Metro station environment and ventilation linkage energy-saving optimization control method based on deep learning

Abstract

The application discloses a subway station environment control and air supply linkage energy-saving optimization control method based on deep learning, which comprises the following steps: preprocessing train operation state, station environment data and environment control equipment data to obtain a time sequence data matrix; decomposing the time sequence data matrix through non-negative matrix decomposition to extract disturbance characteristics in the time sequence data matrix and output a blind source separation result matrix; performing convolution operation on the blind source separation result matrix through a convolutional neural network to extract spatial modes of the disturbance characteristics and generate a spatial feature vector through dimension reduction of pooling; capturing time variation rules of dynamic disturbance through a bidirectional GRU structure in a recurrent neural network to generate a hidden state sequence; introducing an attention mechanism in an output layer of the recurrent neural network to focus on the hidden state when a train enters a station, and generating an optimized disturbance characteristic representation through weight weighting; and taking the optimized disturbance characteristic representation as an input of a full connection layer to generate and output a corresponding energy-saving strategy category.

Description

Technical Field

[0001] This invention belongs to the field of subway station environmental control technology, and more specifically, relates to a deep learning-based energy-saving optimization control method for subway station environmental control and ventilation linkage. Background Technology

[0002] In modern subway station operations, the environmental control and ventilation system has become a key infrastructure to ensure passenger comfort and the stability of station equipment. This system achieves dynamic balance and energy-efficient operation of the station's microenvironment by comprehensively regulating airflow, temperature, humidity, and water cooling circulation.

[0003] Current technologies mostly employ strategies based on traditional control models and empirical parameters, which are effective for adjusting constant passenger flow and stable environmental parameters. However, when trains enter or leave the station, the platform area is subjected to instantaneous pressure waves and thermal disturbances (typical pressure fluctuations range approximately ±50 Pa, and train car temperatures are usually 2-3°C higher than platform temperatures), leading to drastic fluctuations in system energy consumption and control lag. This traditional control method has significant shortcomings in dealing with rapid, nonlinear, and strongly coupled dynamic changes, making it difficult to achieve refined, real-time energy-saving optimization control. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to overcome the aforementioned deficiencies and propose a deep learning-based energy-saving optimization control method for the environmental control and ventilation linkage of subway stations.

[0005] The present invention adopts the following technical solution.

[0006] The first aspect of this invention discloses a deep learning-based energy-saving optimization control method for environmental control and ventilation systems in subway stations, the method comprising:

[0007] Acquire train operation status, station environment data, and environmental control equipment data, and preprocess the train operation status, station environment data, and environmental control equipment data to obtain a time series data matrix;

[0008] The time series data matrix is ​​decomposed by nonnegative matrix factorization to extract the perturbation characteristics in the time series data matrix and output the blind source separation result matrix.

[0009] The blind source separation result matrix is ​​convolved by a convolutional neural network to extract the spatial pattern of the perturbation characteristics, and spatial feature vectors are generated by pooling dimensionality reduction.

[0010] The spatial feature vector is used as the input to the bidirectional GRU structure in the recurrent neural network to capture the time variation law of dynamic perturbation and generate the hidden state sequence.

[0011] Based on the hidden state sequence, an attention mechanism is introduced into the output layer of the recurrent neural network to focus on the hidden state when the train enters the station, and an optimized perturbation feature representation is generated through weighted summation.

[0012] The optimized perturbation feature representation is used as the input to the fully connected layer to generate and output the corresponding energy-saving strategy category;

[0013] The train operation status includes train speed and train arrival time; the station environment data includes air pressure, temperature and humidity; the environmental control equipment data includes fan speed, chilled water flow and air conditioning cooling load; the disturbance characteristics include pressure wave intensity and thermal disturbance mode; and the spatial mode includes pressure wave distribution and heat transfer path in different areas of the subway station.

[0014] Furthermore, the station environment data and environmental control equipment data are data collected by sensors or equipment deployed according to the environmental control requirements of the subway station, used to reflect the pressure waves and thermal disturbances caused by trains entering and leaving the station.

[0015] The process involves acquiring train operation status, station environment data, and environmental control equipment data, and preprocessing this data to obtain a time-series data matrix, including:

[0016] The station environment data and environmental control equipment data are aligned according to timestamps using the same time interval to obtain a collected data vector. The expression of the collected data vector is as follows:

[0017]

[0018] In the formula, Indicates time The collected data vector, For time air pressure data, For time Temperature data, For time Humidity data, For time The fan speed data, For time Cold water flow data, For time Cooling load data;

[0019] High-frequency noise in the acquired data vector is filtered out by low-pass filtering, while signals within a preset frequency range are retained by band-pass filtering, and outliers are removed by a three-standard-deviation rule, with the removed outliers replaced by linear interpolation.

[0020] The collected data vector after filtering, denoising, and outlier removal is sliced ​​into time windows, and the segmented time series is stored in matrix form to obtain the time series data matrix. The expression of the time series data matrix is ​​as follows:

[0021]

[0022] In the formula, The segmented time series data matrix, The data volume of the sampling points is represented in the time series data matrix, where each row represents the time series of one variable and each column represents any time point. All variables.

[0023] Furthermore, the step of decomposing the time series data matrix through nonnegative matrix factorization to extract the perturbation characteristics in the time series data matrix and outputting a blind source separation result matrix includes:

[0024] Obtain the defined dimensions of the time series data matrix, and decompose the time series data matrix based on the defined dimensions using the nonnegative matrix decomposition. The decomposition expression is as follows:

[0025]

[0026] In the formula, The time series data matrix has a defined dimension of m×n; The feature basis matrix has dimensions m×r and is used to represent the weights of the r feature bases, with each row corresponding to m variables; The activation matrix has dimensions r×n and is used to characterize the strength of each feature basis at n time points; r is the number of features in the decomposition, and satisfies... ;

[0027] The matrix reconstruction error is minimized by optimizing the objective function, and the feature basis matrix and activation matrix are updated using an iterative optimization algorithm, which includes a multiplication update rule. The expression for the objective function is:

[0028]

[0029] In the formula, Let f(x) be the Frobenius norm, and let f(x) be the square root of the sum of the squares of the matrix elements. , used for non-negativity constraints on matrix elements.

[0030] Furthermore, the step of decomposing the time series data matrix through nonnegative matrix factorization to extract the perturbation characteristics in the time series data matrix and outputting the blind source separation result matrix further includes:

[0031] The decomposition results of the time series data matrix by the nonnegative matrix decomposition are analyzed, and the activation matrix and the feature basis matrix are normalized respectively, so as to extract the key perturbation time from the activation matrix based on the maximum value and distribution of the variables in the activation matrix, and extract key variables by comparing the column vectors of the feature basis matrix.

[0032] The blind source separation result matrix is ​​reconstructed and output based on the normalized activation matrix and feature basis matrix;

[0033] The normalization expression for the activation matrix is ​​as follows:

[0034]

[0035] In the formula, This is the normalized activation matrix. This represents the maximum value in each row of the activation matrix;

[0036] The normalization expression for the characteristic basis matrix is ​​as follows:

[0037]

[0038] In the formula, The normalized eigenbase matrix, Represents the column sum in the characteristic basis matrix;

[0039] The reconstructed expression for the blind source separation result matrix is:

[0040]

[0041] In the formula, The blind source separation result matrix has dimensions m×n, and each column of the blind source separation result matrix represents a time point. The corresponding blind source characteristics contain key perturbation information.

[0042] Furthermore, the step of performing a convolution operation on the blind source separation result matrix using a convolutional neural network to extract the spatial pattern of the perturbation characteristics, and generating a spatial feature vector through pooling dimensionality reduction, includes:

[0043] The convolutional neural network is invoked to identify local patterns in the blind source separation result matrix through convolution operations, and convolution calculation is performed based on the defined convolution kernel to obtain the output matrix after convolution.

[0044] The spatial feature vector is obtained by processing the output matrix after convolution using a nonlinear activation function, compressing the matrix size through pooling, and then straightening the pooled matrix.

[0045] The expression for the spatial feature vector is:

[0046]

[0047] In the formula, Let p be the spatial eigenvector, and p be the length of the total number of elements in the pooled matrix. This is the matrix after pooling.

[0048] Furthermore, the bidirectional GRU structure includes a forward GRU and a reverse GRU, used to control the transmission of information through update gates and reset gates, wherein the update gates and reset gates are used to determine the retained spatial feature information and forgotten spatial feature information through the Sigmoid activation function;

[0049] The step of using the spatial feature vector as input to a bidirectional GRU structure in a recurrent neural network to capture the temporal variation of dynamic perturbations and generate a hidden state sequence includes:

[0050] The spatial feature vector is used as the input of the recurrent neural network to capture the sequential dependencies in the time series through recursive updates, resulting in a time series input matrix, in which each time step corresponds to a vector.

[0051] The forward GRU and the reverse GRU process the time series input matrix in chronological order to obtain the total hidden state output by the bidirectional GRU structure, and generate a hidden state sequence matrix based on the total hidden state and the time variation law of dynamic perturbation.

[0052] The total hidden state is composed of the hidden states of the forward GRU and the reverse GRU, and the hidden state sequence matrix is ​​used to describe the perturbation characteristics of each time step in the time series.

[0053] Furthermore, based on the hidden state sequence, an attention mechanism is introduced into the output layer of the recurrent neural network to focus on the hidden state when the train enters the station, and an optimized perturbation feature representation is generated through weighted averaging, including:

[0054] The hidden state sequence matrix is ​​used as the input to the attention mechanism introduced by the recurrent neural network to perform a linear transformation on the hidden states in the hidden state sequence matrix, mapping the high-dimensional state to the hidden layer representation of the attention weights.

[0055] The attention weights are calculated by multiplying the hidden layer representation with the global context vector, and the hidden state sequence matrix is ​​weighted and summed to generate the perturbation feature representation.

[0056] The calculation formula for the disturbance feature representation is as follows:

[0057]

[0058] In the formula, This is the perturbation feature representation vector. For time steps, Let be the attention weight at time step t. Let be the hidden state vector at time step t. For global context vectors, This is the weight matrix. The hyperbolic tangent activation function is used. This is the paranoia vector.

[0059] Furthermore, the step of using the optimized perturbation feature representation as input to the fully connected layer to generate and output the corresponding energy-saving strategy category includes:

[0060] The perturbation feature representation is used as the input to the fully connected layer, so as to convert the perturbation feature representation into the input of the classification unit through the linear transformation of the fully connected layer and the nonlinear transformation of the activation function;

[0061] The classification unit is invoked to generate corresponding energy-saving strategy categories based on the number of categories output by the fully connected layer. Energy-saving strategy categories whose category probability exceeds a set threshold are selected as the in-site energy-saving optimization control strategy.

[0062] The probability distribution for generating the energy-saving strategy category is calculated by using the Softmax function to evaluate the output of the fully connected layer, and its expression is as follows:

[0063]

[0064] In the formula, Let c be the probability of class c. The output of the fully connected layer for category c. This represents the total number of categories.

[0065] A second aspect of the present invention discloses a terminal, including a processor and a storage medium; characterized in that:

[0066] The storage medium is used to store instructions;

[0067] The processor is configured to operate according to the instructions to perform the steps of the method described in the first aspect.

[0068] A third aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the steps of the method described in the first aspect.

[0069] The beneficial effects of the present invention are as follows: Compared with the prior art, the present invention has the following advantages:

[0070] By acquiring train operation status, station environment data, and environmental control equipment data, and preprocessing these data, a time-series data matrix is ​​obtained. This matrix is ​​then decomposed using nonnegative matrix factorization (NMF) to extract disturbance characteristics and output a blind source separation result matrix. Subsequently, a convolutional neural network (CNN) is used to convolve the blind source separation result matrix to extract spatial patterns of disturbance characteristics, and pooling is used to reduce dimensionality and generate spatial feature vectors. These spatial feature vectors are then used as input to a bidirectional GRU structure in a recurrent neural network (RNN) to capture the temporal variation patterns of dynamic disturbances and generate a hidden state sequence. Based on this generated hidden state sequence, an attention mechanism is introduced into the output layer of the RNN to focus on the hidden states when the train enters the station, and an optimized disturbance feature representation is generated through weighted summation. Finally, the optimized disturbance feature representation is used as input to a fully connected layer to generate and output the corresponding energy-saving strategy category. This method extracts physical disturbance characteristics using NMF, processes nonlinear dynamic characteristics using an AI model, captures spatial patterns using a CNN, and processes time series data using an RNN, resulting in a more comprehensive feature representation. In addition, by introducing an attention mechanism, key disturbance periods can be highlighted, thereby enhancing the model's ability to respond to complex disturbances. This reduces the need for manual feature selection and enables the generation of energy-saving strategies directly from raw data, achieving more refined and real-time energy-saving optimization control. Attached Figure Description

[0071] Figure 1 This is a flowchart illustrating the energy-saving optimization control method for subway station environmental control and ventilation linkage based on deep learning provided by the present invention. Detailed Implementation

[0072] The present application will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, and should not be construed as limiting the scope of protection of the present application.

[0073] like Figure 1 As shown in one embodiment, a deep learning-based energy-saving optimization control method for environmental control and ventilation linkage in subway stations includes the following steps:

[0074] Step S110: Obtain train operation status, station environment data, and environmental control equipment data, and preprocess the train operation status, station environment data, and environmental control equipment data to obtain a time series data matrix.

[0075] The train operation status includes train speed and train arrival time; the station environment data includes air pressure, temperature and humidity; and the environmental control equipment data includes fan speed, chilled water flow and air conditioning cooling load.

[0076] In some embodiments, the deep learning-based energy-saving optimization control method for subway station environmental control and ventilation linkage provided by the present invention acquires train operation status, station environment data, and environmental control equipment data, and preprocesses the train operation status, station environment data, and environmental control equipment data to obtain a time series data matrix, specifically including the following steps:

[0077] Step S111: Align the station environment data and environmental control equipment data according to timestamps using the same time interval to obtain the collected data vector. The expression for the collected data vector is:

[0078]

[0079] In the formula, Indicates time The collected data vector, For time air pressure data, For time Temperature data, For time Humidity data, For time The fan speed data, For time Cold water flow data, For time The cooling load data.

[0080] Step S112: High-frequency noise in the acquired data vector is filtered out by low-pass filtering, while the signal within the preset frequency range is retained by band-pass filtering, and outliers are removed by a three-standard-deviation rule, and the removed outliers are replaced by linear interpolation.

[0081] Step S113: The collected data vector after filtering, denoising, and outlier removal is sliced ​​into time windows, and the segmented time series is stored in matrix form to obtain the time series data matrix. The expression of the time series data matrix is:

[0082]

[0083] In the formula, The segmented time series data matrix, The data volume of the sampling points is represented in the time series data matrix, where each row represents the time series of one variable and each column represents any point in time. All variables.

[0084] Among them, the station environment data and environmental control equipment data are data collected by sensors or equipment deployed according to the environmental control requirements of the subway station, which are used to reflect the pressure waves and thermal disturbances caused by trains entering and leaving the station.

[0085] Step S120: Decompose the time series data matrix by nonnegative matrix decomposition to extract the perturbation characteristics in the time series data matrix and output the blind source separation result matrix.

[0086] In some embodiments, the deep learning-based energy-saving optimization control method for subway station environmental control and ventilation linkage provided by the present invention decomposes the time series data matrix through non-negative matrix decomposition to extract the disturbance characteristics in the time series data matrix and outputs a blind source separation result matrix, specifically including the following steps:

[0087] Step S121: Obtain the defined dimensions of the time series data matrix, and decompose the time series data matrix based on the defined dimensions using nonnegative matrix decomposition. The decomposition expression is as follows:

[0088]

[0089] In the formula, This is a time series data matrix with the defined dimension m×n; The feature basis matrix has dimensions m×r and is used to represent the weights of the r feature bases, with each row corresponding to m variables; The activation matrix has dimensions r×n and is used to characterize the strength of each feature basis at n time points; r is the number of features in the decomposition, and satisfies... .

[0090] Step S122: Minimize the matrix reconstruction error by optimizing the objective function, and update the feature basis matrix and activation matrix using an iterative optimization algorithm. The iterative optimization algorithm includes a multiplication update rule, and the expression for the objective function is:

[0091]

[0092] In the formula, Let f(x) be the Frobenius norm, and let f(x) be the square root of the sum of the squares of the matrix elements. , used for non-negativity constraints on matrix elements.

[0093] In some embodiments, the deep learning-based energy-saving optimization control method for subway station environmental control and ventilation linkage provided by the present invention decomposes the time series data matrix through non-negative matrix decomposition to extract the disturbance characteristics in the time series data matrix and outputs a blind source separation result matrix. Specifically, it also includes the following steps:

[0094] Step S123: Analyze the decomposition results of the time series data matrix by nonnegative matrix decomposition, and normalize the activation matrix and the characteristic basis matrix respectively, so as to extract the key perturbation moment from the activation matrix based on the maximum value and distribution of the variables in the activation matrix, and extract the key variables by comparing the column vectors of the characteristic basis matrix.

[0095] Step S124: Reconstruct and output the blind source separation result matrix based on the normalized activation matrix and feature basis matrix.

[0096] The normalization expression for the activation matrix is ​​as follows:

[0097]

[0098] In the formula, This is the normalized activation matrix. This represents the maximum value in each row of the activation matrix.

[0099] The normalization expression for the characteristic basis matrix is:

[0100]

[0101] In the formula, The normalized eigenbase matrix, This represents the column sum in the characteristic basis matrix.

[0102] The reconstruction expression for the blind source separation result matrix is:

[0103]

[0104] In the formula, This is a blind source separation result matrix with dimensions m×n. Each column of the blind source separation result matrix represents a time point. The corresponding blind source characteristics contain key perturbation information.

[0105] Step S130: The blind source separation result matrix is ​​convolved by a convolutional neural network to extract the spatial pattern of perturbation characteristics, and spatial feature vectors are generated by pooling dimensionality reduction.

[0106] The disturbance characteristics include pressure wave intensity and thermal disturbance mode, while the spatial mode includes pressure wave distribution and heat transfer path in different areas of the subway station.

[0107] In some embodiments, the deep learning-based energy-saving optimization control method for subway station environmental control and ventilation linkage provided by the present invention extracts spatial patterns of disturbance characteristics by performing convolution operations on the blind source separation result matrix through a convolutional neural network, and generates spatial feature vectors through pooling dimensionality reduction. Specifically, it includes the following steps:

[0108] Step S131: Call the convolutional neural network to identify local patterns in the blind source separation result matrix through convolution operation, and perform convolution calculation based on the defined convolution kernel to obtain the output matrix after convolution.

[0109] Step S132 involves processing the output matrix after convolution using a nonlinear activation function, compressing the matrix size using pooling operations, and straightening the pooled matrix to obtain the spatial feature vector.

[0110] The expression for the spatial feature vector is:

[0111]

[0112] In the formula, Let p be the spatial eigenvector, and p be the length of the total number of elements in the pooled matrix. This is the matrix after pooling.

[0113] Step S140: The spatial feature vector is used as the input to the bidirectional GRU structure in the recurrent neural network to capture the time variation law of dynamic perturbation and generate the hidden state sequence.

[0114] In some embodiments, the deep learning-based energy-saving optimization control method for subway station environmental control and ventilation linkage provided by the present invention uses spatial feature vectors as input to a bidirectional GRU structure in a recurrent neural network to capture the time variation law of dynamic disturbances and generate a hidden state sequence, specifically including the following steps:

[0115] It should be noted that the bidirectional GRU structure includes a forward GRU and a reverse GRU, which are used to control the transmission of information through update gates and reset gates. The update gates and reset gates are used to determine the retained spatial feature information and the forgotten spatial feature information through the Sigmoid activation function.

[0116] Step S141: The spatial feature vector is used as the input of the recurrent neural network to capture the sequential dependencies in the time series through recursive updates, resulting in a time series input matrix. Each time step in the time series input matrix corresponds to a vector.

[0117] Step S142: The time series input matrix is ​​processed by the forward GRU and the reverse GRU in chronological order to obtain the total hidden state output by the bidirectional GRU structure, and the hidden state sequence matrix is ​​generated based on the time variation law of the total hidden state and the dynamic perturbation.

[0118] The total hidden state is composed of the hidden states of the forward GRU and the reverse GRU. The hidden state sequence matrix is ​​used to describe the perturbation characteristics of each time step in the time series.

[0119] Step S150: Based on the hidden state sequence, an attention mechanism is introduced into the output layer of the recurrent neural network to focus on the hidden state when the train enters the station, and an optimized perturbation feature representation is generated by weighting.

[0120] In some embodiments, the deep learning-based energy-saving optimization control method for subway station environmental control and ventilation linkage provided by the present invention, based on the hidden state sequence, introduces an attention mechanism into the output layer of the recurrent neural network to focus on the hidden state when the train enters the station, and generates an optimized disturbance feature representation through weighted averaging, specifically including the following steps:

[0121] Step S151: The hidden state sequence matrix is ​​used as the input to the attention mechanism introduced by the recurrent neural network to perform a linear transformation on the hidden states in the hidden state sequence matrix, mapping the high-dimensional state to the hidden layer representation of the attention weights.

[0122] Step S152: The hidden layer representation is multiplied by the global context vector to calculate the attention weights, and the hidden state sequence matrix is ​​weighted and summed to generate the perturbation feature representation.

[0123] The calculation formula for the disturbance characteristic representation is as follows:

[0124]

[0125] In the formula, This is the perturbation feature representation vector. For time steps, Let be the attention weight at time step t. Let be the hidden state vector at time step t. For global context vectors, This is the weight matrix. The hyperbolic tangent activation function is used. This is the paranoia vector.

[0126] Step S160: The optimized perturbation feature representation is used as the input of the fully connected layer to generate and output the corresponding energy-saving strategy category.

[0127] In some embodiments, the deep learning-based energy-saving optimization control method for subway station environmental control and ventilation linkage provided by the present invention uses the optimized disturbance feature representation as the input of the fully connected layer to generate and output the corresponding energy-saving strategy category, specifically including the following steps:

[0128] Step S161: The perturbation feature representation is used as the input of the fully connected layer, so as to convert the perturbation feature representation into the input of the classification unit through the linear transformation of the fully connected layer and the nonlinear transformation of the activation function.

[0129] Step S163: The classification unit is called to generate corresponding energy-saving strategy categories based on the number of categories output by the fully connected layer. Energy-saving strategy categories whose category probability exceeds a set threshold are selected as the in-station energy-saving optimization control strategies.

[0130] The probability distribution for generating energy-saving strategy categories is calculated using the Softmax function on the output of the fully connected layer, and its expression is as follows:

[0131]

[0132] In the formula, Let c be the probability of class c. The output of the fully connected layer for category c. This represents the total number of categories.

[0133] The aforementioned deep learning-based energy-saving optimization control method for subway station environmental control and ventilation linkage acquires train operation status, station environment data, and environmental control equipment data. These data are preprocessed to obtain a time-series data matrix. Non-negative matrix factorization is then used to decompose the time-series data matrix to extract disturbance characteristics and output a blind source separation result matrix. Subsequently, a convolutional neural network is used to perform a convolution operation on the blind source separation result matrix to extract spatial patterns of disturbance characteristics, and pooling is used to reduce dimensionality and generate spatial feature vectors. These spatial feature vectors are then used as input to a bidirectional GRU structure in a recurrent neural network to capture the temporal variation patterns of dynamic disturbances and generate a hidden state sequence. Based on this generated hidden state sequence, an attention mechanism is introduced into the output layer of the recurrent neural network to focus on the hidden states when the train enters the station, and an optimized disturbance feature representation is generated through weighted summation. Finally, the optimized disturbance feature representation is used as input to a fully connected layer to generate and output the corresponding energy-saving strategy category. This method extracts physical disturbance characteristics through Non-Fast Motion (NMF), combines it with an AI model to handle nonlinear dynamic characteristics, and integrates CNN to capture spatial patterns and RNN to process time series data, forming a more comprehensive feature representation. Furthermore, by introducing an attention mechanism, key disturbance periods can be highlighted, thereby enhancing the model's response to complex disturbances. This reduces the need for manual feature selection and enables direct generation of energy-saving strategies from raw data, achieving more refined and real-time energy-saving optimization control.

[0134] In a specific embodiment, the deep learning-based energy-saving optimization control method for subway station environmental control and ventilation linkage provided by the present invention includes steps 1 to 6:

[0135] Step 1: Data Acquisition and Preprocessing.

[0136] The system collects train operation status (such as speed and arrival time), environmental data (air pressure, temperature and humidity), and environmental control equipment data (fan speed, chilled water flow, and air conditioning cooling load). The collected data is then filtered to remove noise, outliers are removed, and time series are segmented to form a high-quality time series data matrix.

[0137] Specifically, this includes steps 1.1 to 1.4:

[0138] Step 1.1, Deployment of data acquisition equipment.

[0139] Based on the environmental control requirements of the subway station, sensors and equipment are deployed to collect data to reflect the pressure waves and thermal disturbances caused by trains entering and exiting, specifically including environmental data and environmental control equipment data.

[0140] Step 1.2: Data is synchronized with timestamp markers.

[0141] Align the raw data from each acquisition device according to the timestamp, and unify the sampling time interval to Δt seconds. Let the time series be t = t0, t1, ..., t n The time interval satisfies t i+1 -t i =△t, typically taking the value △t=1s, therefore the expression for the collected data vector is:

[0142]

[0143] In the formula, Indicates time The collected data vector, For time Air pressure data, in Pa; For time Temperature data, in °C; For time Humidity data, in % %. For time The fan speed data, in RPM; For time Cold water flow rate data, in m³ / h; For time The cooling load data is in kW.

[0144] Step 1.3, Data denoising and outlier handling.

[0145] Specifically, for Perform filtering and outlier handling:

[0146] Low-pass filtering: filters out high-frequency noise, while using a band-pass filter to preserve the frequency range. to The signal.

[0147] Outlier removal: The three-standard-deviation rule is used. Outliers are then removed and replaced with linear interpolation. In the formula, The mean of the sequence. is the standard deviation of the sequence.

[0148] Step 1.4, Time Series Segmentation.

[0149] right Perform time window slicing, with each segment having a length of L and corresponding to M sampling points, i.e., L = M × Δt. Store the segmented event sequence in the following matrix form:

[0150]

[0151] In the formula, The segmented time series data matrix, To represent the amount of data at each sampling point, in a time series data matrix, each row represents the time series of a single variable (such as air pressure, temperature, etc.), and each column represents a specific time point. All variables.

[0152] Step 2: Data decomposition and blind source feature extraction.

[0153] Nonnegative matrix factorization (NMF) is used to decompose the preprocessed data, extract key perturbation characteristics (such as pressure wave intensity and thermal perturbation mode), and output the blind source separation result matrix.

[0154] Specifically, this includes steps 2.1 to 2.3:

[0155] Step 2.1 introduces nonnegative matrix factorization (NMF).

[0156] It should be noted that Nonnegative Matrix Factorization (NMF) is a method that decomposes a nonnegative matrix into the product of two nonnegative matrices, which is used to extract hidden features, i.e., blind source separation.

[0157] Specifically, let the input be a time series data matrix. Its dimension is defined as m×n, and it can be decomposed into the following form using NMF:

[0158]

[0159] In the formula, This is a time series data matrix with the defined dimension m×n; The feature basis matrix has dimensions m×r and is used to represent the weights of the r feature bases, with each row corresponding to m variables; The activation matrix has dimensions r×n and is used to characterize the strength of each feature basis at n time points; r is the number of features in the decomposition, and satisfies... .

[0160] In this embodiment, the decomposition objective of NMF is to minimize the matrix reconstruction error by optimizing the following objective function:

[0161]

[0162] In the formula, Let f(x) be the Frobenius norm, and let f(x) be the square root of the sum of the squares of the matrix elements. , used for non-negativity constraints on matrix elements.

[0163] Update using iterative optimization methods (such as multiplicative update rules). and :

[0164]

[0165] Step 2.2: Extract key perturbation characteristics.

[0166] The NMF decomposition results are analyzed, and key perturbation characteristics are extracted from the activation function H:

[0167] Each row of H represents the intensity change of a characteristic basis over time. By observing the location and distribution of the maximum values ​​in H, key disturbance moments, such as pressure wave peaks and heat flow peaks, can be determined. Each column in the characteristic basis matrix W represents the contribution of different variables to a certain characteristic basis. By comparing the column vectors of W, variables with high contributions, such as fan speed and air pressure, can be extracted.

[0168] In this embodiment, the expression for the normalization of the activation matrix is:

[0169]

[0170] In the formula, This is the normalized activation matrix. This represents the maximum value of each row in the activation matrix H.

[0171] The eigenvalue basis matrix W is normalized to highlight the contribution of variables. The expression for normalizing the eigenvalue basis matrix is ​​as follows:

[0172]

[0173] In the formula, The normalized eigenbase matrix is ​​formed by normalizing the sum of each column. This represents the column sum in the characteristic basis matrix.

[0174] Step 2.3: Construct the blind source separation result matrix.

[0175] Combination and Reconstruct the blind source separation result matrix The reconstruction expression for the blind source separation result matrix is:

[0176]

[0177] In the formula, This is a blind source separation result matrix with dimensions m×n. Each column of the blind source separation result matrix represents a time point. The corresponding blind source characteristics contain key perturbation information.

[0178] Blind source separation result matrix The element meaning is represented as:

[0179]

[0180] In the formula, For the i-th variable at time t j The blind source characteristic value, Let be the contribution value of the i-th variable to the characteristic basis r. The characteristic basis r at time point t j The activation value.

[0181] Step 3, spatial feature extraction.

[0182] Convolutional neural networks (CNNs) are used to perform convolution operations on the blind source separation result matrix to extract spatial patterns of perturbation characteristics, such as pressure wave distribution and heat transfer paths in different regions. Spatial feature vectors are then generated by pooling dimensionality reduction.

[0183] Specifically, this includes steps 3.1 through 3.5:

[0184] Step 3.1, introduce a convolutional neural network (CNN).

[0185] A convolutional neural network (CNN) is a deep learning model used to extract spatial characteristics of data. It identifies local patterns in the blind source separation result matrix S through convolution operations, such as pressure wave distribution and heat transfer paths.

[0186] In this embodiment, This indicates that the i-th variable at time point t j The perturbation characteristics are given by the input matrix S, which is the blind source separation result matrix output from step 2, and has dimensions m×n.

[0187]

[0188] Step 3.2, convolution operation.

[0189] Define a convolution kernel K of size p×q:

[0190]

[0191] Convolution kernel weights This indicates that it is initialized with random values ​​and adjusted through training.

[0192] The formula for calculating convolution is:

[0193]

[0194] In the formula, The output after convolution represents the nth... Feature values ​​at each position, This is the starting index of the convolution window. The kernel stride 's' determines the step size for each movement; a typical value is s=1.

[0195] Step 3.3, Activation function.

[0196] The matrix after convolution After processing with a nonlinear activation function (such as ReLU), the nonlinear expressive power of the model is increased. The activation expression is:

[0197]

[0198] In the formula, Let be the activated matrix element, representing the _th Nonlinear characteristics of location.

[0199] Step 3.4, pooling dimensionality reduction.

[0200] Pooling operations extract salient features from local regions by compressing the matrix size, while reducing computational cost. The pooling window size is defined as p. p ×q p The typical value is 2×2. The pooling operation formula takes max pooling as an example:

[0201]

[0202] In the formula, These are the elements of the pooled matrix. This refers to the index within the pooling window. The pooling stride is s. p The pooling window step size is determined by the number of steps, typically s. p =2.

[0203] Step 3.5, Spatial feature vector generation.

[0204] The pooled matrix Straighten to spatial feature vectors The expression is:

[0205]

[0206] In the formula, Let p be the spatial eigenvector, and p be the length of the total number of elements in the pooled matrix. This is the pooled matrix. The pooling matrix is ​​the first... Each element. The final output spatial feature vector describes the main spatial pattern features extracted from the blind source separation matrix.

[0207] Step 4, Time Series Modeling.

[0208] Spatial feature vectors are input into a bidirectional GRU structure in a recurrent neural network (RNN) to capture the temporal variation of dynamic disturbances, such as the propagation speed of pressure waves and the duration of thermal disturbances, and generate hidden state sequences.

[0209] Specifically, this includes steps 4.1 to 4.4:

[0210] Step 4.1, introduce a recurrent neural network (RNN).

[0211] Recurrent Neural Networks (RNNs) are deep learning models used to process sequential data. They capture the dependencies between time series data through recursive updates, with the spatial feature vector generated in step 3 as input. At each moment t i Corresponding to a vector x t ,Right now:

[0212]

[0213] in, The input matrix is ​​a time series matrix with dimensions T×d, where T is the number of time steps, d is the dimension of the feature vectors, and x... t Let be the input vector at time t.

[0214] Step 4.2, bidirectional GRU structure.

[0215] GRU (Gated Recurrent Unit) is an improved RNN unit that uses update gates and reset gates to control the transmission of information. A bidirectional GRU structure consists of two independent GRU units:

[0216] Forward GRU: Processes the sequence sequentially from t=1 to t=T.

[0217] Reverse GRU: Processes the sequence sequentially from t=T to t=1.

[0218] The state update formula for each GRU unit is as follows:

[0219] The update gate determines how much previous information to retain; the expression is:

[0220]

[0221] The reset gate determines how much previous information is forgotten; the expression is:

[0222]

[0223] Candidate state: Calculate the candidate hidden state at the current time step, its expression is:

[0224]

[0225] The implicit state update expression is:

[0226]

[0227] in, , The outputs for the update gate and reset gate are respectively, with a range of [range value missing]. , Let be the hidden state vector at the current time step. These are candidate latent states. , , These are the weight matrix, the randomly initialized matrix, and the matrix learned through training, respectively. The Sigmoid activation function has a range of [range missing]. , It is a hyperbolic tangent function, with a range of . .

[0228] The final output of the bidirectional GRU is:

[0229]

[0230] In the formula, This represents the implicit state of a positive GRU. This is the implicit state of the reverse GRU.

[0231] Step 4.3: Capture the temporal variation pattern of dynamic disturbances.

[0232] The event variation patterns of dynamic disturbances (such as pressure wave propagation and thermal disturbance duration) are captured through the implicit state sequence of the bidirectional GRU:

[0233]

[0234] in, , representing the implicit state at time t. The hidden state sequence matrix has dimension 1. , This represents the hidden state dimension of a GRU cell.

[0235] By using the time dimension of H, we can capture the propagation pattern of disturbances over time, for example:

[0236] Pressure wave propagation: Identifying the intensity and propagation speed of waves from the changing trends of implicit states.

[0237] Thermal disturbance duration: The duration of the disturbance effect is determined by the degree of activation of the continuous hidden states.

[0238] Step 4.4, hidden state sequence representation.

[0239] The final hidden state sequence matrix H describes the perturbation characteristics at each time step in the time series:

[0240]

[0241] in, Let be the i-th hidden state value at time step t, representing the contribution of that dimension to the perturbation at the current time step. For time steps, This represents the implicit state dimension of a bidirectional RGU.

[0242] Step 5: Optimize the attention mechanism.

[0243] Based on the hidden state sequence, an attention mechanism is introduced into the output layer of the RNN to focus on the hidden state at key time points (such as the moment the train enters the station) and generate an optimized perturbation feature representation through weighted summation.

[0244] Specifically, this includes steps 5.1 to 5.5:

[0245] Step 5.1, introduce the attention mechanism.

[0246] Attention mechanism is a technique for optimizing sequence modeling. It highlights the contribution of hidden states at key time points by calculating the importance weight of each time step in the time series. Its input is the hidden state sequence matrix H generated in step 4.

[0247] Step 5.2: Calculate the weights of key time points.

[0248] First, perform a hidden state transformation, that is, transform the hidden state... By applying a linear transformation, the high-dimensional state is mapped to the hidden layer representation of the attention weights:

[0249]

[0250] In the formula, Let be the hidden layer representation vector, with dimension . , This is the weight matrix, with dimension 1. , The paranoia vector has dimension . , This is the hyperbolic tangent activation function, used to introduce nonlinearity.

[0251] Secondly, the weight vector is calculated, and With global context vector Inner product, calculating attention score :

[0252]

[0253] In the formula, The global context vector has dimensions of . Used to measure the correlation of time steps. Let be the attention weight at time step t, with a range of . And all The sum is 1.

[0254] Step 5.3: Calculate the perturbation feature representation.

[0255] The perturbation feature representation is the result of a weighted summation of the hidden state sequence H, highlighting the contribution at key time points. Its expression is:

[0256]

[0257] In the formula, Let be the perturbation feature representation vector, with dimension . , Let be the hidden state vector at time step t. Let be the attention weight at time step t.

[0258] Final output It represents the optimized characteristics of the entire time series and can capture the impact of key time points on disturbances.

[0259] Step 5.4, describe the perturbation feature representation.

[0260] Perturbation feature representation The specific form is:

[0261]

[0262] The i-th eigenvalue of the perturbation feature representation vector is a feature component obtained by time-weighted summation, such as:

[0263]

[0264] in, This represents the i-th hidden state value at time step t.

[0265] Step 5.5, summarize the optimization of the attention mechanism.

[0266] In the implicit state of key time points, the key time points are weighted through the attention mechanism. Enhance its hidden state The contribution of this data is used to achieve optimization. During the process of capturing dynamic disturbances, through... Weighted Implicit State Disturbance characteristics It can accurately capture the dynamic changes at different time steps in a time series.

[0267] Therefore, the final formula for optimizing the attention mechanism is as follows:

[0268]

[0269] In the formula, This is the perturbation feature representation vector. For time steps, Let be the attention weight at time step t. Let be the hidden state vector at time step t. For global context vectors, This is the weight matrix. The hyperbolic tangent activation function is used. This is the paranoia vector.

[0270] Step 6: Energy-saving strategy generation and classification.

[0271] The optimized feature representation is input into the fully connected layer, and combined with the classification module to generate specific energy-saving strategy categories (such as adjusting fan speed and optimizing chilled water supply).

[0272] Input: Optimized perturbation feature representation.

[0273] Output: Energy-saving strategy category.

[0274] Specifically, this includes steps 6.1 to 6.3:

[0275] Step 6.1, introduce a fully connected layer.

[0276] A fully connected layer is a neural network layer used to combine input features; its role is to combine the optimized feature representations. This is converted into the input of the classification module, whose input is the perturbation feature representation output from step 5, with dimensions of [missing information]. By employing linear transformations of the fully connected layer and nonlinear transformations of the activation function, the input features are represented as the input to the classification module, with the following formula:

[0277]

[0278] in, This is the output of the fully connected layer, with dimension C, representing the number of categories in the classification module. This is the weight matrix, with dimension 1. Used for feature weighting. Let C be the paranoia vector.

[0279] Step 6.2, Classification Module.

[0280] The classification module is used to classify according to The value generates the energy-saving strategy category, which refers to a specific control and adjustment method, such as adjusting the fan speed or optimizing the chilled water supply.

[0281] In this embodiment, the classification module uses the Softmax classifier to call the Softmax function to output the fully connected layer. Transform into probability distributions for each category:

[0282]

[0283] In the formula, Let c be the probability of category c, with a range of 1 / 2. , The output of the fully connected layer for category c. This represents the total number of categories.

[0284] Then, a category determination is performed, and the category with the highest probability is selected as the final energy-saving strategy:

[0285]

[0286] in, For the index of predicted energy-saving strategy categories, To return The category index corresponding to the maximum value.

[0287] Step 6.3, Energy-saving strategy category definition.

[0288] Energy saving strategy categories Describe the specific control and adjustment behaviors, for example:

[0289] Category c=1: Increase fan speed.

[0290] Category c=2: Reduce fan speed.

[0291] Category c=3: Optimize cold water supply.

[0292] Category c=4: Adjust the air conditioning cooling load. ......

[0294] The classification results are associated with specific control devices and are executed through the device interface.

[0295] Beneficial effects:

[0296] Combining mechanism and data-driven approaches: Physical perturbation characteristics are extracted through NMF, and nonlinear dynamic characteristics are processed by combining AI models.

[0297] Joint modeling of spatial-temporal characteristics: CNN captures spatial patterns, RNN processes time series, forming a comprehensive feature representation.

[0298] Attention optimization: Highlighting key perturbation periods to enhance the model's ability to respond to complex perturbations.

[0299] End-to-end optimization: Reduce manual feature selection and generate energy-saving strategies directly from raw data.

[0300] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0301] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0302] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0303] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0304] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0305] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0306] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0307] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0308] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A deep learning-based energy-saving optimization control method for air-water linkage of environmental control in a subway station, characterized in that, The method comprises: acquiring train operation state, station environment data and air conditioning equipment data, and preprocessing the train operation state, station environment data and air conditioning equipment data to obtain a time series data matrix; wherein the train operation state comprises train speed and train arrival time, the station environment data comprises air pressure, temperature and humidity, and the air conditioning equipment data comprises fan speed, cold water flow and air conditioning refrigeration load; performing non-negative matrix decomposition on the time series data matrix to extract disturbance characteristics in the time series data matrix and output a blind source separation result matrix; the disturbance characteristics comprise pressure wave intensity and thermal disturbance mode; the non-negative matrix decomposition on the time series data matrix to extract disturbance characteristics in the time series data matrix comprises: analyzing the non-negative matrix decomposition result, extracting key disturbance characteristics from an activation function, each row of the activation function representing the intensity change of a characteristic base over time, and determining the key disturbance time by observing the position and distribution of the maximum value in the activation function, the key disturbance time comprising pressure wave peak value and thermal flow peak value; performing convolution operation on the blind source separation result matrix by a convolutional neural network to extract spatial modes of the disturbance characteristics, and generating a spatial feature vector by dimension reduction through pooling; the spatial modes comprise pressure wave distribution and heat transfer path in different areas of the subway station; inputting the spatial feature vector into a bidirectional GRU structure in a recurrent neural network as input to capture the time variation law of dynamic disturbance and generate a hidden state sequence; the time variation law of dynamic disturbance comprises: pressure wave transmission speed and thermal disturbance duration; the time variation law of dynamic disturbance is captured by the hidden state sequence of the bidirectional GRU: wherein, denotes the hidden state at time t, t = 1, 2,..., T, is a hidden state sequence matrix of dimension , is the hidden state dimension of the GRU unit, is the hidden state of the forward GRU, is the hidden state of the backward GRU; the propagation law of disturbance over time is captured through the time dimension of H, including identifying the fluctuation intensity and propagation speed of the pressure wave from the change trend of the hidden state, and determining the duration of the disturbance influence by the activation degree of the continuous hidden state; based on the hidden state sequence, introducing an attention mechanism in the output layer of the recurrent neural network to focus on the hidden state when the train enters the station, and generating an optimized disturbance feature representation by weight weighting; inputting the optimized disturbance feature representation into a fully connected layer as input to generate and output a corresponding energy-saving strategy category, the energy-saving strategy category comprising: adjusting the fan speed, optimizing the cold water flow and adjusting the air conditioning refrigeration load.

2. The deep learning-based energy-saving optimization control method for air-water linkage of a subway station according to claim 1, characterized in that, The station environment data and air conditioning equipment data are data collected by sensors or equipment arranged according to the air conditioning requirements of the subway station, for reflecting the pressure wave and thermal disturbance caused by the arrival and departure of the train; the acquisition of train operation state, station environment data and air conditioning equipment data, and the preprocessing of the train operation state, station environment data and air conditioning equipment data to obtain a time series data matrix, comprises: aligning the station environment data and air conditioning equipment data according to the time stamp by using the same time interval to obtain a collection data vector, the expression of the collection data vector being: In the formula, indicates the acquisition data vector at time , is the air pressure data at time , is the temperature data at time , is the humidity data at time , is the fan speed data at time , is the cold water flow data at time , is the refrigeration load data at time ; The high-frequency noise in the collected data vector is filtered out by low-pass filtering, while the signal in a preset frequency range is retained by a band-pass filter, and the abnormal values are removed by a three-sigma rule, and the removed abnormal values are replaced by linear interpolation; The collected data vector after filtering and denoising and abnormal value removal is time window sliced, and the segmented time series is stored in a matrix form to obtain a time series data matrix, and an expression of the time series data matrix is: In the formula, is the time series data matrix after segmentation, is the data amount of the sampling point, in the time series data matrix, each row represents the time series of a variable, and each column represents all variables at any time point .

3. The deep learning-based energy-saving optimization control method for the air-water linkage of the environmental control of a subway station according to claim 2, characterized in that, The time series data matrix is decomposed by the non-negative matrix factorization to extract the disturbance characteristics in the time series data matrix, and a blind source separation result matrix is output, including: The definition dimension of the time series data matrix is obtained, and the time series data matrix is decomposed based on the definition dimension by the non-negative matrix factorization, and an expression of the decomposition is: In the formula, is the time series data matrix, with dimensions defined as m x n; is a feature basis matrix, with dimensions m x r, used to represent the weights of r feature bases, and each row corresponds to m variables; is an activation matrix, with dimensions r x n, used to represent the intensity of each feature base at n time points; r is the number of features decomposed, and satisfies ; The matrix reconstruction error is minimized by optimizing the objective function, and the feature basis matrix and the activation matrix are updated by an iterative optimization algorithm, and the iterative optimization algorithm includes a multiplication update rule, and an expression of the objective function is: wherein is the Frobenius norm, representing the square root of the sum of the squares of the matrix elements, and , for non-negativity constraints on the matrix elements.

4. The deep learning-based metro station HVAC and air conditioning water linkage energy-saving optimization control method according to claim 3, characterized in that, The time series data matrix is decomposed by the non-negative matrix factorization to extract the disturbance characteristics in the time series data matrix, and a blind source separation result matrix is output, further including: The decomposition result of the time series data matrix by the non-negative matrix factorization is analyzed, and the activation matrix and the feature basis matrix are normalized to extract a key disturbance time from the activation matrix based on the maximum value and distribution of variables in the activation matrix, and to extract a key variable by comparing column vectors of the feature basis matrix; The blind source separation result matrix is reconstructed and output based on the normalized activation matrix and feature basis matrix; An expression of the normalization processing of the activation matrix is: wherein is the normalized activation matrix, denotes the maximum value of each row in the activation matrix; An expression of the normalization processing of the feature basis matrix is: In the formula, is the normalized feature basis matrix, denotes a column in the feature basis matrix. A reconstruction expression of the blind source separation result matrix is: In the formula, is the blind source separation result matrix, with a dimension of mxn, and each row of the blind source separation result matrix represents a time point Corresponding blind source characteristics contain key disturbance information.

5. The deep learning-based metro station HVAC and air conditioning water linkage energy-saving optimization control method according to claim 4, characterized in that, The blind source separation result matrix is convolved by the convolutional neural network to extract the spatial mode of the disturbance characteristics, and a spatial feature vector is generated by pooling dimension reduction, including: The convolutional neural network is called to identify the local mode in the blind source separation result matrix by convolution operation, and a convolution kernel is defined to perform convolution calculation to obtain a convolution output matrix; The convolution output matrix is processed by a nonlinear activation function, and the matrix size is compressed by a pooling operation, and the pooled matrix is flattened to obtain the spatial feature vector; An expression of the spatial feature vector is: In the formula, is a spatial feature vector, and the length is the total number of elements p x q of the pooled matrix, is a pooled matrix.

6. The deep learning-based metro station HVAC and air conditioning water linkage energy-saving optimization control method according to claim 5, characterized in that, The bidirectional GRU structure includes a forward GRU and a reverse GRU, which are used to control the transmission of information by update gates and reset gates, and the update gates and the reset gates are used to determine the retained spatial feature information and the forgotten spatial feature information by a Sigmoid activation function; The spatial feature vector is input into the bidirectional GRU structure in the recurrent neural network to capture the time variation law of the dynamic disturbance and generate a hidden state sequence, including: The spatial feature vector is taken as an input of the recurrent neural network to capture the forward and backward dependencies in the time series in a recursive update manner, so as to obtain a time series input matrix, each time corresponding to a vector in the time series input matrix; The time series input matrix is processed in a time sequence by the forward GRU and the backward GRU respectively to obtain a total hidden state output by the bidirectional GRU structure, and a hidden state sequence matrix is generated based on the total hidden state and a time variation law of dynamic disturbance; The total hidden state is composed of the hidden state of the forward GRU and the hidden state of the backward GRU, and the hidden state sequence matrix is used to describe the disturbance characteristics of each time step in the time series.

7. The deep learning-based metro station HVAC and air conditioning water linkage energy-saving optimization control method according to claim 6, characterized in that, The attention mechanism is introduced into the output layer of the recurrent neural network based on the hidden state sequence to focus on the hidden state when the train enters the station, and an optimized disturbance feature representation is generated by weighting, including: The hidden state sequence matrix is taken as an input of the attention mechanism introduced by the recurrent neural network to perform linear transformation on the hidden state in the hidden state sequence matrix, and map the high-dimensional state to a hidden layer representation of the attention weight; The inner product of the hidden layer representation and a global context vector is calculated to obtain the attention weight, and the hidden state sequence matrix is weighted and summed to calculate and generate the disturbance feature representation; The calculation formula of the disturbance feature representation is: wherein, is a perturbation feature representation vector, is a time step, is an attention weight at the t-th time step, is a hidden state vector at the t-th time step, is a global context vector, is a weight matrix, is a hyperbolic tangent activation function, is a bias vector.

8. The deep learning-based metro station HVAC and air conditioning water linkage energy-saving optimization control method according to claim 7, characterized in that, The optimized disturbance feature representation is taken as an input of the full connection layer to generate and output a corresponding energy-saving strategy category, including: The disturbance feature representation is taken as an input of the full connection layer to convert the disturbance feature representation into an input of a classification unit through linear transformation of the full connection layer and nonlinear transformation of an activation function; The classification unit is called to generate a corresponding energy-saving strategy category according to the number of categories output by the full connection layer, so as to select an energy-saving strategy category with a category probability exceeding a set threshold as a station energy-saving optimization control strategy based on the energy-saving strategy category; The probability distribution of the energy-saving strategy category is obtained by calculating the output of the full connection layer by a Softmax function, and the expression is: wherein is the probability of class c, is the output of the fully connected layer for class c, is the total number of classes.

9. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is used to operate according to the instructions to perform the steps of the method according to any one of claims 1-8.

10. A computer readable storage medium having stored thereon a computer program, characterized in that The program is executed by the processor to implement the steps of the method according to any one of claims 1-8.

Images