A wisdom station SCADA data early warning method and system based on long convolution and data control gate

Abstract

The application provides a wisdom station SCADA data early warning method and system based on long convolution and data control gate, the method comprises the following steps: constructing a physical model to organize a wind field signal; inputting the wind field signal into a trained linear response filter to obtain a prediction result; comparing the prediction result with an actual signal to obtain a first difference; comparing the first difference with a set difference, and if the difference is greater than the set difference, early warning is performed; the set difference is the difference between the prediction result output by the linear response filter during training and the actual signal. The application has the advantages that: the physical model is used to uniformly set the data style, and the standardization processing of the whole field data assets is realized; compared with the traditional attention mechanism, the prediction method of the application has low complexity, and the long convolution fully utilizes the length of the signal, so compared with the local convolution calculation, the long convolution has better signal feature learning ability.

Description

Technical Field

[0001] This application belongs to the field of wind power technology, specifically relating to a smart wind farm SCADA data early warning method and system based on long convolution and data control gates. Background Technology

[0002] As the construction of wind farms continues to deepen, a wind farm often has wind turbines from different manufacturers and of different models. In order to meet the users' requirements for comprehensive monitoring and management of wind farms, a unified platform is needed to monitor and manage the various wind turbines and external equipment such as access control, fire protection, condition monitoring, and transformer substations.

[0003] A comprehensive intelligent monitoring system covering the entire wind farm faces challenges due to the wide scope and diverse categories of management, involving numerous and complex equipment types. Furthermore, current integrated systems are based on SCADA (Supervisory Control and Data Acquisition) data, and different manufacturers' units often use their own specified data formats. Additionally, manufacturers' external equipment monitoring systems operate independently, further complicating the construction of an intelligent monitoring system. From an inspection perspective, the types of maintenance items are numerous, and maintenance standards and cycles are inconsistent, necessitating a management and regular inspection dispatch system. For example, major components requiring maintenance include: nacelle control cabinets, hub control cabinets, tower base control cabinets, and tower base converter cabinets. These involve specific maintenance items such as: generator speed monitoring panel fixing bolts, pitch bearing seal and visual inspection, pitch reducer oil level checks, and nacelle lighting maintenance. Each maintenance item has a defined maintenance cycle, making the overall maintenance management of the wind farm a complex issue. Besides routine inspections, monitoring components should also provide operational status updates, alarm detection, and information on whether personnel are operating in compliance with regulations, as well as timely management. From the perspective of reducing personnel workload, the system platform needs to achieve standardized information across the entire facility, integrated system functions, visualized operational status, and intelligent operation and maintenance management.

[0004] In the AI ​​field, SCADA data warning algorithms mainly include the base methods of Transformer and Convergence. The attention operator of Transformer depends on the scale and the content learned from the background, so its time cost increases quadratically with the scale. To reduce the time cost, popular methods use linearized, low-scale, and feature-sparse approximation methods, such as Generating long sequences with sparse transformers, Linformer: Self-attention with linear complexity, and the efficient transformer. However, the drawback is the trade-off between rich feature representation and computational speed. Summary of the Invention

[0005] The purpose of this application is to overcome the shortcomings of different manufacturers' units often using their own specified data formats and the fact that manufacturers' external equipment monitoring systems are also independent, resulting in complex and diverse data types, management, and communication protocols. The application aims to improve the accuracy of SCADA data early warning and reduce computing costs.

[0006] To achieve the above objectives, this application proposes a smart station SCADA data early warning method based on long convolution and data control gates, including:

[0007] Constructing a model to organize wind field signals;

[0008] The wind field signal is input into a trained linear response filter to obtain the prediction result;

[0009] The predicted result is compared with the actual signal to obtain the first difference;

[0010] The first difference is compared with the set difference. If the difference is greater than the set difference, an alert is issued.

[0011] The set difference is the difference between the predicted result output by the linear response filter during training and the actual signal.

[0012] As an improvement to the above method, the calculation process of the linear response filter includes:

[0013] Step 1: Convert the input signal u∈R L×D Perform a linear transformation to generate a linearly transformed tensor. The tensor shape is L×(N+1)D; where L represents the signal length, D represents the signal type, N is a positive integer, and R represents a real number;

[0014] tensor Each vector in the second dimension is convolved in one dimension to generate the convolved variable z.

[0015] Divide the variable z into N+1 equal parts, and use a data-driven feature x for each of the first N parts. i Let i = 1, 2, ..., N represent the current predicted feature quantity, and let v represent the last block. All blocks have the same shape, R. L×D :

[0016] Step 2: Use sine and cosine functions to generate the positional encoding variable t;

[0017] A feedforward neural network (FFN), consisting of a linear transformation and an activation function, is used to perform a linear transformation on the position-encoded variable t to generate an intermediate variable. This intermediate variable is then input into a ReLU activation function to generate the variable calculated by the FFN.

[0018] Modify variables A third dimension is extended and transposed to generate the filter variable h; the first dimension of variable h is divided into N equal blocks, each block consisting of the filter variable h from a single iteration. i i = 1, 2, ..., N represents the variable h. i The shape is R D×L :

[0019] Step 3: Calculate the single-iteration filter h using Fast Fourier Transform. n The matrix multiplication between the current predicted feature v and the matrix x is used to obtain a weight matrix, which is then used to drive the feature x. i The latest predicted feature value v is obtained by weighting the features at the position of the feature value in the sequence; the calculation is repeated N times, and the last updated predicted feature value v is taken as the final prediction result.

[0020] As an improvement to the above method, the tensor A one-dimensional convolution operation is performed on each vector in the second dimension, using a convolution kernel size of 3.

[0021] As an improvement to the above method, the step of generating the position encoding variable t using sine and cosine functions includes:

[0022] The odd-numbered feature bits t of the position encoding variable t pos,2i+1 =sin(ω) i ·pos), even-numbered characteristic bits t pos,2i =cos(ω i ·pos), i = 0, 1, ..., D e / 2, pos is the position index from 0 to L-1, D eThis is a location embedding feature.

[0023] This application also provides a smart station SCADA data early warning system based on long convolution and data control gates, implemented using the above method, the system comprising:

[0024] The construct model module is used to organize wind field signals in the construct model;

[0025] The prediction module is used to input the wind field signal into the trained linear response filter to obtain the prediction result; compare the prediction result with the actual signal to obtain the first difference; compare the first difference with a set difference, and if the difference is greater than the set difference, an early warning is issued; the set difference is the difference between the prediction result output by the linear response filter during training and the actual signal.

[0026] Compared with existing technologies, the advantages of this application are:

[0027] 1. A unified data style is set using a physical model. The physical model includes equipment type, equipment name, data measurement point name, control action, alarm value, and equipment attributes. The physical model provides support for asset and model management. The configuration management module is divided into asset tree, model management, model grouping, page configuration, alarm reminder configuration, and maintenance records. It groups and manages data by type, which is beneficial for sorting out and managing the complex data of the wind farm station and realizing standardized processing of data assets across the entire site.

[0028] 2. The time complexity of the long convolutional data control gate algorithm is O(NLlog2L), where L is the length of the input signal and N is the number of repetitions in each Fast Fourier Transform (FFT) unit. Compared to the traditional attention mechanism, O(L... 2 The advantage of the method in this application is its low complexity. Moreover, long convolution makes full use of the signal length, so it has a better signal feature learning ability compared with local convolution calculation. Attached Figure Description

[0029] Figure 1 The diagram shown is the overall architecture of the wind farm intelligent monitoring system.

[0030] Figure 2 The diagram shows a flowchart of a smart facility SCADA data early warning method based on long convolution and data control gates; where resurrence represents recursion and filters represent filters.

[0031] Figure 3 The image shows the effect of the transformation replacing the attention operator. Detailed Implementation

[0032] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0033] This invention proposes a smart station SCADA data early warning method and system based on long convolution and data control gates. The method uses a physical model to uniformly set the data style. The physical model includes equipment type, equipment name, data measurement point name, control action, alarm value, and equipment attributes. The physical model provides support for asset and model management. The technical features of the early warning method include: (1) data control, which enables the attention mechanism to realize linear operators with rich data control through matrix factorization; (2) sublinear parameter scaling, which decouples the parameters of the attention layer from the sequence length, thereby allocating more parameters to the feedforward convolutional network; (3) unrestricted text content, which preserves the characteristics of the attention mechanism for given fixed input data, that is, the interdependencies between texts without any distance constraints.

[0034] The predicted scenarios in this application include: wind turbine blade angle imbalance, wind turbine blade icing, wind turbine derating, abnormal low-speed bearing temperature in the gearbox, abnormal high-speed bearing temperature in the gearbox, gearbox heat dissipation, abnormal nacelle cabinet temperature, abnormal nacelle ambient temperature, abnormal non-drive end bearing temperature in the generator, abnormal drive end bearing temperature in the generator, three-phase imbalance in generator stator coil temperature, abnormal generator stator winding temperature, imbalance between main shaft speed and generator speed, abnormal pitch capacitor temperature, abnormal pitch motor performance, abnormal pitch motor temperature, abnormal pitch inverter temperature, wind vane freezing, anemometer malfunction, and abnormal wind turbine power.

[0035] like Figure 1 As shown, the intelligent wind farm SCADA data early warning method and system based on long convolution and data control gates provided in this application are part of the wind farm intelligent monitoring system. The entire architecture of the wind farm intelligent monitoring system is based on the data style of the object model. The backend system receives object model data and reorganizes the data according to the logical storage style of asset tree, model tree, model grouping, alarm configuration, and maintenance records, and then stores and analyzes the data. The backend system can also realize online and offline development, automatic deployment, and environmental isolation functions. The frontend system interface displays information according to different functional dimensions, equipment placement locations, and scheduling plans, and summarizes it into four functional modules: operation monitoring, daily inspection, site security, and maintenance. Finally, the homepage provides an overview of wind farm information. The early warning method of this application is located in the algorithm analysis part of the backend system of the overall system architecture. It adopts a long convolution and data control gate architecture, based on a loop calculation architecture of two efficient sub-quadratic basic modules. One basic module is an implicit long convolution and the other is a matrix point-by-point multiplication gate module. This method can be represented as the product of a diagonal matrix controlled by the input data and a Topplitz matrix, such as... Figure 2 As shown. This method features sublinear parameter scaling and unrestricted text association, which is similar to traditional attention mechanisms but with lower time complexity.

[0036] Example 1

[0037] The intelligent SCADA data early warning method for power stations based on long convolution and data control gates provided in this application includes:

[0038] Step S1: Design the object model to organize the original data. The organization name can be divided into asset tree, model management, model grouping, page configuration, alarm reminder configuration, and maintenance record, supporting the calling and querying of upper-level modules. The object model data information organization table records the information header of the data, and provides information query fields and constraint configuration information for obtaining data points.

[0039] Each organization name is equivalent to a data organization information table, which is filled out and entered on the front end. This table is used to query location information and configure data constraints. Only with this data information table can one query a specific data location, extract data, obtain the warning threshold, constraints, and warning scheduling configuration for that location.

[0040] The asset type supports adding new regions, intervals, and equipment. On the asset details page, you can view basic information and the attributes of the patrol equipment.

[0041] The asset tree contains multiple connected wind farms, and each wind farm contains a list of equipment. The equipment list fields include: equipment name, equipment code, farm name, region, type, and number of inspection points. Users can view, delete, bind to models, associate equipment, add components, and add inspection points. Model binding requires binding data from the group's industry data platform, including: equipment asset ID, ERP asset ID, all-in-one machine asset ID, and KKS code. Equipment association includes model classification and selecting the equipment to associate. Adding components includes equipment identifier and equipment name. Adding inspection points includes equipment identifier, equipment name, parent component, inspection type, inspection point level, inspection standard, meter type, auxiliary equipment type, data collection / saving file, and identification type.

[0042] Model management includes: model identifier, model name, model category, model version, and model operations. The model category allows selection of model classifications, including main equipment and auxiliary equipment. Model operations include a maintenance item data table, enabling seamless navigation between model management and maintenance item management. The maintenance item table includes attribute configuration and inspection point configuration. Attribute configuration includes parameter name, field type, and whether a field is required. Inspection point configuration includes point identifier, mail code, mail name, point name, inspection type, point level, inspection standard, meter type, auxiliary facility type, data collection / saving file, identification type, phase, lower warning limit, upper warning limit, general lower limit, general upper limit, severe upper limit, severe lower limit, critical upper limit, and critical lower limit.

[0043] Models can be grouped, enabling recombination between different models, and the recombination information is then distributed to the sites recorded in the asset tree. The data includes model group code, model group name, application site, start / stop status, and the site to which the data was distributed.

[0044] The alarm notification configuration allows you to fill in the alarm content, notification method, and notification frequency.

[0045] Maintenance records describe the maintenance history of the equipment, covering information such as site, area, maintenance interval, equipment information, maintenance content, and maintenance time.

[0046] Step S2: Design a learnable linear response filter h(t), and express the filter γ in terms of the learnable parameter θ at time step t. θ (t), this parameter is not explicitly given as a hyperparameter but is an implicit parameter.

[0047] A filter is equivalent to the convolution kernel of a neural network or a black box. When a vector is input, the filter outputs a vector. The error between the output vector and the true vector is calculated. The deviation of the convolution kernel is calculated in reverse according to the chain rule, and the convolution kernel is updated.

[0048] The linear response filter h(t) is γ θ (t) can be expressed by a first-order difference equation:

[0049] x t+1 =Ax t +Bu t Equations of state

[0050] y t =Cx t +Du t Output equation

[0051] Let x0 = 0,

[0052]

[0053] In the above formula, t represents time, and u t For input quantity, y t For the output quantity, x t δ is an intermediate state variable. t-n A is the Kronecker function; A, B, C, and D are the parameters of the linear filter to be learned, and the elements in the four matrices are updated during the training process.

[0054] convolution The time complexity in a basic convolutional neural network is O(L). 2L is the size of the matrix or the length of the input vector. The matrix is ​​diagonalized using the Cooley-Tukey Fast Fourier Transform (FFT), i.e.:

[0055] y = S h *u

[0056] =W -1 D H W*u

[0057] =iFFT(D H FFT(u))

[0058] The time cost of FFT convolution is O(Llog2L).

[0059] The effect of transformation replacing attention operator in matrix diagonalization is as follows: Figure 3 As shown, the calculation process using the attention operator is as follows:

[0060] First, calculate the attention matrix A(u):

[0061]

[0062] Where, u∈R L×D It is the input vector; M q M k ∈R D×D It is a learnable linear projection matrix, and SoftMax is the activation function.

[0063] Next, calculate the output:

[0064] y=A(u)uM v ,

[0065] Among them, M v ∈R D×D It is a learnable linear projection matrix.

[0066] The time complexity of the entire calculation process is O(L). 2 ).

[0067] The key vector and query vector of traditional attention operators are no longer discussed here.

[0068]

[0069] in, v is the score vector of the attention mechanism.

[0070] In S2.4, x n ∈R L×D The matrix formed by placing the vector (D=1) on the diagonal of the matrix.

[0071] If we use the attention mechanism to calculate the expression for y, it is y = A(u)u M v Then use The expression is more complex than using A(u)uM v Expressions save computational costs.

[0072] Step S2 specifically includes:

[0073] Step S2.1 Projection;

[0074] 1) Input signal u∈R L×D Perform a linear transformation to generate a linearly transformed tensor. The tensor shape is L×(N+1)D; L represents the signal length, D represents the signal type, and R represents a real number.

[0075]

[0076] 2) To Each vector in the second dimension (whose shape is R) L Performing a one-dimensional convolution operation generates a convolution variable z, the shape of which is R. L×(N+1)D The one-dimensional convolution kernel w used in convolution h The size is 3.

[0077] w h ∈R 3 Local convolution kernel

[0078] 3) Divide the variable z into blocks, into N+1 equal blocks. Each of the first N blocks uses a data-driven feature x. i Let i = 1, 2, ..., N represent the current predicted feature quantity, and let v represent the last block. All blocks have the same shape, R. L×D :

[0079] Decompose z as x 1 x 2 ... x N ,vR L×(N+1)D →R L×D ,R L×D ,…,R L×D ,R L×D

[0080] Step S2.2: Generate the filter;

[0081] 1) Use sine and cosine functions to generate position encoding variables t, let pos be the position index from 0 to L-1, and the odd feature bits t pos,2i+1 =sin(ω) i ·pos), even-numbered characteristic bits t pos,2i =cos(ωi ·pos), i = 0, 1, ..., D e / 2, D e This is a location embedding feature; the final generated feature has a length of L and an encoded feature size of D. e Feature location encoding variables:

[0082]

[0083] 2) A feedforward neural network (FFN) consisting of a linear transformation and an activation function performs a linear transformation on the variable t to generate an intermediate variable, then adds a ReLU-corrected linear unit activation function to the intermediate variable, and finally generates the variable calculated by the FFN.

[0084]

[0085] 3) Regarding variables The shape is modified to extend a third dimension, and then transposed to generate the filter variable h. The first dimension of the variable h is divided into N equal blocks, each block consisting of the filter variable h from a single iteration. i i = 1, 2, ..., N represents the variable h. i The shape is R D×L :

[0086]

[0087] Decompose h into h 1 h 2 ... h N

[0088] Step 2.3 Forward computation;

[0089] Forward computation requires the use of variables {x} generated by projection and generative filtering. i ,v} and {h i}

[0090] Calculate the single-iteration filter h using Fast Fourier Transform. n The matrix multiplication between the current predicted feature v and the matrix x is used to obtain a weight matrix, which is then used to drive a specific data feature x. n The latest predicted feature value v is obtained by weighting the features at different positions in the sequence. This process is repeated N times, and the last updated predicted feature value v is used as the final predicted feature value of the sequence.

[0091]

[0092] Step 2.4 compares the final sequence prediction feature with the actual value, and compares the difference with the standard deviation obtained during training. If the comparison result exceeds the set threshold, an early warning is issued.

[0093] Let's illustrate the use of linear response filters for early warning in a real-world scenario:

[0094] 1. Wind speed-power warning:

[0095] During training, take at least three months of wind speed-power data under normal operating conditions, input wind speed u, output power y, and calculate the difference between the output y and the actual power y'.

[0096] During prediction, the system takes 100 sampled wind speed data as input, outputs the predicted power, compares it with the actual power value, and then compares the comparison value with the difference calculated during training. If the difference exceeds the set threshold, an early warning is issued.

[0097] 2. Gearbox high-speed shaft bearing temperature warning:

[0098] During training:

[0099] (1) Input short-term 10-minute data, one data point per second, take 10,000 10-minute temperature data randomly collected over at least 12 months under normal working conditions, input the temperature data of the first 9 minutes, output the temperature data of the last minute, and the difference between the temperature and the actual temperature;

[0100] (2) Input mid-term 1-day data, one data point per minute, take 10,000 1-day temperatures randomly collected over at least 12 months under normal operating conditions, input the temperature data for the first 23 hours, output the temperature data for the last hour, and the difference between the temperature and the actual temperature.

[0101] (3) Input long-term 7-day data, with one data point every 5 minutes. Take 10,000 7-day temperatures randomly collected over at least 12 months under normal operating conditions. Input the temperature data for the first 6 days and output the temperature data for the last day, along with the difference between the temperature and the actual temperature.

[0102] When a warning is issued:

[0103] (1) Short-term warning: Input the current 9-minute temperature data, output the 1-minute prediction, compare it with the actual temperature value, and then compare it with the difference obtained during training. If it exceeds the set threshold, a warning will be issued.

[0104] (2) Mid-term warning: Input the current 23 hours of data, output the 1-hour prediction, compare it with the actual temperature value, and then compare it with the difference obtained during training. If it exceeds the set threshold, a warning will be issued.

[0105] (3) Long-term warning: Input the current 6 days of data, output the 1-day prediction, compare it with the actual temperature value, and then compare it with the difference obtained during training. If it exceeds the set threshold, a warning will be issued.

[0106] In the smart station SCADA data early warning method based on long convolution and data control gate provided in this application, the object model can make the organizational data structure flexible and convenient for upper-level applications to extract model data structures; the early warning method uses a data control operator to replace the traditional attention mechanism operator. This operator is composed of multiple cyclically multiplied long convolutions and data control gates. This method no longer depends on the key vector and query vector of the attention mechanism, thus decoupling the input vector size and parameter size of the attention mechanism.

[0107] Example 2

[0108] This application also provides a smart station SCADA data early warning system based on long convolution and data control gates, implemented using the above method, the system comprising:

[0109] The construct model module is used to organize wind field signals in the construct model;

[0110] The prediction module is used to input the wind field signal into the trained linear response filter to obtain the prediction result; compare the prediction result with the actual signal to obtain the first difference; compare the first difference with a set difference, and if the difference is greater than the set difference, an early warning is issued; the set difference is the difference between the prediction result output by the linear response filter during training and the actual signal.

[0111] This application may also provide a computer device, including: at least one processor, memory, at least one network interface, and a user interface. The various components in this device are coupled together via a bus system. It is understood that the bus system is used to implement communication between these components. In addition to a data bus, the bus system also includes a power bus, a control bus, and a status signal bus.

[0112] The user interface can include a display, keyboard, or clicking device. Examples include a mouse, trackball, touchpad, or touchscreen.

[0113] It is understood that the memory in the embodiments disclosed in this application may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory may be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DRRAM). The memories described herein are intended to include, but are not limited to, these and any other suitable types of memory.

[0114] In some implementations, the memory stores elements such as executable modules or data structures, or subsets thereof, or extended sets thereof: operating systems and applications.

[0115] The operating system includes various system programs, such as the framework layer, core library layer, and driver layer, used to implement various basic business functions and handle hardware-based tasks. The application programs include various applications, such as media players and browsers, used to implement various application functions. Programs implementing the methods of the embodiments of this disclosure can be included in the application programs.

[0116] In the above embodiments, the processor can also invoke programs or instructions stored in memory, specifically programs or instructions stored in an application program, for the following purposes:

[0117] Follow the steps described above.

[0118] The above methods can be applied to or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above methods can be completed through integrated logic circuits in the processor's hardware or through software instructions. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the disclosed methods, steps, and logic block diagrams. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the disclosed methods can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0119] It is understood that the embodiments described in this application can be implemented using hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described in this application, or combinations thereof.

[0120] For software implementation, the technology of this application can be implemented by executing the functional modules (e.g., procedures, functions, etc.) of this application. The software code can be stored in memory and executed by a processor. The memory can be implemented in the processor or externally.

[0121] This application may also provide a non-volatile storage medium for storing a computer program. When the computer program is executed by a processor, it can implement the steps in the above method embodiments.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of this application do not depart from the spirit and scope of the technical solutions of this application, and should all be covered within the scope of the claims of this application.

Claims

1. A smart facility SCADA data early warning method based on long convolution and data control gating, comprising: Constructing a model to organize wind field signals; The wind field signal is input into a trained linear response filter to obtain the prediction result; The predicted result is compared with the actual signal to obtain the first difference; The first difference is compared with the set difference. If the difference is greater than the set difference, an alert is issued. The set difference value is the difference between the predicted result output by the linear response filter during training and the actual signal; The calculation process of the linear response filter includes: Step 1: Input signal Perform a linear transformation to generate a linearly transformed tensor. , The shape of the tensor is ;in, L Indicates the signal length. D Indicates the type of signal. N It is a positive integer. R Represent real numbers; tensor Perform a one-dimensional convolution operation on each vector in the second dimension to generate the convolved variable. ; variables z Divide into equal parts N +1 piece, front N Each block uses a data-driven feature. i=1,2,..,N It means that the last piece is used This represents the current predicted feature quantity, and all blocks have the same shape. : Step 2: Use sine and cosine functions to generate positional encoding variables t ; The position-encoding variable is achieved using a feedforward neural network (FFN) consisting of a linear transformation and an activation function. t A linear transformation is performed to generate intermediate variables, which are then input into the ReLU activation function to generate the variables after FFN calculation. ; Modify variables The third dimension is extended and transposed to generate filter variables. ; to variable h The first dimension is divided into equal parts N Blocks, each block consisting of a single-iteration filter variable i=1,2,...,N Indicates variables The shape is : Step 3: Calculate the single-iteration filter using Fast Fourier Transform. and the current predicted feature quantity v The matrix multiplication between the two results in a weight matrix, which is then used to drive the features of the data. The latest predicted feature value is obtained by weighting the features at the position of the feature value in the sequence. v ; Loop calculation N Next, the current predicted feature quantity updated last time. v As the final prediction result; The use of sine and cosine functions to generate position coding variables t ,include: The location encoding variable t odd feature bits sin( ) Even-numbered characteristic bits cos( )，i=0,1,…, , , pos From 0 to L -1 position index, This is a location embedding feature.

2. The intelligent station SCADA data early warning method based on long convolution and data control gates according to claim 1, characterized in that, The tensor A one-dimensional convolution operation is performed on each vector in the second dimension, using a convolution kernel size of 3.

3. A smart station SCADA data early warning system based on long convolution and data control gating, implemented based on the method described in any one of claims 1-2, characterized in that, The system includes: The construct model module is used to organize wind field signals in construct models; and The prediction module is used to input the wind field signal into the trained linear response filter to obtain the prediction result; compare the prediction result with the actual signal to obtain the first difference; compare the first difference with a set difference, and if the difference is greater than the set difference, an early warning is issued; the set difference is the difference between the prediction result output by the linear response filter during training and the actual signal.

Images