Underground structure energy consumption management method and system based on energy recovery

Abstract

The invention relates to the technical field of energy consumption management, in particular to an underground structure energy consumption management method and system based on energy recovery. According to the technical scheme, the method comprises the steps of deploying a multi-source sensor network in a key area of an underground structure, performing spatial-temporal feature fusion processing on sensor data, constructing a hybrid prediction model, designing an optimization controller based on a reinforcement learning algorithm, and adjusting operation parameters of an energy recovery system in a graded manner according to real-time load prediction deviation. Operation parameters of the energy recovery system can be adjusted in a grading mode according to real-time load deviation, and dynamic power distribution is achieved; all layers of the management system work cooperatively, an FPGA module preprocesses data, equipment digital twinborn simulation analysis is performed, and a fault diagnosis module early warns equipment faults. On the whole, underground environment information can be accurately grasped, energy recovery efficiency is improved, energy consumption is reduced, equipment loss is reduced, stable and efficient operation of the system is guaranteed, and the energy utilization level of an underground structure is improved.

Description

Technical Field

[0001] This invention relates to the field of energy management technology, and in particular to a method and system for energy management of underground structures based on energy recovery. Background Technology

[0002] With the rapid development of urban rail transit, subways, as an efficient and convenient mode of transportation, are experiencing a continuous increase in passenger volume. Underground structures, such as subway stations, have a very high proportion of HVAC (heating, ventilation, and air conditioning) energy consumption in their total energy consumption. Statistics show that in some large subway stations, HVAC system energy consumption can even reach over 60% of the total energy consumption. Because subway operations are highly time-sensitive, peak hours see large passenger flows, and human heat dissipation becomes a significant heat source, leading to a sharp increase in the load on the ventilation system; while during off-peak hours, passenger flow decreases sharply, and the load drops significantly. This dramatic fluctuation in load presents a significant challenge to the energy consumption management of subway stations.

[0003] Traditional energy management methods for subway stations have many drawbacks. Firstly, in the data acquisition phase, the number and type of sensors used are limited, failing to comprehensively capture complex and ever-changing environmental information. For example, relying on only a few temperature and humidity sensors makes it difficult to accurately obtain temperature distribution in different areas of the platform, let alone effectively monitor the distribution of passenger flow heat sources. Secondly, in terms of energy recovery and equipment control, there is a lack of dynamic adjustment mechanisms. Waste heat recovery devices typically operate according to fixed patterns, unable to adjust recovery power in a timely manner based on real-time passenger flow and load changes. This results in a situation where, during peak hours, a large amount of waste heat cannot be recovered in time, requiring refrigeration units to consume additional electricity for cooling; while during off-peak hours, waste heat recovery devices may over-operate, leading to energy waste.

[0004] With the continuous development of emerging technologies such as the Internet of Things (IoT) and artificial intelligence (AI), new ideas and methods have been provided for solving the energy consumption management problem in subway stations. IoT technology enables the interconnection of various sensors, collecting massive amounts of environmental parameters, passenger flow and heat source distribution data, and equipment operating status data in real time. By constructing an IoT-based dynamic monitoring network, comprehensive and real-time perception of the complex environment within subway stations can be achieved. Meanwhile, AI algorithms, such as reinforcement learning, have demonstrated powerful advantages in handling complex decision-making problems. Using reinforcement learning algorithms, the control strategy of energy recovery devices can be dynamically optimized based on real-time collected data, achieving precise matching between the energy recovery system and the building load.

[0005] However, applying IoT and AI technologies to subway station energy management still faces numerous challenges. In terms of data processing, the collected multi-source heterogeneous data suffers from noise interference and data gaps. Efficiently cleaning, fusing, and extracting features from this data is crucial for achieving accurate prediction and optimized control. Regarding model building and training, the complex environment of subway stations and the numerous factors influencing load changes necessitate in-depth research into establishing accurate predictive models and ensuring rapid convergence to the optimal solution in actual operation. Furthermore, in terms of system integration, the organic integration of multiple subsystems such as energy recovery systems, ventilation systems, and refrigeration systems to achieve coordinated operation is also an urgent problem to be solved.

[0006] Therefore, this application proposes a method and system for energy consumption management of underground structures based on energy recovery. Summary of the Invention

[0007] The purpose of this invention is to address the problem in the background art that the ventilation system load of subway stations fluctuates greatly due to the peak and valley changes in passenger flow during operation, and that existing waste heat recovery devices cannot dynamically adjust the recovery power according to real-time passenger flow, resulting in energy waste or additional energy consumption of the refrigeration unit. This invention proposes an energy consumption management method and system for underground structures based on energy recovery.

[0008] In a first aspect, this application provides a method for energy consumption management of underground structures based on energy recovery, comprising the following steps:

[0009] S1. Deploy a multi-source sensor network in key areas of the underground structure to collect environmental parameters, passenger flow heat source distribution and equipment operation status data in real time;

[0010] S2. Perform spatiotemporal feature fusion processing on sensor data to extract a multidimensional vector containing time period features, spatial correlation features, and device coupling features;

[0011] S3. Construct a hybrid forecasting model, predict the basic load through an LSTM neural network, and combine it with the XGBoost algorithm to correct the heat dissipation associated with passenger flow, and dynamically weight the output of the comprehensive load forecasting result;

[0012] S4. Design and optimize the controller based on reinforcement learning algorithm, with the optimization objectives of recovery efficiency, energy consumption reduction rate and equipment operation frequency, and generate control strategy for heat recovery device;

[0013] S5. Adjust the operating parameters of the energy recovery system according to the real-time load forecast deviation and realize dynamic power distribution through industrial communication protocol.

[0014] Optionally, S1 specifically includes: constructing a multi-source heterogeneous sensor monitoring network, deploying temperature and humidity sensors, CO2 concentration sensors, infrared thermal imagers, and equipment current transformers in the target area, and synchronously collecting environmental parameters, passenger flow heat source distribution, and equipment operating status data, wherein the target area includes at least the subway platform level, transfer hall, and equipment room.

[0015] The deployment method of the infrared thermal imager includes: arranging 3 thermal imagers at the vertices of an equilateral triangle along the main passageway on the ceiling of the station hall, with the side length of the triangle being 12±0.5 meters, the elevation angle of each detection point being adjusted to 30°±2°, and the horizontal scanning angle being set to 120°.

[0016] The dynamic heat source is extracted using an adaptive background difference algorithm, and the calculation formula is as follows:

[0017]

[0018] Where I(x,y,t) is the radiant intensity value of the (x,y) coordinates at time t, N=30 is the number of background modeling frames, and G σ (x,y) is the Gaussian filter kernel with a standard deviation σ = 1.5 pixels, ΔT is the temperature difference score, and k is the historical frame index;

[0019] When ΔT>5℃, it is determined to be an effective human heat source, and the spatial positioning error is less than 0.3 meters.

[0020] Optionally, in step S2, the spatiotemporal feature fusion processing of the sensor data specifically includes the following steps:

[0021] S2a. Perform sliding window standardization on the raw data, with the window length set to 15 minutes, and use the Z-score normalization method to eliminate dimensional differences;

[0022] S2b. Extract multi-dimensional feature vectors, including time dimension features, spatial topology features, and device association features. The time dimension features include historical 24-hour load data of the same period, and the spatial topology features are established by using the Delaunay triangulation algorithm to build a sensor spatial association matrix.

[0023] Optionally, in step S3, constructing the hybrid prediction model specifically includes the following steps:

[0024] S3a. An LSTM neural network is used to establish a basic load prediction module. The network structure contains 3 hidden layers, each with 128 neurons. The input layer receives the feature vector sequence of the previous 2 hours.

[0025] The LSTM neural network employs an improved structure with an attention mechanism, specifically including:

[0026] An attention layer is added after the output of the LSTM hidden layer to calculate the context vector c at the i-th time step. i :

[0027]

[0028] Where T is the total number of time steps, and the attention weight α is... ij Calculated in the following way:

[0029]

[0030] Among them, e ij To score attention, This represents the hidden state vector at the current moment. This represents the implicit state vector at a historical moment. For trainable weight matrix, For paranoia, The attention score vector is fed into a fully connected layer, and the activation function is LeakyReLU (α = 0.01).

[0031] S3b. An XGBoost algorithm is used to establish a passenger flow correlation correction module. The CO2 concentration change rate is used as an indirect passenger flow indicator. A gradient boosting decision tree model is constructed to predict the human body heat dissipation in the next 30 minutes.

[0032] S3c. The outputs of the two modules are fused using an adaptive weighting algorithm, and the weighting coefficients are dynamically adjusted according to the prediction error.

[0033] Optionally, the S4 step of designing and optimizing the controller based on a reinforcement learning algorithm specifically includes the following steps:

[0034] S4a. Define the state space as [current recovered power, predicted load deviation, equipment efficiency coefficient], and the action space as [heat exchanger power adjustment amount, water pump speed adjustment level, bypass valve opening percentage];

[0035] S4b. Design the reward function R = α × recovery efficiency + β × energy consumption reduction rate - γ × equipment operation frequency, where α = 0.6, β = 0.3, γ = 0.1; the recovery efficiency η is calculated as follows:

[0036]

[0037] in, To actually recover heat, Q max Q represents the theoretical maximum recovery capacity. baseline Based on the recovered heat, c p T is the specific heat capacity of air at constant pressure, ρ is the density of air, and T is the specific heat capacity of air at constant in T outThese are the inlet and outlet temperatures of the heat exchanger, The real-time air volume is calculated using the following formula:

[0038]

[0039] Where t is the current time, k is an integer loop variable, and 168 represents the upper limit of the loop in the summation operation, which is the moving average of the actual values ​​at the same time in the past 7 days.

[0040] Q max =P rated ·Δt·COP represents the theoretical maximum recovery capacity, P rated =50kW is the rated power of the heat recovery device, and COP=3.2 is the equipment performance coefficient.

[0041] S4c. The agent is trained using the Deep Deterministic Policy Gradient (DDPG) algorithm, with an experience replay pool capacity of 10,000 entries and a critic network learning rate of 0.001.

[0042] Optionally, S5 includes implementing a dynamic control strategy, specifically including:

[0043] S5a. Execute a control decision every 5 minutes to generate a power allocation scheme for the heat recovery device;

[0044] S5b. When the predicted load deviation exceeds 15%, the emergency adjustment mode is triggered, and the operating status of the plate heat exchanger is adjusted first.

[0045] Emergency response mode includes a tiered response strategy:

[0046] Level 1 response (10% < deviation ≤ 15%):

[0047] Adjust the plate heat exchanger fan speed to 75% of the rated speed.

[0048] Cooling water flow rate adjustment formula:

[0049]

[0050] in, Q represents the adjusted flow rate. pred To predict the load, ΔQ = Q actual -Q pred For real-time load deviation, The reference water flow rate is ΔQ, and the load deviation is ΔQ.

[0051] Secondary response (deviation > 15%):

[0052] Activate the emergency cold storage tank to release the cooling capacity Q. storage =min(ΔQ·Δt,Q) max), Δt is the time interval, Q max This represents the theoretical maximum heat recovery.

[0053] Trigger the frequency increase command for the ventilation system; frequency adjustment amount:

[0054] f new =f current +K p ·ΔQ K p =0.05Hz / kW

[0055] Among them, f new To set the new frequency value, f current K represents the current frequency. p The proportionality constant is given, with the constraint: 20℃ ≤ T. supply ≤25℃, T supply This refers to the supply air temperature.

[0056] S5c sends PWM control signals to the frequency converter driver via the Modbus-TCP protocol to achieve precise adjustment of reclaimed power.

[0057] S5d. Execute the exception handling mechanism. When abnormal sensor data or equipment failure is detected, the data repair and security control mode is activated. The exception handling mechanism includes:

[0058] Data reliability assessment uses Mahalanobis distance detection:

[0059]

[0060] Where x is the observation vector, μ is the homogeneous vector, and ∑ is the covariance matrix, when When an anomaly is detected, n is the feature dimension. Let T be the chi-square critical value, and T be the matrix transpose operation;

[0061] The state equation for Kalman filter repair:

[0062] x k =Ax k-1 +Bu k +w k

[0063] z k =Hx k +v k

[0064] Where, x k Let A be the state variables, including temperature, humidity, and CO2 concentration, and B be the input matrix. k Let H represent the control input vector of the system at time k, and let w be the observation matrix. k For process noise, vk To observe noise.

[0065] Secondly, this application provides an energy management system for underground structures based on energy recovery, comprising:

[0066] Sensor network layer: includes temperature and humidity detection terminals, multispectral passenger flow counters, and equipment status monitoring units;

[0067] Edge computing layer: Equipped with an FPGA data preprocessing module and a timing database, wherein the FPGA module has a built-in sliding window differential coding circuit;

[0068] Cloud analytics layer: Deploy load forecasting engines, reinforcement learning decision models, and device digital twins;

[0069] The execution control layer includes intelligent frequency converter cabinets, electric regulating valve groups, and fault diagnosis modules;

[0070] The edge computing layer is connected to the sensor network layer via the OPC UA protocol, and the cloud analytics layer interacts with the edge computing layer via the MQTT protocol.

[0071] Optionally, the FPGA data preprocessing module includes:

[0072] The hardware logic implementation of the parallelized sliding window differential encoder is as follows:

[0073]

[0074] The dynamic range adjustment circuit has the following adjustment rules:

[0075]

[0076] Among them, Gain current Gain is the current gain value. new The adjusted gain value, σ 2 The variance of the signal, 2V 2 For the set variance threshold, the temperature drift compensation module uses the following compensation formula:

[0077] V corrected =V raw ×(1+αΔT+βΔT 2 )

[0078] Among them, V corrected V is the voltage value after temperature drift compensation. raw The original voltage signal is given, ΔT is the temperature change, and α and β are compensation coefficients.

[0079] α = 2.1 × 10 -4 / ℃, β=8.7×10 -7 / ℃2.

[0080] Optionally, the three-dimensional thermodynamic simulation model of the device's digital twin adopts an unsteady heat transfer equation:

[0081]

[0082] Where ρ is the material density, c p q is the specific heat capacity, k is the thermal conductivity, and q is the thermal conductivity. gen For the heat source generation term, q loss This is the heat loss term, where T is temperature and t is time. For Hamiltonian operators, Let T be the partial derivative of temperature T with respect to time t;

[0083] Boundary conditions:

[0084] Platform wall: Type III boundary conditions

[0085] Where k is the thermal conductivity. Let T represent the directional derivative of temperature T along the wall normal n, and h be the convective heat transfer coefficient. air For air temperature, T wall The temperature of the platform wall;

[0086] Equipment surface: Second type boundary condition q″=520W / m 2 q″ is the heat flux density;

[0087] Mesh generation: unstructured mesh, minimum size 0.1m, time step Δt = 10s.

[0088] Optionally, the three-level early warning mechanism of the fault diagnosis module includes:

[0089] Level 1 Warning: Generate equipment efficiency trend chart:

[0090]

[0091] Where, η trend (t) represents the trend value of equipment efficiency at time t. Let η be a coefficient, k be an integer variable, Δt be the time interval, and η(t-kΔt) be the equipment efficiency at time t-kΔt. For equipment efficiency trend η trend The reciprocal of (t), when the slope Maintenance reminders will be issued in a timely manner;

[0092] Level 2 warning: Perform wavelet packet decomposition on the vibration signal and extract the energy ratio in the 3-5kHz frequency band.

[0093]

[0094] Where f is the frequency, PSD(f) is the power spectral density function, and 3k, 5k, and 10k are 3000Hz, 5000Hz, and 10000Hz, respectively.

[0095] When E ratio Bearing wear is determined when the percentage is >15%;

[0096] Level 3 warning: Insulation resistance is measured using the DC superposition method.

[0097]

[0098] Among them, V DC For DC superposition voltage, I leakage For leakage current, R shunt As the shunt resistor, when R iso The circuit is triggered to stop when the resistance is less than 50MΩ.

[0099] Compared with the prior art, this application includes at least one of the following beneficial technical effects:

[0100] It provides accurate data collection on environmental parameters, passenger flow and heat source distribution, and equipment operating status. By performing spatiotemporal feature fusion processing on this data, it can extract multi-dimensional vectors of time period features, spatial correlation features, and equipment coupling features, providing accurate and comprehensive data support for subsequent prediction and control, and improving the ability to grasp information about complex underground environments.

[0101] The constructed hybrid prediction model combines LSTM neural networks and XGBoost algorithms. The LSTM neural network uses an improved structure with an attention mechanism to predict the base load, which can more accurately predict the comprehensive load of underground structures, making energy supply and equipment regulation more aligned with actual needs and avoiding energy waste or insufficient supply.

[0102] By rationally defining the state space and action space, designing a scientific reward function, and training the agent using a deep deterministic policy gradient algorithm, the operating parameters of the heat recovery device can be dynamically adjusted according to real-time conditions, thereby improving recovery efficiency, reducing system energy consumption, reducing unnecessary equipment actions, and extending equipment lifespan.

[0103] Precise power regulation is achieved through the Modbus-TCP protocol, and an anomaly handling mechanism is implemented. Maslow's distance detection and Kalman filtering are used to repair abnormal data, ensuring stable and efficient system operation.

[0104] Data is collected through a sensor network layer, preprocessed and stored by an edge computing layer, analyzed and made by a cloud analytics layer, and controlled by an execution control layer. These layers interact via OPC UA and MQTT protocols, enabling efficient data transmission and overall system optimization, thus improving system efficiency and reliability.

[0105] The system's fault diagnosis module employs a three-level early warning mechanism, which helps to promptly identify potential equipment problems, schedule maintenance in advance, reduce equipment downtime, ensure the continuous and stable operation of the system, and lower maintenance costs.

[0106] This invention collects data by deploying a multi-source sensor network in key areas of underground structures. After spatiotemporal feature fusion, a hybrid prediction model is constructed to accurately predict load. Based on reinforcement learning, an optimized controller is designed to generate control strategies for the heat recovery device. The system can adjust the operating parameters of the energy recovery system in stages according to real-time load deviations, achieving dynamic power allocation. Its management system enables collaborative work across all layers, with an FPGA module preprocessing data, digital twin simulation analysis of equipment, and a fault diagnosis module providing early warnings of equipment failures. Overall, it can accurately grasp underground environmental information, improve energy recovery efficiency, reduce energy consumption, minimize equipment wear, ensure stable and efficient system operation, and enhance the energy utilization level of underground structures. Attached Figure Description

[0107] Figure 1 A flowchart illustrating an energy consumption management method for underground structures based on energy recovery;

[0108] Figure 2 This is a schematic diagram of a principle-based energy consumption management system for underground structures based on energy recovery.

[0109] Figure 3 This is a schematic diagram of an energy consumption management method for underground structures based on energy recovery. Detailed Implementation

[0110] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0111] Example 1

[0112] like Figure 1 As shown, this invention proposes an energy consumption management method for underground structures based on energy recovery. The steps are described in detail below.

[0113] S1. Deploy a multi-source sensor network in key areas of the underground structure to collect environmental parameters, passenger flow heat source distribution, and equipment operation status data in real time; specifically, this includes: constructing a multi-source heterogeneous sensor monitoring network, deploying temperature and humidity sensors, CO2 concentration sensors, infrared thermal imagers, and equipment current transformers in the target area, and synchronously collecting environmental parameters, passenger flow heat source distribution, and equipment operation status data, wherein the target area includes at least the subway platform level, transfer hall, and equipment room;

[0114] The deployment method of the infrared thermal imager includes: arranging 3 thermal imagers at the vertices of an equilateral triangle along the main passageway on the ceiling of the station hall, with the side length of the triangle being 12±0.5 meters, the elevation angle of each detection point being adjusted to 30°±2°, and the horizontal scanning angle being set to 120°.

[0115] The dynamic heat source is extracted using an adaptive background difference algorithm, and the calculation formula is as follows:

[0116]

[0117] Where I(x,y,t) is the radiant intensity value of the (x,y) coordinates at time t, N=30 is the number of background modeling frames, and G σ (x,y) is the Gaussian filter kernel with a standard deviation σ = 1.5 pixels, ΔT is the temperature difference score, and k is the historical frame index;

[0118] A temperature ΔT > 5℃ is considered a valid human heat source, with a spatial positioning error of less than 0.3 meters. This embodiment achieves real-time and accurate monitoring of environmental parameters, passenger flow heat source distribution, and equipment status. The equilateral triangle layout of the infrared thermal imager, combined with an adaptive background difference algorithm, enables millimeter-level positioning of human heat sources (error < 0.3 meters). A multi-dimensional data acquisition system is established to provide basic data support for subsequent analysis.

[0119] S2. Perform spatiotemporal feature fusion processing on the sensor data to extract a multidimensional vector containing time period features, spatial correlation features, and device coupling features; the spatiotemporal feature fusion processing of the sensor data specifically includes the following steps:

[0120] S2a. Perform sliding window standardization on the raw data, with the window length set to 15 minutes, and use the Z-score normalization method to eliminate dimensional differences;

[0121] S2b. Extract multi-dimensional feature vectors, including time-dimensional features, spatial topology features, and equipment correlation features. The time-dimensional features include historical 24-hour load data within the same period. The spatial topology features are established using the Delaunay triangulation algorithm to create a sensor spatial correlation matrix. In this embodiment, sliding window standardization eliminates differences in data dimensions, improving data comparability. By fusing time-period features (24-hour load patterns), spatial correlation features (Delaunay triangulation modeling), and equipment coupling features, the data representation capability is enhanced, providing the load forecasting model with high-dimensional feature vectors containing spatiotemporal coupling relationships.

[0122] S3. Construct a hybrid forecasting model, predicting the base load using an LSTM neural network, and combining it with the XGBoost algorithm to correct passenger flow-related heat dissipation, dynamically weighting the output of the comprehensive load forecast result; the construction of the hybrid forecasting model specifically includes the following steps:

[0123] S3a. An LSTM neural network is used to establish a basic load prediction module. The network structure contains 3 hidden layers, each with 128 neurons. The input layer receives the feature vector sequence of the previous 2 hours.

[0124] The LSTM neural network employs an improved structure with an attention mechanism, specifically including:

[0125] An attention layer is added after the output of the LSTM hidden layer to calculate the context vector c at the i-th time step. i :

[0126]

[0127] Where T is the total number of time steps, and the attention weight α is... ij Calculated in the following way:

[0128]

[0129] Among them, e ij To score attention, This represents the hidden state vector at the current moment. This represents the implicit state vector at a historical moment. For trainable weight matrix, For paranoia, The attention score vector is fed into a fully connected layer, and the activation function is LeakyReLU (α = 0.01).

[0130] S3b. An XGBoost algorithm is used to establish a passenger flow correlation correction module. The CO2 concentration change rate is used as an indirect passenger flow indicator. A gradient boosting decision tree model is constructed to predict the human body heat dissipation in the next 30 minutes.

[0131] S3c. The outputs of the two modules are fused using an adaptive weighting algorithm, with the weight coefficients dynamically adjusted based on the prediction error. An LSTM neural network captures load change patterns at key time steps through an attention mechanism. The XGBoost algorithm utilizes the CO2 concentration change rate to correct passenger flow-related heat dissipation predictions, improving the accuracy of dynamic load prediction. The dynamic weighted fusion strategy achieves collaborative optimization prediction of both base load and dynamic load.

[0132] S4. Design an optimized controller based on reinforcement learning algorithms, with recovery efficiency, energy consumption reduction rate, and equipment operation frequency as optimization objectives, and generate a control strategy for the heat recovery device; the design of the optimized controller based on reinforcement learning algorithms specifically includes the following steps:

[0133] S4a. Define the state space as [current recovered power, predicted load deviation, equipment efficiency coefficient], and the action space as [heat exchanger power adjustment amount, water pump speed adjustment level, bypass valve opening percentage];

[0134] S4b. Design the reward function R = α × recovery efficiency + β × energy consumption reduction rate - γ × equipment operation frequency, where α = 0.6, β = 0.3, γ = 0.1; the recovery efficiency η is calculated as follows:

[0135]

[0136] in, To actually recover heat, Q max Q represents the theoretical maximum recovery capacity. baseline Based on the recovered heat, c p T is the specific heat capacity of air at constant pressure, ρ is the density of air, and T is the specific heat capacity of air at constant in T out These are the inlet and outlet temperatures of the heat exchanger, The real-time air volume is calculated using the following formula:

[0137]

[0138] Where t is the current time, k is an integer loop variable, and 168 represents the upper limit of the loop in the summation operation, which is the moving average of the actual values ​​at the same time in the past 7 days.

[0139] Q max =P rated ·Δt·COP represents the theoretical maximum recovery capacity, P rated =50kW is the rated power of the heat recovery device, and COP=3.2 is the equipment performance coefficient.

[0140] S4c. The agent is trained using the Deep Deterministic Policy Gradient (DDPG) algorithm, with an experience replay pool capacity of 10,000 entries and a critic network learning rate of 0.001. A dynamic control policy generation mechanism is established to adapt to complex load fluctuations.

[0141] S5. Adjust the operating parameters of the energy recovery system according to the real-time load forecast deviation in stages, and realize dynamic power allocation through industrial communication protocols, including the implementation of dynamic control strategies, specifically including:

[0142] S5a. Execute a control decision every 5 minutes to generate a power allocation scheme for the heat recovery device;

[0143] S5b. When the predicted load deviation exceeds 15%, the emergency adjustment mode is triggered, and the operating status of the plate heat exchanger is adjusted first.

[0144] Emergency response mode includes a tiered response strategy:

[0145] Level 1 response (10% < deviation ≤ 15%):

[0146] Adjust the plate heat exchanger fan speed to 75% of the rated speed.

[0147] Cooling water flow rate adjustment formula:

[0148]

[0149] in, Q represents the adjusted flow rate. pred To predict the load, ΔQ = Q actual -Q pred For real-time load deviation, The reference water flow rate is ΔQ, and the load deviation is ΔQ.

[0150] Secondary response (deviation > 15%):

[0151] Activate the emergency cold storage tank to release the cooling capacity Q. storage =min(ΔQ·Δt,Q) max ), Δt is the time interval, Q max This represents the theoretical maximum heat recovery.

[0152] Trigger the frequency increase command for the ventilation system; frequency adjustment amount:

[0153] f new =f current +K p ·ΔQ K p =0.05Hz / kW

[0154] Among them, f new To set the new frequency value, f currentK represents the current frequency. p The proportionality constant is given, with the constraint: 20℃ ≤ T. supply ≤25℃, T supply This refers to the supply air temperature.

[0155] S5c sends PWM control signals to the frequency converter driver via the Modbus-TCP protocol to achieve precise adjustment of reclaimed power.

[0156] S5d. Execute the exception handling mechanism. When abnormal sensor data or equipment failure is detected, the data repair and security control mode is activated. The exception handling mechanism includes:

[0157] Data reliability assessment uses Mahalanobis distance detection:

[0158]

[0159] Where x is the observation vector, μ is the homogeneous vector, and ∑ is the covariance matrix, when When an anomaly is detected, n is the feature dimension. Let T be the chi-square critical value, and T be the matrix transpose operation;

[0160] The state equation for Kalman filter repair:

[0161] x k =Ax k-1 +Bu k +w k

[0162] z k =Hx k +v k

[0163] Where, x k Let A be the state variables, including temperature, humidity, and CO2 concentration, and B be the input matrix. k Let H represent the control input vector of the system at time k, and let w be the observation matrix. k For process noise, v k To monitor noise, a 5-minute dynamic control system is used to achieve fine-grained adjustment of the energy recovery system. A two-stage load deviation response strategy (15% threshold) ensures system stability: the first-stage response (10-15% deviation) adjusts the fan speed and compensates for the cooling water flow (with formula adjustment up to ±80%); the second-stage response (>15% deviation) activates the emergency cold storage tank (with cold energy release accuracy up to ±5%) and rapidly adjusts the frequency of the air supply system (Kp = 0.05Hz / kW); anomaly detection based on Mahalanobis distance and Kalman filter data repair improve system robustness.

[0164] Example 2

[0165] This embodiment provides an energy consumption management system for underground structures based on energy recovery, including:

[0166] Sensor network layer: includes temperature and humidity detection terminals, multispectral passenger flow counters, and equipment status monitoring units;

[0167] Edge computing layer: Equipped with an FPGA data preprocessing module and a timing database, wherein the FPGA module has a built-in sliding window differential coding circuit; the FPGA data preprocessing module includes:

[0168] The hardware logic implementation of the parallelized sliding window differential encoder is as follows:

[0169]

[0170]

[0171] The dynamic range adjustment circuit has the following adjustment rules:

[0172]

[0173] Among them, Gain current Gain is the current gain value. new The adjusted gain value, σ 2 The variance of the signal, 2V 2 For the set variance threshold, the temperature drift compensation module uses the following compensation formula:

[0174] V corrected =V raw ×(1+αΔT+βΔT 2 )

[0175] Among them, V corrected V is the voltage value after temperature drift compensation. raw The original voltage signal is given, ΔT is the temperature change, and α and β are compensation coefficients.

[0176] α = 2.1 × 10 -4 / ℃, β=8.7×10 -7 / ℃2.

[0177] Cloud-based analytics layer: Deploys a load forecasting engine, a reinforcement learning decision model, and digital twins of equipment. The three-dimensional thermodynamic simulation model of the equipment digital twins uses unsteady-state heat transfer equations.

[0178]

[0179] Where ρ is the material density, c p q is the specific heat capacity, k is the thermal conductivity, and q is the thermal conductivity. gen For the heat source generation term, qloss This is the heat loss term, where T is temperature and t is time. For Hamiltonian operators, Let T be the partial derivative of temperature T with respect to time t;

[0180] Boundary conditions:

[0181] Platform wall: Type III boundary conditions

[0182] Where k is the thermal conductivity. Let T represent the directional derivative of temperature T along the wall normal n, and h be the convective heat transfer coefficient. air For air temperature, T wall The temperature of the platform wall;

[0183] Equipment surface: Second type boundary condition q″=520W / m 2 q″ is the heat flux density;

[0184] Mesh generation: unstructured mesh, minimum size 0.1m, time step Δt = 10s;

[0185] The execution control layer includes an intelligent frequency converter cabinet, an electric regulating valve group, and a fault diagnosis module. The fault diagnosis module's three-level early warning mechanism includes:

[0186] Level 1 Warning: Generate equipment efficiency trend chart:

[0187]

[0188] Where, η trend (t) represents the trend value of equipment efficiency at time t. Let η be a coefficient, k be an integer variable, Δt be the time interval, and η(t-kΔt) be the equipment efficiency at time t-kΔt. For equipment efficiency trend η trend The reciprocal of (t), when the slope Maintenance reminders will be issued in a timely manner;

[0189] Level 2 warning: Perform wavelet packet decomposition on the vibration signal and extract the energy ratio in the 3-5kHz frequency band.

[0190]

[0191] Where f is the frequency, PSD(f) is the power spectral density function, and 3k, 5k, and 10k are 3000Hz, 5000Hz, and 10000Hz, respectively.

[0192] When E ratio Bearing wear is determined when the percentage is >15%;

[0193] Level 3 warning: Insulation resistance is measured using the DC superposition method.

[0194]

[0195] Among them, V DC I is a DC superposition voltage. leakage For leakage current, R shunt As the shunt resistor, when R iso The circuit is triggered to stop when the resistance is less than 50MΩ.

[0196] The specific implementation method of the energy consumption management system for underground structures based on energy recovery in this embodiment will be described in detail below.

[0197] I. Sensor Network Layer Implementation

[0198] Equipment selection:

[0199] Temperature and humidity monitoring terminal: Employs SHT30 digital sensor, accuracy ±0.3℃ / ±2% RH

[0200] Multispectral passenger flow counter: Utilizes a TOF laser + RGB camera fusion solution, with a detection range of 0-10 meters.

[0201] Equipment condition monitoring unit: integrates vibration acceleration sensor (ADXL355) and current transformer (LEM LTSR50NP)

[0202] Deployment plan:

[0203] Temperature and humidity sensors are installed along the tunnel wall at 10-meter intervals.

[0204] Passenger flow counters are deployed at key locations such as escalator entrances and turnstiles;

[0205] The equipment status monitoring unit is installed near energy-consuming equipment such as ventilation units and water pumps;

[0206] II. Implementation of Edge Computing Layer

[0207] FPGA hardware design:

[0208] It uses Xilinx Zynq UltraScale+ MPSoC;

[0209] Implement a 16-bit parallel sliding window differential encoder with a 15-stage pipeline structure with a window depth;

[0210] Dynamic range adjustment module:

[0211] Variance calculation is implemented using a 32-bit floating-point arithmetic unit;

[0212] Gain adjustment step: 1.5x / 0.7x, maximum gain limit set to 1000x;

[0213] Temperature compensation module:

[0214] Integrated DS18B20 temperature sensor;

[0215] Hardware multipliers are used to perform quadratic polynomial compensation calculations.

[0216] Time series database configuration:

[0217] The TimescaleDB time-series database is used.

[0218] Data storage strategy: Raw data is retained for 7 days, and preprocessed data is retained for 30 days;

[0219] Data compression: LZ4 compression algorithm is used, with a compression ratio of approximately 3:1;

[0220] III. Implementation of Cloud-based Analytics Layer

[0221] Load forecasting engine:

[0222] The LSTM neural network model is used, and the input features include:

[0223] Historical energy consumption data (past 24 hours);

[0224] Real-time temperature and humidity data;

[0225] Passenger flow density data;

[0226] Forecast period: 15-minute increments, forecasting load for the next 4 hours;

[0227] Reinforcement learning decision-making model:

[0228] State space: includes 12-dimensional features such as current energy consumption, equipment status, and environmental parameters; Action space: frequency of frequency converter (20-60Hz) and opening degree of regulating valve (0-100%); Reward function: a weighted value based on energy recovery efficiency and equipment life loss;

[0229] Digital twin construction:

[0230] A three-dimensional thermodynamic model was established using COMSOL Multiphysics;

[0231] Mesh generation: Tetrahedral mesh, minimum size 0.1m, boundary layer refinement; Solver configuration:

[0232] Time step: 10 seconds;

[0233] Steady-state solver: PARDISO direct solver;

[0234] Transient solver: BDF method;

[0235] IV. Implementation of the execution control layer;

[0236] Intelligent frequency converter cabinet:

[0237] Select ABB ACS880 series frequency converters;

[0238] Communication protocol: Modbus RTU over RS485;

[0239] Control mode: Closed-loop vector control that accepts commands from the cloud;

[0240] Electric regulating valve assembly:

[0241] SMC VX3000 series valve positioners are used;

[0242] Control accuracy: ±0.5% opening degree;

[0243] Drive method: 4-20mA analog signal;

[0244] Fault diagnosis module:

[0245] Level 1 Warning:

[0246] Data acquisition frequency: 1Hz

[0247] Trend calculation window: 5 time points (Δt = 1 hour) Level 2 alert:

[0248] Vibration signal sampling rate: 48kHz

[0249] Wavelet packet decomposition levels: 3

[0250] Energy ratio calculation window: 1 second

[0251] Level 3 Warning:

[0252] DC superimposed voltage: 500V DC;

[0253] Insulation resistance measurement cycle: 1 time / hour;

[0254] V. System Integration and Implementation

[0255] Communication architecture:

[0256] Edge layer to cloud: MQTT protocol, QoS1 guaranteed transmission; Device layer to edge layer: CAN bus, 1Mbps baud rate; Backup communication: LoRaWAN wide area network link;

[0257] Security mechanisms:

[0258] Data encryption: AES-256 encrypted transmission;

[0259] Identity authentication: Two-way authentication based on X.509 certificates;

[0260] Anomaly traffic detection: Deploy an intrusion detection system (Snort); Testing and verification:

[0261] Laboratory testing: Establishing a 1:10 scale model;

[0262] On-site testing: A typical subway station was selected for a 6-month trial operation; verification indicators:

[0263] Energy recovery efficiency ≥35%;

[0264] Fault warning accuracy rate ≥98%;

[0265] System response time ≤ 200ms;

[0266] VI. Implementation of Operation and Maintenance Management

[0267] Visualization platform:

[0268] 3D visualization interface: CesiumJS engine;

[0269] Data dashboard: Grafana real-time monitoring;

[0270] Alarm push notifications: DingTalk robot + SMS dual channels;

[0271] Maintenance strategy:

[0272] Preventive maintenance: Predictive maintenance based on a three-level early warning system;

[0273] Knowledge base management: Establish a database of equipment failure cases;

[0274] Remote upgrade: Supports OTA firmware updates;

[0275] This implementation method covers the entire lifecycle management from hardware selection to system operation and maintenance, and achieves efficient energy consumption management and equipment health monitoring of underground structures through multi-level collaborative optimization.

[0276] The above specific embodiments are merely several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A method for energy consumption management of underground structures based on energy recovery, characterized in that, Includes the following steps: S1. Deploy a multi-source sensor network in key areas of the underground structure to collect environmental parameters, passenger flow heat source distribution and equipment operation status data in real time; S2. Perform spatiotemporal feature fusion processing on sensor data to extract a multidimensional vector containing time period features, spatial correlation features, and device coupling features; S3. Construct a hybrid forecasting model, predict the basic load through an LSTM neural network, and combine it with the XGBoost algorithm to correct the heat dissipation associated with passenger flow, and dynamically weight the output of the comprehensive load forecasting result; S4. Design and optimize the controller based on reinforcement learning algorithm, with the optimization objectives of recovery efficiency, energy consumption reduction rate and equipment operation frequency, and generate control strategy for heat recovery device; S5. Adjust the operating parameters of the energy recovery system according to the real-time load forecast deviation and realize dynamic power distribution through industrial communication protocol.

2. The energy consumption management method for underground structures based on energy recovery according to claim 1, characterized in that, S1 specifically includes: constructing a multi-source heterogeneous sensor monitoring network, deploying temperature and humidity sensors, CO2 concentration sensors, infrared thermal imagers, and equipment current transformers in the target area, and synchronously collecting environmental parameters, passenger flow heat source distribution, and equipment operation status data, wherein the target area includes at least the subway platform level, transfer hall, and equipment room. The deployment method of the infrared thermal imager includes: arranging 3 thermal imagers at the vertices of an equilateral triangle along the main passageway on the ceiling of the station hall, with the side length of the triangle being 12±0.5 meters, the elevation angle of each detection point being adjusted to 30°±2°, and the horizontal scanning angle being set to 120°. The adaptive background subtraction algorithm is used to extract dynamic heat sources. The calculation formula is as follows: Where I(x,y,t) is the radiant intensity value of the (x,y) coordinates at time t, N=30 is the number of background modeling frames, and G σ (x,y) is the Gaussian filter kernel with a standard deviation σ = 1.5 pixels, ΔT is the temperature difference score, and k is the historical frame index; When ΔT>5℃, it is determined to be an effective human body heat source, and the spatial positioning error is less than 0.3 meters.

3. The energy consumption management method for underground structures based on energy recovery according to claim 1, characterized in that, In step S2, the spatiotemporal feature fusion processing of the sensor data specifically includes the following steps: S2a. Perform sliding window standardization on the raw data, with the window length set to 15 minutes, and use the Z-score normalization method to eliminate dimensional differences; S2b. Extract multi-dimensional feature vectors, including time dimension features, spatial topology features, and device association features. The time dimension features include historical 24-hour load data of the same period, and the spatial topology features are established by using the Delaunay triangulation algorithm to build a sensor spatial association matrix.

4. The energy consumption management method for underground structures based on energy recovery according to claim 1, characterized in that, In step S3, constructing the hybrid prediction model specifically includes the following steps: S3a. An LSTM neural network is used to establish a basic load prediction module. The network structure contains 3 hidden layers, each with 128 neurons. The input layer receives the feature vector sequence of the previous 2 hours. The LSTM neural network employs an improved structure with an attention mechanism, specifically including: An attention layer is added after the output of the LSTM hidden layer to calculate the context vector c at the i-th time step. i : Where T is the total number of time steps, and the attention weight α is... ij Calculated in the following way: Among them, e ij To score attention, This represents the hidden state vector at the current moment. This represents the implicit state vector at a historical moment. For trainable weight matrix, For paranoia, The attention score vector is fed into a fully connected layer, and the activation function is LeakyReLU. S3b. An XGBoost algorithm is used to establish a passenger flow correlation correction module. The CO2 concentration change rate is used as an indirect passenger flow indicator. A gradient boosting decision tree model is constructed to predict the human body heat dissipation in the next 30 minutes. S3c. The outputs of the two modules are fused using an adaptive weighting algorithm, and the weighting coefficients are dynamically adjusted according to the prediction error.

5. The energy consumption management method for underground structures based on energy recovery according to claim 1, characterized in that, The S4 step involves designing and optimizing the controller based on a reinforcement learning algorithm, specifically including the following steps: S4a. Define the state space as [current recovered power, predicted load deviation, equipment efficiency coefficient], and the action space as [heat exchanger power adjustment amount, water pump speed adjustment level, bypass valve opening percentage]; S4b. Design the reward function R = α × recovery efficiency + β × energy consumption reduction rate - γ × equipment operation frequency, where α = 0.6, β = 0.3, γ = 0.1; the recovery efficiency η is calculated as follows: in, To actually recover heat, Q max Q represents the theoretical maximum recovery capacity. baseline Based on the recovered heat, c p T is the specific heat capacity of air at constant pressure, ρ is the density of air, and T is the specific heat capacity of air at constant pressure. in T out These are the inlet and outlet temperatures of the heat exchanger, The real-time air volume is calculated using the following formula: Where t is the current time, k is an integer loop variable, and 168 represents the upper limit of the loop in the summation operation, which is the moving average of the actual values ​​at the same time in the past 7 days. Q max =P rated ·Δt·COP represents the theoretical maximum recovery capacity, P rated =50kW is the rated power of the heat recovery device, and COP=3.2 is the coefficient of performance of the equipment; S4c. The agent is trained using a deep deterministic policy gradient algorithm, with an experience replay pool capacity of 10,000 entries and a critic network learning rate of 0.

001.

6. The energy consumption management method for underground structures based on energy recovery according to claim 1, characterized in that, S5 includes implementing a dynamic control strategy, specifically including: S5a. Execute a control decision every 5 minutes to generate a power allocation scheme for the heat recovery device; S5b. When the predicted load deviation exceeds 15%, the emergency adjustment mode is triggered, and the operating status of the plate heat exchanger is adjusted first. Emergency response mode includes a tiered response strategy: Level 1 response, 10% < deviation ≤ 15% Adjust the plate heat exchanger fan speed to 75% of the rated speed; Cooling water flow rate adjustment formula: in, Q represents the adjusted flow rate. pred To predict the load, ΔQ = Q actual -Q pred For real-time load deviation, The reference water flow rate is ΔQ, and the load deviation is ΔQ. Level 2 response, deviation >15% Activate the emergency cold storage tank to release the cooling capacity Q. storage =min(ΔQ·Δt,Q) max ), Δt is the time interval, Q max This represents the theoretical maximum heat recovery. Trigger the frequency increase command for the ventilation system; frequency adjustment amount: f new =f current +K p ·ΔQ K p =0.05Hz / kW Among them, f new Set the value for the new frequency, f current K represents the current frequency. p The proportionality constant is given, with the constraint: 20℃ ≤ T. supply ≤25℃, T supply This refers to the supply air temperature. S5c. Sends PWM control signals to the frequency converter driver via Modbus-TCP protocol; S5d. Execute the exception handling mechanism. When abnormal sensor data or equipment failure is detected, the data repair and security control mode is activated. The exception handling mechanism includes: Data reliability assessment uses Mahalanobis distance detection: Where x is the observation vector, μ is the homogeneous vector, and ∑ is the covariance matrix, when When an anomaly is detected, n is the feature dimension. Let T be the chi-square critical value, and T be the matrix transpose operation; The state equation for Kalman filter repair: x k =Ax k-1 +Bu k +w k z k =Hx k +v k Where, x k Let A be the state variables, including temperature, humidity, and CO2 concentration, and B be the input matrix. k Let H represent the control input vector of the system at time k, and let w be the observation matrix. k For process noise, v k To observe noise.

7. An energy consumption management system for underground structures based on energy recovery, characterized in that, include: Sensor network layer: includes temperature and humidity detection terminals, multispectral passenger flow counters, and equipment status monitoring units; Edge computing layer: Equipped with an FPGA data preprocessing module and a timing database, wherein the FPGA module has a built-in sliding window differential coding circuit; Cloud analytics layer: Deploy load forecasting engines, reinforcement learning decision models, and device digital twins; The execution control layer includes intelligent frequency converter cabinets, electric regulating valve groups, and fault diagnosis modules; The edge computing layer is connected to the sensor network layer via the OPC UA protocol, and the cloud analytics layer interacts with the edge computing layer via the MQTT protocol.

8. The energy management system for underground structures based on energy recovery according to claim 7, characterized in that, The FPGA data preprocessing module includes: Parallelized sliding window differential encoder, The dynamic range adjustment circuit has the following adjustment rules: Among them, Gain current Gain is the current gain value. new The adjusted gain value, σ 2 The variance of the signal, 2V 2 For the set variance threshold, the temperature drift compensation module uses the following compensation formula: V corrected =V raw ×(1+αΔT+βΔT 2 ) Among them, V corrected V is the voltage value after temperature drift compensation. raw The original voltage signal is given, ΔT is the temperature change, and α and β are compensation coefficients. α＝2.1×10 -4 / ℃,β＝8.7×10 -7 / ℃2。 9. The energy management system for underground structures based on energy recovery according to claim 8, characterized in that, The three-dimensional thermodynamic simulation model of the device's digital twin adopts an unsteady heat transfer equation: Where ρ is the material density, c p q is the specific heat capacity, k is the thermal conductivity, and q is the thermal conductivity. gen For the heat source generation term, q loss This is the heat loss term, where T is temperature and t is time. For Hamiltonian operators, Let T be the partial derivative of temperature T with respect to time t; Boundary conditions: Platform wall: Type III boundary conditions Where k is the thermal conductivity. Let T represent the directional derivative of temperature T along the wall normal n, and h be the convective heat transfer coefficient. air For air temperature, T wall The temperature of the platform wall; Equipment surface: Second type boundary condition q″=520W / m 2 q″ is the heat flux density; Mesh generation: unstructured mesh, minimum size 0.1m, time step Δt = 10s.

10. The energy management system for underground structures based on energy recovery according to claim 9, characterized in that, The three-level early warning mechanism of the fault diagnosis module includes: Level 1 Warning: Generate equipment efficiency trend chart: Where, η trend (t) represents the trend value of equipment efficiency at time t. Let η be a coefficient, k be an integer variable, Δt be the time interval, and η(t-kΔt) be the equipment efficiency at time t-kΔt. For equipment efficiency trend η trend The reciprocal of (t), when the slope Maintenance reminders will be issued in a timely manner; Level 2 warning: Perform wavelet packet decomposition on the vibration signal and extract the energy ratio in the 3-5kHz frequency band. Where f is the frequency, PSD(f) is the power spectral density function, and 3k, 5k, and 10k are 3000Hz, 5000Hz, and 10000Hz, respectively. When E ratio Bearing wear is determined when the percentage is >15%; Level 3 warning: Insulation resistance is measured using the DC superposition method. Among them, V DC I is a DC superposition voltage. leakage For leakage current, R shunt As the shunt resistor, when R iso The circuit is triggered to stop when the resistance is less than 50MΩ.

Images