on pre-stored parameters to obtain the SCADA feature vector.

10. The PLC-based real-time load prediction system for wind turbine units according to claim 8, characterized in that, The residual model generation module is further used for: The difference between the final predicted load value and the actual load value is calculated to obtain the prediction error; The prediction error is stored in an error history buffer, which includes the most recent N prediction errors. Statistical analysis was performed on the error history buffer to obtain the time series of the error moving average; If the absolute value of the moving average error exceeds a preset threshold M times consecutively, it is determined that the base model has drifted significantly and the update process of the residual model is triggered. Extract recent feature-target pairs and update the parameters of the residual model based on the recent feature-target pairs using stochastic gradient descent.

Images