A deep learning-based spatiotemporal prediction method for complex construction ventilation environment of underground cavern groups

Abstract

The application discloses a kind of underground cavern group complex construction ventilation environment space-time prediction method based on deep learning, including data acquisition, data cleaning, data noise reduction and normalization processing: construct and include sequence perception global Token generation module, graph structure learning module, graph aggregation module, time coding module and trend perception attention module's space-time prediction model: normalized data is divided into training set, verification set and test set, constructs joint loss function, utilizes optimization algorithm to train the space-time prediction model, and adopts swarm intelligence optimization algorithm to automatically optimize hyperparameter;Real-time monitoring data is input after pre-processing into trained space-time prediction model, and the wind speed, dust concentration prediction value of future time step is output, and the operating frequency of fan of ventilation system is dynamically adjusted according to prediction value.

Description

Technical Field

[0001] This invention belongs to the field of water conservancy and hydropower engineering construction technology, specifically involving a spatiotemporal prediction method for complex construction ventilation environment of underground cavern groups based on deep learning. Background Technology

[0002] In water conservancy and hydropower projects, the excavation of large underground cavern complexes, such as underground powerhouses and ventilation chambers in hydropower stations, using the drill-and-blast method generates large amounts of dust and harmful gases. These pollutants pose serious safety hazards to construction workers. The vast spaces and complex structures of underground cavern complexes increase the distance dust travels, and the winding ventilation paths easily lead to dust retention and swirling, making effective removal difficult. Dynamic changes in ventilation environmental parameters directly affect construction safety and progress.

[0003] Current technologies primarily rely on real-time monitoring and reactive response strategies: sensors are deployed at key locations to monitor pollutant concentrations and ventilation strategies are adjusted accordingly. However, this traditional "real-time monitoring first, then taking action" approach is inherently delayed and cannot effectively prevent safety accidents caused by excessive pollutant concentrations. Furthermore, monitoring data is often affected by factors such as mechanical vibrations from blasting impacts and unstable signal transmission, resulting in high-frequency noise and strong fluctuations. Random high-frequency noise distorts the data, while local fluctuations caused by changes in construction conditions, such as dust from manual or mechanical operations or blasting dust, contain crucial information. However, existing methods often treat both as noise, leading to information loss.

[0004] While mainstream computational fluid dynamics numerical simulation methods offer physical interpretability, they rely on empirical assumptions, simplify boundary conditions, and incur high computational costs, failing to capture dynamic factors and meet real-time prediction requirements. Traditional time-series prediction models have limited ability to model nonlinear dynamic responses and multi-sensor spatiotemporal dependencies, and lack robustness when dealing with highly volatile and non-stationary data. Although deep learning methods introduced in recent years have made progress, they lack preprocessing strategies for high-frequency noise and strong fluctuations, and cannot fully capture the spatiotemporal correlations of multi-sensor networks.

[0005] Therefore, a robust spatiotemporal prediction method is needed to predict the ventilation environment of underground cavern groups in advance, identify potential pollution peaks, support timely ventilation, dust reduction or personnel evacuation and other measures, and ensure construction safety, reduce energy consumption and guarantee progress. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and, addressing the noise interference, data fluctuations, and incompleteness of monitoring data in complex construction environments, provide a deep learning-based spatiotemporal prediction method for ventilation environments in complex underground cavern complexes. By integrating advanced noise reduction technologies and an improved deep learning model, the method enhances the accuracy and robustness of predictions for parameters such as dust concentration and wind speed. This method achieves stable predictions under complex working conditions, addressing high-frequency noise, local fluctuations, and distribution shifts in monitoring data. It provides a scientific basis for decision-making in intelligent ventilation systems, ensuring the safe and efficient operation of the construction environment. Furthermore, this method is applicable to engineering scenarios such as underground powerhouses in hydropower stations and tunnel excavation, aiming to improve the real-time monitoring and prediction capabilities of the construction environment.

[0007] The objective of this invention is achieved through the following technical solution: A spatiotemporal prediction method for complex construction ventilation environment in underground cavern groups based on deep learning, comprising the following steps: S1. Data Acquisition and Preprocessing: Several sensors were deployed along the ventilation paths and in the work areas of the underground cavern complex to collect time-series data on wind speed and dust concentration. After cleaning the time-series data, a composite noise reduction model was used for processing. The composite noise reduction model processing included: decomposing the time series into modal functions and residuals; performing anomaly detection on the high-frequency modal functions to obtain anomaly point sets; and distinguishing between isolated noise and clustered fluctuations in the anomaly point sets based on density clustering, retaining clustered fluctuations and removing isolated noise to obtain the denoised sequence; finally, the denoised sequence was subjected to Reversible Instance Normalization (RevIN) processing to generate normalized data. S2. Construct the ISageformer prediction model: A spatiotemporal prediction model is constructed, comprising a sequence-aware global token generation module, a graph structure learning module, a graph aggregation module, a time encoding module, and a trend-aware attention module; wherein, the spatiotemporal prediction model integrates a reversible instance normalization RevIN mechanism at both the input and output ends to handle data distribution shifts; S3. Model Training and Optimization: Normalized data is divided into training set, validation set and test set, a joint loss function is constructed, the spatiotemporal prediction model is trained using optimization algorithm, and the hyperparameters of the composite noise reduction model in step S1 are automatically optimized using swarm intelligence optimization algorithm. S4. Prediction and Application: After preprocessing the real-time monitoring data in step S1, it is input into the trained spatiotemporal prediction model, which outputs the predicted values ​​of wind speed and dust concentration for future time steps, and dynamically adjusts the operating frequency of the ventilation system's fans based on the predicted values.

[0008] Furthermore, in step S1, the FOX-VMD-IF-DBSCAN model is used as the composite noise reduction model; specifically, it includes: Variational Mode Decomposition (VMD) is used to analyze the cleaned time series. Decomposed into K intrinsic mode functions and residual ; The Isolation Forest IF algorithm is used to detect outliers in high-frequency mode functions and generate an outlier set. ; Using the DBSCAN algorithm A secondary discrimination is performed, calculating the density of data points. If the density is greater than a set threshold, it is determined to be a clustered fluctuation function. Those that are retained will be retained; otherwise, they will be considered isolated noise and removed. Reconstructing the denoised sequence: ; This is the denoised sequence. This is the preserved clustered wave function.

[0009] Furthermore, the FOX optimization algorithm is used to determine the number of modes K in variational mode decomposition and the outlier ratio in the isolated forest algorithm. Neighborhood radius of the DBSCAN algorithm The optimization is performed; the objective function of the optimization is to minimize the permutation entropy PE of the denoised sequence; the position update formula of the FOX algorithm is: ;in, Where T is the population location and T is the target location. and Here, r is the control parameter, and r is a random vector.

[0010] Furthermore, in step S2, the sequence-aware global token generation module normalizes the input... Add a learnable global token G, and generate the initial embedding using the following formula. , ; The graph structure learning module generates an adjacency matrix A by learning the dependencies between nodes using a multilayer perceptron (MLP). .

[0011] Furthermore, in step S2, the processing procedure of the trend-aware attention module is as follows: Enhanced embedding by extracting graphs using causal convolution Local trends : ; in, This represents the graph-enhanced embedding features obtained after processing by the graph structure learning module and the graph aggregation module; This represents a one-dimensional causal convolution operation; k represents the kernel size of the causal convolution. This represents the local trend features extracted by causal convolution; Calculate the query matrix based on local trends. Key matrix Sum matrix And calculate the attention score: ; Where d is the embedding dimension.

[0012] Furthermore, in step S1, the reversible instance normalization RevIN includes normalization and denormalization; Standardization process: ; Denormalization process: ; in, This is the denoised sequence. For normalized sequences, For inverse normalized sequences, The mean of the sequence. The standard deviation of the sequence. Learnable affine parameters; generate normalized data .

[0013] Furthermore, in step S3, the joint loss function L is composed of a weighted average of the mean squared error (MSE) and the mean absolute error (MAE): , These are the weighting coefficients; In step S4, the specific steps for dynamically adjusting the operating parameters of the ventilation system based on the predicted values ​​are as follows: Based on dust concentration prediction values and target concentration Calculate the fan operating frequency f: ; Where k is the scaling factor; when the dust concentration is predicted... An alarm is triggered when the threshold is exceeded.

[0014] This invention also provides a deep learning-based spatiotemporal prediction device for complex construction ventilation environments in underground cavern groups, comprising: The data acquisition and preprocessing module is used to deploy several sensors along the ventilation path and work area of ​​the underground cavern complex to collect time-series data on wind speed and dust concentration. After cleaning the time-series data, a composite noise reduction model is used for processing. The composite noise reduction model processing includes: decomposing the time series into modal functions and residuals; performing anomaly detection on the high-frequency modal functions to obtain anomaly point sets; and distinguishing between isolated noise and clustered fluctuations in the anomaly point sets based on density clustering. Clustered fluctuations are retained while isolated noise is removed to obtain the denoised sequence. Finally, the denoised sequence is subjected to Reversible Instance Normalization (RevIN) processing to generate normalized data. The ISageformer prediction model building module is used to construct a spatiotemporal prediction model that includes a sequence-aware global token generation module, a graph structure learning module, a graph aggregation module, a time encoding module, and a trend-aware attention module. The spatiotemporal prediction model integrates a reversible instance normalization RevIN mechanism at both the input and output ends to handle data distribution shifts. The model training and optimization module is used to divide the normalized data into training set, validation set and test set, construct joint loss function, train the spatiotemporal prediction model using optimization algorithm, and automatically optimize the hyperparameters of the composite noise reduction model in step S1 using swarm intelligence optimization algorithm. The prediction and application module is used to input the real-time monitoring data into the trained spatiotemporal prediction model after preprocessing in step S1, output the predicted values ​​of wind speed and dust concentration for future time steps, and dynamically adjust the operating frequency of the ventilation system's fans based on the predicted values.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the deep learning-based spatiotemporal prediction method for complex construction ventilation environment of underground cavern groups.

[0016] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the deep learning-based method for spatiotemporal prediction of ventilation environment in complex construction of underground cavern groups.

[0017] Compared with the prior art, the beneficial effects of the technical solution of the present invention are as follows: 1. Existing technologies often misclassify high-frequency data fluctuations caused by construction as noise, leading to the loss of critical information. This invention introduces a FOX-VMD-IF-DBSCAN composite noise reduction model; it decomposes the signal through VMD, detects anomalies using IF, and innovatively introduces DBSCAN for secondary discrimination based on density differences. This enables the model to accurately distinguish between meaningless sensor random noise (isolated points) and meaningful construction activity fluctuations (clustered points), effectively reducing noise while fully preserving local fluctuation information reflecting the construction status. Finally, by accurately modeling the complex spatiotemporal relationships between multiple sensors, dynamic features are captured, improving the ability to process high-noise and incomplete data. Experiments show that this strategy improves prediction accuracy by an average of 38% in Gaussian and Lévy noise environments, significantly enhancing prediction accuracy and anti-interference capabilities under complex working conditions.

[0018] 2. To address the issue of data distribution drift over time due to changes in the construction progress of underground engineering projects, this invention introduces Reversible Instance Normalization (RevIN) at the model input and output to dynamically align the data distribution; simultaneously, it combines causal convolution in the ISageformer prediction model to extract local trends. This combination effectively solves the performance degradation problem of traditional models under non-stationary, highly volatile data, enabling them to maintain stable predictive performance even in the face of blasting impacts, mechanical disturbances, or even partial data loss (e.g., a 2% missing rate). It enhances the model's robustness to non-stationary data and distribution shifts, enabling it to adapt to the complex working conditions of underground cavern groups, such as blasting impacts and mechanical disturbances.

[0019] 3. By introducing global tokens and adaptive graph structure learning, this invention can not only capture the temporal characteristics of a single sensor, but also automatically learn and model the dynamic topological relationships between sensors in different spatial locations (such as the directional migration of dust along the ventilation path and the periodic impact of construction activities). This overcomes the shortcomings of traditional methods in handling the spatial correlation of complex cavern groups, achieving accurate modeling of the spatiotemporal dependencies of multiple sensors; the prediction results support variable frequency operation of fans, dynamic adjustment of ventilation strategies, reduced energy consumption, and improved system economy and operating efficiency. Furthermore, it is applicable to the monitoring and prediction of various pollutants and parameters.

[0020] 4. By using the FOX optimization algorithm to minimize permutation entropy, hyperparameters are automatically optimized, avoiding the subjectivity and inefficiency of manual parameter tuning. It provides real-time early warning and optimization guidance, improving the universality and automation level of the method in different engineering scenarios (such as tunnels and coal mines) and lowering the application threshold.

[0021] 5. Based on high-precision advance prediction (e.g., 30 minutes in advance), this method can guide the variable frequency operation of fans. While ensuring that dust concentration does not exceed standards, it avoids the energy waste caused by traditional full-speed operation or delayed adjustment. Examples show that it can reduce energy consumption by 15%, achieving intelligent ventilation decisions and improving energy efficiency. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the method of the present invention. Detailed Implementation

[0023] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0024] Example 1 This embodiment provides a spatiotemporal prediction method for complex construction ventilation environments in underground cavern groups based on deep learning. (See...) Figure 1 Specifically, it includes: S1. Data Acquisition and Preprocessing: S101. Data Acquisition: Monitoring data on ventilation environment of underground cavern complexes exhibits spatiotemporal correlation, including dust concentration. and wind speed The data exhibits directional migration along the ventilation path and shows periodicity and fluctuation due to construction activities. A multi-sensor network is deployed along key ventilation paths, such as ventilation / safety tunnels, main transformer exhaust tunnels, and exhaust overhead tunnels, as well as in the work area, to generate multivariate time-series data. .

[0025] in, For the number of sensors, For time step, The sensor measures dust concentration using laser scattering and wind speed using ultrasonic waves, with a sampling interval of 2 minutes. Data is transmitted via RS485 to a wireless DTU and then uploaded to a cloud database. Based on IoT technology, redundant sensor placement mitigates electromagnetic interference and improves data reliability.

[0026] S102. Preliminary Data Cleaning: Raw Data It contains outliers and duplicate records. The cleaning process is based on statistical anomaly detection: calculating the sequence mean. and standard deviation Remove excess Outliers; use hash functions to detect duplicate records and generate clean data. .in, The mean of the dataset. The standard deviation of the dataset. This is the cleaned data matrix.

[0027] S103. Data Denoising Processing: The data contains high-frequency noise and local fluctuations. A FOX-VMD-IF-DBSCAN composite denoising model is proposed, which combines multi-scale decomposition and anomaly clustering to remove noise and retain fluctuations.

[0028] (1) VMD decomposition: Variational mode decomposition transforms time series Decomposed into One intrinsic mode function and residual : ; in, For the input time series, For the first One modal function, For residuals, Let be the number of modes. The optimization objective is to minimize the modal bandwidth. ; in, For the center frequency, For time derivative, The imaginary unit, Modulation factor. High frequency. Includes noise, low frequency Reflecting trends.

[0029] (2) IF Anomaly Detection: The Isolation Forest IF algorithm is used in high-frequency... Anomalies are detected in the middle, an isolation tree is constructed, and anomaly scores are calculated. ; in, These are abnormal scores. For the sample The length of the isolation path, For the desired path length, This represents the average path length. For the sample size, [This represents the proportion of anomalies.] Output the set of anomalies. .

[0030] (3) DBSCAN secondary discrimination: The DBSCAN algorithm is used to perform secondary discrimination on A_IF, calculate the density of data points, and distinguish them based on density clustering. Isolated noise and clustered fluctuations within the [structure / data]. Density definition: ; in, For data points, radius The neighborhood point set within, For the neighborhood volume, [0.1, 1.0] represents the neighborhood radius. MinPts is the minimum number of neighborhood points threshold for density discrimination; when the number of neighborhood points of data point p... If the value is not less than MinPts, the data point is considered to belong to clustered fluctuations; otherwise, it is considered isolated noise. Clustered fluctuations are retained. Remove isolated noise and reconstruct the denoised sequence: ; in, This is the denoised sequence. For the preserved clustered wave modes.

[0031] (4) FOX optimization: The FOX optimization algorithm is used to determine the number of modes K in variational mode decomposition and the anomaly ratio in the isolated forest algorithm. Neighborhood radius of the DBSCAN algorithm Optimization is performed; the Extreme Fox optimization algorithm is optimized. , , Updated formula: ; in, For the i-th population position, For the target location, and For control parameters, Let be a random vector. The objective is to minimize the permutation entropy. ; Where PE is the permutation entropy. Arrangement mode The probability is calculated and iterated 50 times.

[0032] (5) Missing value handling and normalization: Reversible instance normalization is used to process the denoised sequence. : Standardization: ; Denormalization: ; in, For normalized sequences, For inverse normalized sequences, The mean of the sequence. The standard deviation of the sequence. Learnable affine parameters; generate normalized data .

[0033] S2. Construct the ISageformer prediction model: Basic framework: ISageformer combines GNN and Transformer to process... This captures spatial topology and temporal dependencies. For the number of sensors, For time step, This refers to the number of channels. It includes: (201) Sequence-aware global token generation module; add global Embedded: ; in, For initial embedding, To enable the learning of global tokens, For normalized input.

[0034] (202) Graph structure learning module: Using each sensor as a node, the module learns the dependencies between nodes through a multilayer perceptron (MLP) to generate an adjacency matrix A. ; in, The adjacency matrix is ​​used, and MLP stands for Multilayer Perceptron. This is the initial embedding.

[0035] (203) Graph aggregation module: using normalized graph Laplace: ; in, To normalize the Laplace matrix, For degree matrix, This is an adjacency matrix. Multi-level aggregation: ; in, For the first Layer embedding, It is the ReLU activation function. As the weight matrix, generate graph-enhanced embeddings. .

[0036] (204) Timing Encoding Module: Transformer Encoder Processing ,generate: ; in, , , These represent the query, key, and value matrices, respectively, with Conv1D representing a one-dimensional convolution. Multi-head attention: ; Output: ; in, For the first Pay attention to the head. For single-head dimension, For the number of heads, To output weights, For attention output.

[0037] (205) Improved trend-aware attention module: RevIN Integration: In Initial Embedding and the output of the time encoding module RevIN is applied to alleviate distribution shift. Among other things, For initial embedding, For attention output, Or These are the mean, standard deviation, and affine parameters.

[0038] Multi-head attention mechanism: Causal convolution extracts local trends. ; in, This represents the graph-enhanced embedding features obtained after processing by the graph structure learning module and the graph aggregation module; This represents a one-dimensional causal convolution operation; k represents the kernel size of the causal convolution. This represents the local trend features extracted by causal convolution.

[0039] attention: ; in, This is a local trend. , , For trend-based queries, key, value, For the embedded dimension.

[0040] 3. Model Training and Optimization: Loss functions: using mean squared error (MSE) and mean absolute error (MAE): ; Constructing joint loss: ; in, For predicted values, For the true value, For the number of sensors, For time step, These are the weighting coefficients.

[0041] Using the Adam optimizer: ; ; in, First-order momentum, It is a second-order momentum. For gradient, , The attenuation rate, Let be the learning rate, and be a small constant. , For model parameters, This is the first-order momentum estimate from the previous iteration. It is a numerically stable term.

[0042] Early stopping and overfitting prevention: Stop if the validation set loss does not improve for 5 consecutive epochs.

[0043] Hyperparameter tuning principles and implementation: Hyperparameters include encoder layer number, attention head number, hidden dimension, global token number, node embedding dimension, nearest neighbor number, and graph aggregation depth.

[0044] Using FOX optimization to minimize and verify the MAE dataset partitioning principle: It is divided into training set, validation set, and test set. This is normalized data.

[0045] S4. Ventilation Environment Prediction and Application: First, a ventilation environment prediction is performed, by inputting normalized data. Output: ; in, For predicted values, For normalized input, To predict the time step.

[0046] Then, the application is based on the predicted values. Used for threshold alarms and ventilation optimization: ; in, For the frequency of the fan, To predict dust concentration, For the target concentration, This is a scaling factor. It integrates with a cloud system to achieve automated control.

[0047] Example 2 This embodiment provides supplementary explanations to the above-mentioned spatiotemporal prediction method for complex construction ventilation environments in underground cavern groups based on deep learning, using specific applications and data, as follows: In a large-scale underground cavern project at a hydropower station in Southwest China, the project involves a complex network of underground caverns, including the main powerhouse, ventilation and safety caverns, main transformer exhaust caverns, and exhaust surface caverns. The project is massive in scale, generating a large amount of dust during excavation. Affected by blasting impacts, mechanical vibrations, and construction activities, the monitoring data exhibits high-frequency noise and strong fluctuations. To achieve robust spatiotemporal prediction of dust concentration and wind speed, the following steps are employed to implement this method: S1. Data Acquisition and Preprocessing: S101. Data Acquisition: Deploy 10-15 sensors at key nodes along the ventilation path (such as cavern entrances, blasting areas, and near fans), covering approximately 500m of the ventilation duct. Use dust sensors based on the laser scattering principle to measure PM concentration. Ultrasonic anemometers measure wind speed The sampling interval was 2 minutes, and the collection period was 7 consecutive days, generating a data matrix of approximately 5000 time steps. The data is transmitted to the wireless DTU via an RS485 wired connection and then uploaded to the cloud database.

[0048] S102. Preliminary cleaning: Cleaning the original... Calculate the mean and standard deviation Remove approximately 5% of outliers; use a hash function to remove duplicate records, generating clean data. The amount of data was reduced to about 95%, and the data quality was improved.

[0049] S103. Noise Reduction Processing: VMD decomposition: For each sequence Decomposed into Modality and residual High-frequency modes capture noise, while low-frequency modes retain trends.

[0050] IF anomaly detection: High frequency detection Generate an anomaly set from the anomalies in the data. Identify The data contains potential noise or fluctuations.

[0051] DBSCAN discrimination: based on Clustering with MinPts=3 Preserving clustered fluctuations Remove isolated noise and generate a denoised sequence. .

[0052] FOX optimization: Initialize population size to 20, optimize after 50 iterations. , , The objective is to minimize the permutation entropy PE, ensuring improved noise reduction performance under Gaussian noise. Lévy noise enhancement .

[0053] Permutation entropy was chosen as the optimization objective because it can quantitatively characterize the randomness and complexity of a time series, is extremely sensitive to high-frequency random noise, and exhibits strong robustness to structural fluctuations caused by construction activities, making it suitable as an evaluation metric for noise reduction effectiveness. S104. Missing Value Handling and Normalization: right Apply RevIN to process approximately Missing values, normalized generation Inverse normalization restores the original distribution.

[0054] S2. Model Building: ISageformer Construction: Embedding generate (Global Token count: 1), graph structure learning generation softmax(MLP The graph is finally generated after a 3-layer aggregation depth. Generated with 2 Transformer layers and 8 attention heads. .

[0055] Improved trend-aware attention module: First, graph-enhanced embeddings are applied. One-dimensional causal convolution is used to extract local trend features, resulting in a local trend representation. The convolution kernel size is 3; then, based on the local trend features... Query vectors, key vectors, and value vectors are constructed, and time-series features are weighted and aggregated through an attention mechanism to form a trend-aware attention output; Trend-aware attention incorporates local trend information, enabling attention weights to adaptively reflect short-term change patterns in ventilation environmental parameters.

[0056] S3. Model Training: The dataset is divided into a 70% training set, a 15% validation set, and a 15% test set; the Adam optimizer is used to initialize the learning rate. Training for 20 epochs, joint loss MSE + 0.3MAE, early stopping to prevent overfitting, hidden dimension 512, monitoring the mean absolute error (MAE) of the validation set.

[0057] S4. Prediction and Application: Input test set The ISageformer prediction model outputs a 30-minute forecast. MAE decreased by 38%. (Forecast) For threshold alarm (over 50mg / (Early warning), adjust fan frequency With k=1.2, the fan can be operated by frequency conversion, reducing energy consumption by 15%, optimizing the ventilation scheme, and ensuring that the risk of dust exceeding the standard is reduced.

[0058] This embodiment verifies the effectiveness of the method under complex operating conditions, supporting engineering safety and schedule.

[0059] Example 3 Based on the same inventive concept, this application also provides a spatiotemporal prediction device for complex construction ventilation environment of underground cavern groups based on deep learning, which can be used to implement the method described in the above embodiments, and specifically includes the following: The data acquisition and preprocessing module is used to deploy several sensors along the ventilation path and work area of ​​the underground cavern complex to collect time-series data on wind speed and dust concentration. After cleaning the time-series data, a composite noise reduction model is used for processing. The composite noise reduction model processing includes: decomposing the time series into modal functions and residuals; performing anomaly detection on the high-frequency modal functions to obtain anomaly point sets; and distinguishing between isolated noise and clustered fluctuations in the anomaly point sets based on density clustering; retaining clustered fluctuations and removing isolated noise to obtain the denoised sequence; finally, performing Reversible Instance Normalization (RevIN) processing on the denoised sequence to generate normalized data. The ISageformer prediction model building module is used to construct a spatiotemporal prediction model that includes a sequence-aware global token generation module, a graph structure learning module, a graph aggregation module, a time encoding module, and a trend-aware attention module. The spatiotemporal prediction model integrates a reversible instance normalization RevIN mechanism at both the input and output ends to handle data distribution shifts. The model training and optimization module is used to divide the normalized data into training set, validation set and test set, construct joint loss function, train the spatiotemporal prediction model using optimization algorithm, and automatically optimize the hyperparameters of the composite noise reduction model in step S1 using swarm intelligence optimization algorithm. The prediction and application module is used to input the real-time monitoring data into the trained spatiotemporal prediction model after preprocessing in step S1, output the predicted values ​​of wind speed and dust concentration for future time steps, and dynamically adjust the operating frequency of the ventilation system's fans based on the predicted values.

[0060] Preferably, embodiments of this application also provide a specific implementation of an electronic device capable of implementing all steps in the deep learning-based spatiotemporal prediction method for complex construction ventilation environments in underground cavern groups described in the above embodiments. The electronic device specifically includes the following: Processor, memory, communications interface, and bus; The processor, memory, and communication interface communicate with each other via a bus; the communication interface is used to realize information transmission between server-side devices, metering devices, and user-side devices.

[0061] The processor is used to call the computer program in the memory. When the processor executes the computer program, it implements all the steps in the spatiotemporal prediction method for complex construction ventilation environment of underground cavern groups based on deep learning in the above embodiments.

[0062] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the deep learning-based spatiotemporal prediction method for complex construction ventilation environment of underground cavern groups in the above embodiments. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements all steps of the deep learning-based spatiotemporal prediction method for complex construction ventilation environment of underground cavern groups in the above embodiments.

[0063] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, hardware + program embodiments are relatively simple in description because they are fundamentally similar to method embodiments; relevant parts can be referred to the descriptions in the method embodiments.

[0064] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0065] While this application provides method operation steps as shown in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-inventive labor. The order of steps listed in the embodiments is merely one possible execution order among many and does not represent the only execution order. In actual device or client product execution, the method can be executed in the order shown in the embodiments or drawings or in parallel (e.g., in a parallel processor or multi-threaded processing environment).

[0066] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0067] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0068] This invention is not limited to the embodiments described above. The above description of specific embodiments is intended to illustrate and explain the technical solutions of this invention. The specific embodiments described above are merely illustrative and not restrictive. Without departing from the spirit and scope of the claims, those skilled in the art can make many specific modifications based on the teachings of this invention, and these modifications all fall within the scope of protection of this invention.

Claims

1. A spatiotemporal prediction method for complex construction ventilation environment of underground cavern groups based on deep learning, characterized in that, Includes the following steps: S1. Data Acquisition and Preprocessing: Several sensors were deployed in the ventilation paths and work areas of the underground cavern complex to collect time-series data on wind speed and dust concentration. After cleaning the time-series data, a composite noise reduction model is used for processing. The composite denoising model processing includes: decomposing the time series into mode functions and residuals, performing anomaly detection on the high-frequency mode functions to obtain anomaly point sets, and distinguishing isolated noise and clustered fluctuations in the anomaly point sets based on density clustering, retaining clustered fluctuations and removing isolated noise to obtain the denoised sequence; finally, performing reversible instance normalization RevIN processing on the denoised sequence to generate normalized data. S2. Construct the ISageformer prediction model: A spatiotemporal prediction model is constructed, comprising a sequence-aware global token generation module, a graph structure learning module, a graph aggregation module, a time encoding module, and a trend-aware attention module; wherein, the spatiotemporal prediction model integrates a reversible instance normalization RevIN mechanism at both the input and output ends to handle data distribution shifts; S3. Model Training and Optimization: Normalized data is divided into training set, validation set and test set, a joint loss function is constructed, the spatiotemporal prediction model is trained using optimization algorithm, and the hyperparameters of the composite noise reduction model in step S1 are automatically optimized using swarm intelligence optimization algorithm. S4. Prediction and Application: After preprocessing the real-time monitoring data in step S1, it is input into the trained spatiotemporal prediction model, which outputs the predicted values ​​of wind speed and dust concentration for future time steps, and dynamically adjusts the operating frequency of the ventilation system's fans based on the predicted values.

2. The spatiotemporal prediction method for complex construction ventilation environment of underground cavern groups based on deep learning according to claim 1, characterized in that, In step S1, the FOX-VMD-IF-DBSCAN model is used as the composite noise reduction model; specifically, it includes: Variational Mode Decomposition (VMD) is used to analyze the cleaned time series. Decomposed into K intrinsic mode functions and residual ; The Isolation Forest IF algorithm is used to detect outliers in high-frequency mode functions and generate an outlier set. ; Using the DBSCAN algorithm A secondary discrimination is performed, calculating the density of data points. If the density is greater than a set threshold, it is determined to be a clustered fluctuation function. Those that are retained will be retained; otherwise, they will be considered isolated noise and removed. Reconstructing the denoised sequence: ; This is the denoised sequence. This is the preserved clustered wave function.

3. The spatiotemporal prediction method for complex construction ventilation environment of underground cavern groups based on deep learning according to claim 2, characterized in that, The FOX optimization algorithm was used to determine the number of modes K in variational mode decomposition and the outlier ratio in the Isolation Forest algorithm. Neighborhood radius of the DBSCAN algorithm The optimization is performed; the objective function of the optimization is to minimize the permutation entropy PE of the denoised sequence; the position update formula of the FOX algorithm is: ;in, Where T is the population location and T is the target location. and Here, r is the control parameter, and r is a random vector.

4. The spatiotemporal prediction method for complex construction ventilation environment of underground cavern groups based on deep learning according to claim 1, characterized in that, In step S2, the sequence-aware global token generation module normalizes the input. Add a learnable global token G, and generate the initial embedding using the following formula. , ; The graph structure learning module generates an adjacency matrix A by learning the dependencies between nodes using a multilayer perceptron (MLP). .

5. The spatiotemporal prediction method for complex construction ventilation environment of underground cavern groups based on deep learning according to claim 1, characterized in that, In step S2, the processing procedure of the trend-aware attention module is as follows: Enhanced embedding by extracting graphs using causal convolution Local trends : ； in, This represents the graph-enhanced embedding features obtained after processing by the graph structure learning module and the graph aggregation module; This represents a one-dimensional causal convolution operation; k represents the kernel size of the causal convolution. This represents the local trend features extracted by causal convolution; Calculate the query matrix based on local trends. Key matrix Sum matrix And calculate the attention score: ； Where d is the embedding dimension.

6. The spatiotemporal prediction method for complex construction ventilation environment of underground cavern groups based on deep learning according to claim 1, characterized in that, In step S1, the reversible instance normalization RevIN includes normalization and denormalization; Standardization process: ； Denormalization process: ； in, This is the denoised sequence. For normalized sequences, For inverse normalized sequences, The mean of the sequence. The standard deviation of the sequence. Learnable affine parameters; generate normalized data .

7. The spatiotemporal prediction method for complex construction ventilation environment of underground cavern groups based on deep learning according to claim 1, characterized in that, In step S3, the joint loss function L is composed of a weighted average of the mean squared error (MSE) and the mean absolute error (MAE): , These are the weighting coefficients; In step S4, the specific steps for dynamically adjusting the operating parameters of the ventilation system based on the predicted values ​​are as follows: Based on dust concentration prediction values and target concentration Calculate the fan operating frequency f: ; Where k is the scaling factor; when the dust concentration is predicted... An alarm is triggered when the threshold is exceeded.

8. A spatiotemporal prediction device for complex construction ventilation environment of underground cavern groups based on deep learning, characterized in that, include: The data acquisition and preprocessing module is used to deploy several sensors in the ventilation path and work area of ​​the underground cavern complex to collect time-series data on wind speed and dust concentration. After cleaning the time-series data, a composite noise reduction model is used for processing. The composite denoising model processing includes: decomposing the time series into mode functions and residuals, performing anomaly detection on the high-frequency mode functions to obtain anomaly point sets, and distinguishing isolated noise and clustered fluctuations in the anomaly point sets based on density clustering, retaining clustered fluctuations and removing isolated noise to obtain the denoised sequence; finally, performing reversible instance normalization RevIN processing on the denoised sequence to generate normalized data. The ISageformer prediction model building module is used to construct a spatiotemporal prediction model that includes a sequence-aware global token generation module, a graph structure learning module, a graph aggregation module, a time encoding module, and a trend-aware attention module. The spatiotemporal prediction model integrates a reversible instance normalization RevIN mechanism at both the input and output ends to handle data distribution shifts. The model training and optimization module is used to divide the normalized data into training set, validation set and test set, construct joint loss function, train the spatiotemporal prediction model using optimization algorithm, and automatically optimize the hyperparameters of the composite noise reduction model in step S1 using swarm intelligence optimization algorithm. The prediction and application module is used to input the real-time monitoring data into the trained spatiotemporal prediction model after preprocessing in step S1, output the predicted values ​​of wind speed and dust concentration for future time steps, and dynamically adjust the operating frequency of the ventilation system's fans based on the predicted values.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the spatiotemporal prediction method for complex construction ventilation environment of underground cavern groups based on deep learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the spatiotemporal prediction method for complex construction ventilation environment of underground cavern groups based on deep learning as described in any one of claims 1 to 7.

Images