A digital substation parallel control management system based on artificial intelligence

Abstract

The present invention relates to the field of artificial intelligence, and specifically refers to an artificial intelligence-based digital substation parallel control management system, comprising a data acquisition module, a data processing module, an artificial intelligence analysis module, a parallel control module, a digital twin module and a communication module. The present invention introduces a window-by-window attention long-short-term memory model in the artificial intelligence analysis module, and processes multidimensional data streams more efficiently through window-by-window processing, thereby improving data analysis capabilities; a whale optimization algorithm is used in the window-by-window attention long-short-term memory model, and the initial diversity and global search capabilities of the population are improved through chaotic mapping and random reverse learning, and the dynamic update capability of the window-by-window attention long-short-term model at different stages is enhanced through the golden sine, and adaptive disturbances are added to enhance the exploration capability.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and specifically refers to a digital substation parallel control management system based on artificial intelligence. Background Art

[0002] With the rapid development of smart grids, substations, as key nodes for power transmission and distribution, have become increasingly complex and automated. In general systems, the analysis and prediction of data in substations have problems such as low prediction accuracy and inability to process multi-dimensional time series data. Ordinary methods such as linear regression or simple LSTM cannot efficiently capture the long-term and short-term dependencies in substation equipment data, and linear models cannot fully model the nonlinear relationships in these data, while simple LSTM may not perform well when processing large-scale multi-dimensional data. In general models, traditional optimization algorithms are usually used to update model parameters, such as stochastic gradient descent and traditional particle swarm optimization. There are problems such as insufficient exploration of the search space and insufficient diversity maintenance. Relying on gradient information for search, the search of certain local areas is insufficient, and convergence to poor areas in the early stages affects the overall optimization effect. Summary of the invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a digital substation parallel control management system based on artificial intelligence. In view of the problems of low prediction accuracy and inability to process multi-dimensional time series data in analysis and prediction in general systems, the present invention introduces a window-by-window attention long-short-term memory model in the artificial intelligence analysis module. By dynamically adjusting the focus on different time windows and combining with optimized model parameters, it can more accurately capture short-term fluctuations and long-term trends in substation equipment data. Through window-by-window processing, multi-dimensional data streams can be processed more efficiently to improve data analysis capabilities. In view of the problems of insufficient search space exploration and insufficient diversity maintenance in traditional optimization algorithms in general models, the present invention uses a whale optimization algorithm combined with chaotic mapping, random reverse learning, golden sine and adaptive distributed perturbation in the window-by-window attention long-short-term memory model. The initial diversity and global search capability of the population are improved through chaotic mapping and random reverse learning, the dynamic update capability of the window-by-window attention long-short-term model at different stages is enhanced through the golden sine, and the exploration capability is enhanced by adding adaptive perturbation.

[0004] The technical solution adopted by the present invention is as follows: The present invention provides an artificial intelligence-based digital substation parallel control management system, including a data acquisition module, a data processing module, an artificial intelligence analysis module, a parallel control module, a digital twin module and a communication module, specifically including the following contents:

[0005] The data acquisition module collects time series data from various devices in the substation and marks the time stamp, including real-time data and historical data. The data types include device status information, power load, voltage, current, temperature and humidity;

[0006] The data processing module preprocesses the time series data collected by the data collection module, including data cleaning, outlier detection and data standardization operations;

[0007] The artificial intelligence analysis module uses a window-by-window attention long short-term memory model to analyze and predict the preprocessed data;

[0008] The parallel control module uses parallel computing technology to synchronously dispatch and control various types of equipment in the substation;

[0009] The digital twin module creates a virtual model of the substation and simulates the operating status through real-time data;

[0010] The communication module uses the TCP communication protocol to achieve data transmission between the above modules and between the substation and external equipment.

[0011] Furthermore, the artificial intelligence analysis module uses the window-by-window attention long short-term memory model to analyze and predict the preprocessed data. The specific implementation steps are as follows:

[0012] Step S1: Data preparation, defining timestamps to identify specific time points of each preprocessed time series data, arranging them in order of timestamps to obtain arranged time series data;

[0013] Step S2: Split the preprocessed time series data into independent time windows, set the size of the time window, and define the number of consecutive timestamps in each time window as the time step;

[0014] Step S3: Generate training samples, extract input features and output labels according to the time window, and use the sliding window method to generate training samples from the arranged time series data. Each training sample includes the input features and corresponding output labels within the time window;

[0015] Step S4: Dataset division: divide the training samples into training set and validation set with a ratio of 70% and 30%;

[0016] Step S5: Model construction, constructing a window-by-window attention long short-term memory model;

[0017] Step S6: Model training and verification, training the window-by-window attention long short-term memory model and verifying it using the verification set;

[0018] Step S7: Implement prediction.

[0019] Furthermore, in step S5, a window-by-window attention long short-term memory model is constructed, which specifically includes the following contents:

[0020] Input layer: receives input features within the time window and sets the input dimension to three dimensions, including time step, number of features, and batch size;

[0021] LSTM layer: sets the number of LSTM units and activation function, processes the input features within the time window, and outputs the hidden state of each time step;

[0022] Fully connected layer: Set 64 neurons to map the three-dimensional data output by the LSTM layer into a two-dimensional array.

[0023] Window-by-window attention mechanism: Introduce a non-overlapping attention mechanism to perform independent attention calculations in each time window, calculate attention weights for input features at each time step, and weight each time step of the LSTM output to obtain a weighted hidden state;

[0024] Output layer: maps the weighted hidden states to the final predicted output.

[0025] Further, step S6 specifically includes the following steps:

[0026] Step S61: Initialize the model, set the training cycle and the number of iterations, and initialize the parameters of the window-by-window attention long short-term memory model, including all parameters in the input layer, LSTM layer, fully connected layer, window-by-window attention mechanism, and input layer;

[0027] Step S62: forward propagation, input the training set, pass through the input layer, LSTM layer, fully connected layer, window-by-window attention mechanism and input layer, and obtain the prediction result through the output layer;

[0028] Step S63: define a loss function, use the mean square error as the loss function, and calculate the error of the prediction result obtained by forward propagation through the loss function;

[0029] Step S64: Back propagation, calculate the gradient of each model parameter according to the loss function, and pass the error obtained in step S53 from the output layer back to the input layer;

[0030] Step S65: Update the model parameters, use the whale optimization algorithm combined with the four strategies to update the parameters of the window-by-window attention long short-term memory model to obtain a new parameter combination;

[0031] Step S66: Iterate the training process. After each training cycle, use the validation set to validate the model. When the number of iterations is reached, stop the training and output the final window-by-window attention long short-term memory model.

[0032] Step S67: Evaluate the performance and use the loss function to evaluate the window-by-window attention long short-term memory model.

[0033] Further, step S65, using the whale optimization algorithm combined with the four strategies, specifically includes the following steps:

[0034] Step S651: Initialize the population, set the population size, maximum number of iterations and number of individuals, and use chaotic mapping and random reverse learning strategy to initialize the population position of the whale optimization algorithm. The formula used is as follows:

[0035] ;

[0036] ;

[0037] in, For the The location of an individual, is a random number in the interval [0,1], For the The reverse solution position of each individual, is the lower boundary of the search space, is the upper boundary of the search space;

[0038] Step S652: Fitness calculation, calculating the fitness value of each individual, and selecting the one with the highest fitness as the current optimal solution;

[0039] Step S653: Position update: the position is updated according to the hunting behavior of the whale optimization algorithm combined with the golden sine strategy. The formula used is as follows:

[0040] ;

[0041] in, For the updated The location of an individual, is the current optimal solution, is the radius, and is a random angle;

[0042] Step S654: Adaptive perturbation: When the number of iterations reaches half of the maximum number, an adaptive t-distribution perturbation is added. The formula used is as follows:

[0043] ;

[0044] in, For the The new position of each individual after the disturbance, is a random number in the interval [0,1], is the current iteration number, is a random number that follows a uniform distribution, is a random number in the range of (0,1);

[0045] Step S655: competition mechanism, updating the position and speed of individuals based on the comparison of fitness values, and individuals with high fitness replace individuals with low fitness;

[0046] Step S656: Iteration process, iterating step S652 to step S655 until the maximum number of iterations is reached, and outputting the individual with the highest fitness as the parameter combination.

[0047] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0048] (1) In view of the problems of low prediction accuracy and inability to process multi-dimensional time series data in general system analysis and prediction, the present invention introduces a window-by-window attention long short-term memory model in the artificial intelligence analysis module. By dynamically adjusting the focus on different time windows and combining optimized model parameters, it can more accurately capture the short-term fluctuations and long-term trends in substation equipment data. Through window-by-window processing, it can more efficiently process multi-dimensional data streams and improve data analysis capabilities.

[0049] (2) In view of the problems of insufficient search space exploration and insufficient diversity maintenance in traditional optimization algorithms in general models, the present invention uses the whale optimization algorithm combined with chaotic mapping, random reverse learning, golden sine and adaptive distributed perturbation in the window-by-window attention long-term and short-term memory model. The initial diversity and global search ability of the population are improved by chaotic mapping and random reverse learning, the dynamic update ability of the window-by-window attention long-term and short-term model at different stages is enhanced by the golden sine, and the exploration ability is enhanced by adding adaptive perturbation. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A schematic diagram of a digital substation parallel control management system based on artificial intelligence proposed by the present invention;

[0051] Figure 2 A schematic diagram of the steps of analyzing and predicting the preprocessed data using the window-by-window attention long short-term memory model in the present invention;

[0052] Figure 3 It is a flowchart of the specific steps of step S6 in the present invention.

[0053] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0055] Example 1, see Figure 1 The present invention provides an artificial intelligence-based digital substation parallel control management system, including a data acquisition module, a data processing module, an artificial intelligence analysis module, a parallel control module, a digital twin module and a communication module, specifically including the following contents:

[0056] The data acquisition module collects time series data from various devices in the substation and marks the time stamp, including real-time data and historical data. The data types include device status information, power load, voltage, current, temperature and humidity;

[0057] The data processing module preprocesses the time series data collected by the data collection module, including data cleaning, outlier detection and data standardization operations;

[0058] The artificial intelligence analysis module uses a window-by-window attention long short-term memory model to analyze and predict the preprocessed data;

[0059] The parallel control module uses parallel computing technology to synchronously dispatch and control various types of equipment in the substation;

[0060] The digital twin module creates a virtual model of the substation and simulates the operating status through real-time data;

[0061] The communication module uses the TCP communication protocol to achieve data transmission between the above modules and between the substation and external equipment.

[0062] Example 2, see Figure 2 This embodiment is based on the above embodiment. The artificial intelligence analysis module uses the window-by-window attention long short-term memory model to analyze and predict the preprocessed data. The specific implementation steps are as follows:

[0063] Step S1: Data preparation, defining timestamps to identify specific time points of each preprocessed time series data, arranging them in order of timestamps to obtain arranged time series data;

[0064] Step S2: Split the preprocessed time series data into independent time windows, set the size of the time window, and define the number of consecutive timestamps in each time window as the time step;

[0065] Step S3: Generate training samples, extract input features and output labels according to the time window, and use the sliding window method to generate training samples from the arranged time series data. Each training sample includes the input features and corresponding output labels within the time window;

[0066] Step S4: Dataset division: divide the training samples into training set and validation set with a ratio of 70% and 30%;

[0067] Step S5: Model construction, constructing a window-by-window attention long short-term memory model;

[0068] Step S6: Model training and verification, training the window-by-window attention long short-term memory model and verifying it using the verification set;

[0069] Step S7: Implement prediction.

[0070] In this embodiment, the data is sorted according to the timestamp to ensure that the data is arranged in chronological order. The code used is as follows:

[0071] import pandas as pd

[0072] # Assume the data has been preprocessed and stored in a CSV file

[0073] data = pd.read_csv('processed_data.csv')

[0074] # Add timestamp

[0075] data['timestamp'] = pd.to_datetime(data['timestamp'])

[0076] data = data.sort_values('timestamp')

[0077] Select one hour (3600 seconds) as the time window. If the data is one record per minute, the number of time steps is 60. The code used is as follows:

[0078] # Set the time window and time step

[0079] window_size = 3600 # 1 hour

[0080] time_step = 60 # one data point per minute

[0081] # Split time series data into time windows

[0082] windows = []

[0083] for start in range(0, len(data) - window_size + 1, time_step):

[0084] windows.append(data.iloc[start:start + window_size])

[0085] For each time window, the input features are extracted as the data of the first 59 minutes and the output label is the current value of the 60th minute. The sliding window method is used to generate training samples. The code used is as follows:

[0086] import numpy as np

[0087] X = []

[0088] y = []

[0089] for window in windows:

[0090] # Extract input features (first 59 minutes)

[0091] features = window.iloc[:-1, 1:].values ​​# Assume that the first column is the timestamp and the following columns are features

[0092] target = window.iloc[-1, 1] # target value for the 60th minute

[0093] X.append(features)

[0094] y.append(target)

[0095] X = np.array(X)

[0096] y = np.array(y).

[0097] Embodiment 3: Based on the above embodiment, in step S5, a window-by-window attention long short-term memory model is constructed, which specifically includes the following contents:

[0098] Input layer: receives input features within the time window and sets the input dimension to three dimensions, including time step, number of features, and batch size;

[0099] LSTM layer: sets the number of LSTM units and activation function, processes the input features within the time window, and outputs the hidden state of each time step;

[0100] Fully connected layer: Set 64 neurons to map the three-dimensional data output by the LSTM layer into a two-dimensional array.

[0101] Window-by-window attention mechanism: Introduce a non-overlapping attention mechanism to perform independent attention calculations in each time window, calculate attention weights for input features at each time step, and weight each time step of the LSTM output to obtain a weighted hidden state;

[0102] Output layer: maps the weighted hidden states to the final predicted output.

[0103] In this embodiment, the dimension of the input data is three-dimensional, specifically including the following contents:

[0104] Time steps: There are 60 time steps in each time window, that is, the time span of each time window is 1 hour (for example, data is recorded once every minute, for a total of 60 records);

[0105] Number of features: 5 features per time step, such as voltage, current, temperature, humidity, and pressure;

[0106] Batch size: The batch size is 32, which means that 32 time windows of data are input each time;

[0107] The number of LSTM units is set to 128, the activation function is set to tanh, and the return sequence is set to return_sequences=True;

[0108] The output layer uses a linear activation function to predict future values.

[0109] Example 4, see Figure 3 This embodiment is based on the above embodiment, and step S6 specifically includes the following steps:

[0110] Step S61: Initialize the model, set the training cycle and the number of iterations, and initialize the parameters of the window-by-window attention long short-term memory model, including all parameters in the input layer, LSTM layer, fully connected layer, window-by-window attention mechanism, and input layer;

[0111] Step S62: forward propagation, input the training set, pass through the input layer, LSTM layer, fully connected layer, window-by-window attention mechanism and input layer, and obtain the prediction result through the output layer;

[0112] Step S63: define a loss function, use the mean square error as the loss function, and calculate the error of the prediction result obtained by forward propagation through the loss function;

[0113] Step S64: Back propagation, calculate the gradient of each model parameter according to the loss function, and pass the error obtained in step S53 from the output layer back to the input layer;

[0114] Step S65: Update the model parameters, use the whale optimization algorithm combined with the four strategies to update the parameters of the window-by-window attention long short-term memory model to obtain a new parameter combination;

[0115] Step S66: Iterate the training process. After each training cycle, use the validation set to validate the model. When the number of iterations is reached, stop the training and output the final window-by-window attention long short-term memory model.

[0116] Step S67: Evaluate the performance and use the loss function to evaluate the window-by-window attention long short-term memory model.

[0117] In this embodiment, the training cycle is set to 100 times, and the number of iterations in each cycle is 200 times. The parameters of the aforementioned model, such as the input dimension, the number of LSTM units, and the number of neurons in the fully connected layer, are also applicable to this stage.

[0118] Embodiment 5, based on the above embodiment, step S65, uses the whale optimization algorithm combined with four strategies, specifically including the following steps:

[0119] Step S651: Initialize the population, set the population size, maximum number of iterations and number of individuals, and use chaotic mapping and random reverse learning strategy to initialize the population position of the whale optimization algorithm. The formula used is as follows:

[0120] ;

[0121] ;

[0122] in, For the The location of an individual, is a random number in the interval [0,1], For the The reverse solution position of each individual, is the lower boundary of the search space, is the upper boundary of the search space;

[0123] Step S652: Fitness calculation, calculating the fitness value of each individual, and selecting the one with the highest fitness as the current optimal solution;

[0124] Step S653: Position update: the position is updated according to the hunting behavior of the whale optimization algorithm combined with the golden sine strategy. The formula used is as follows:

[0125] ;

[0126] in, For the updated The location of an individual, is the current optimal solution, is the radius, and is a random angle;

[0127] Step S654: Adaptive perturbation: When the number of iterations reaches half of the maximum number, an adaptive t-distribution perturbation is added. The formula used is as follows:

[0128] ;

[0129] in, For the The new position of each individual after the disturbance, is a random number in the interval [0,1], is the current iteration number, is a random number that follows a uniform distribution, is a random number in the range of (0,1);

[0130] Step S655: competition mechanism, updating the position and speed of individuals based on the comparison of fitness values, and individuals with high fitness replace individuals with low fitness;

[0131] Step S656: Iteration process, iterating step S652 to step S655 until the maximum number of iterations is reached, and outputting the individual with the highest fitness as the parameter combination.

[0132] In this example, the population size is set to 3 individuals (to simplify the calculation presentation), which are , and , the lower boundary of the search space is -10, the upper boundary of the search space is 10, and the random number is 0.2, is 0.5, is 0.8, and the initial position is calculated for each individual:

[0133] ;

[0134] ;

[0135] ;

[0136] Initialize the reverse solution position using a random reverse learning strategy:

[0137] ;

[0138] ;

[0139] ;

[0140] The initial population position is , the reverse position is ;

[0141] Fitness calculation, set a simple fitness function, the fitness value is equal to the sum of the squares of each individual, so the fitness value is the smallest , is 0.04;

[0142] In step S653, the radius is set to 0.5 and the angle is set to random. and Set to and , then the current optimal solution is calculated to be 0.2, and using the position update formula, we get is 0.506, is 0.506, is 0.506, so the updated position is [0.506, 0.506, 0.506];

[0143] In adaptive perturbation, the random number Set to 0.3, is 0.1, is 0.5, the current number of iterations is 25, and the maximum number of iterations is 50, so disturbances are added;

[0144] According to the formula, is -1.739, is -1.739, is -1.739, so the individual position after disturbance is [-1.739, -1.739, -1.739];

[0145] After the adaptive disturbance, the individual fitness value is recalculated and compared with the previous fitness value. The individuals with high fitness replace the individuals with low fitness. The fitness value after the disturbance is calculated to be 3.025, which is greater than 0.04, so the original optimal solution is retained;

[0146] The final position remains unchanged at [0.2,-0.5,-0.809].

[0147] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0148] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0149] The present invention and its embodiments are described above, which is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it and design structural methods and embodiments similar to the technical solution without creativity without departing from the purpose of the invention, they should all fall within the protection scope of the present invention.

Claims

1. An artificial intelligence-based digital substation parallel control management system, characterized by: It includes data acquisition module, data processing module, artificial intelligence analysis module, parallel control module, digital twin module and communication module, specifically including the following contents: The data acquisition module collects time series data from various devices in the substation and marks the time stamp, including real-time data and historical data. The data types include device status information, power load, voltage, current, temperature and humidity; The data processing module preprocesses the time series data collected by the data collection module, including data cleaning, outlier detection and data standardization operations; The artificial intelligence analysis module uses a window-by-window attention long short-term memory model to analyze and predict the preprocessed data; The parallel control module uses parallel computing technology to synchronously dispatch and control various types of equipment in the substation; The digital twin module creates a virtual model of the substation and simulates the operating status through real-time data; The communication module uses TCP communication protocol to realize data transmission between the above modules and between the substation and external equipment; The artificial intelligence analysis module uses the window-by-window attention long short-term memory model to analyze and predict the preprocessed data. The specific implementation steps are as follows: Step S1: Data preparation, defining timestamps to identify specific time points of each preprocessed time series data, arranging them in order of timestamps to obtain arranged time series data; Step S2: Split the preprocessed time series data into independent time windows, set the size of the time window, and define the number of consecutive timestamps in each time window as the time step; Step S3: Generate training samples, extract input features and output labels according to the time window, and use the sliding window method to generate training samples from the arranged time series data. Each training sample includes the input features and corresponding output labels within the time window; Step S4: Dataset division: divide the training samples into training set and validation set with a ratio of 70% and 30%; Step S5: Model construction, constructing a window-by-window attention long short-term memory model; Step S6: Model training and verification, training the window-by-window attention long short-term memory model and verifying it using the verification set; Step S7: Implement prediction; Step S6 specifically includes the following steps: Step S61: Initialize the model, set the training cycle and the number of iterations, and initialize the parameters of the window-by-window attention long short-term memory model, including all parameters in the input layer, LSTM layer, fully connected layer, window-by-window attention mechanism, and input layer; Step S62: forward propagation, input the training set, pass through the input layer, LSTM layer, fully connected layer, window-by-window attention mechanism and input layer, and obtain the prediction result through the output layer; Step S63: define a loss function, use the mean square error as the loss function, and calculate the error of the prediction result obtained by forward propagation through the loss function; Step S64: Back propagation, calculate the gradient of each model parameter according to the loss function, and pass the error obtained in step S53 from the output layer back to the input layer; Step S65: Update the model parameters, use the whale optimization algorithm combined with the four strategies to update the parameters of the window-by-window attention long short-term memory model to obtain a new parameter combination; Step S66: Iterate the training process. After each training cycle, use the validation set to validate the model. When the number of iterations is reached, stop the training and output the final window-by-window attention long short-term memory model. Step S67: Evaluate the performance, using a loss function to evaluate the window-by-window attention long short-term memory model; Step S65, using the whale optimization algorithm combined with the four strategies, specifically includes the following steps: Step S651: Initialize the population, set the population size, maximum number of iterations and number of individuals, and use chaotic mapping and random reverse learning strategy to initialize the population position of the whale optimization algorithm. The formula used is as follows: ； ； in, For the The location of an individual, is a random number in the interval [0,1], For the The reverse solution position of each individual, is the lower boundary of the search space, is the upper boundary of the search space; Step S652: Fitness calculation, calculating the fitness value of each individual, and selecting the one with the highest fitness as the current optimal solution; Step S653: Position update: the position is updated according to the hunting behavior of the whale optimization algorithm combined with the golden sine strategy. The formula used is as follows: ； in, For the updated The location of an individual, is the current optimal solution, is the radius, and is a random angle; Step S654: Adaptive perturbation: When the number of iterations reaches half of the maximum number, an adaptive t-distribution perturbation is added. The formula used is as follows: ； in, For the The new position of each individual after the disturbance, is a random number in the interval [0,1], is the current iteration number, is a random number that follows a uniform distribution, is a random number in the range of (0,1); Step S655: competition mechanism, updating the position and speed of individuals based on the comparison of fitness values, and individuals with high fitness replace individuals with low fitness; Step S656: Iteration process, iterating step S652 to step S655 until the maximum number of iterations is reached and the output fitness is the highest.

2. According to the artificial intelligence-based digital substation parallel control management system of claim 1, it is characterized by: In step S5, a window-by-window attention long short-term memory model is constructed, which specifically includes the following contents: Input layer: receives input features within the time window and sets the input dimension to three dimensions, including time step, number of features, and batch size; LSTM layer: sets the number of LSTM units and activation function, processes the input features within the time window, and outputs the hidden state of each time step; Fully connected layer: Set 64 neurons to map the three-dimensional data output by the LSTM layer into a two-dimensional array. Window-by-window attention mechanism: Introduce a non-overlapping attention mechanism to perform independent attention calculations in each time window, calculate attention weights for input features at each time step, and weight each time step of the LSTM output to obtain a weighted hidden state; Output layer: maps the weighted hidden states to the final predicted output.

Images