Deep underground space environment comfort level self-adaptive regulation and control system and method based on improved gating circulation unit

Abstract

The invention discloses a deep underground space environment comfort adaptive regulation and control system and method based on an improved gating circulation unit, and the method comprises the steps: collecting multi-source environment parameters of a deep underground space in real time through a multi-source data collection and preprocessing module, and converting the parameters into a time sequence window to generate time sequence window data; then, according to the data acquired and processed by the multi-source data acquisition and preprocessing module, an environment parameter prediction module extracts dynamic fractal features of the multi-source environment data through an improved gating cycle unit (E-GRU), and generates a future environment parameter prediction value; and then, through a reinforcement learning strategy generation module, based on an environmental parameter prediction value generated by an environmental parameter prediction module and real-time energy consumption, generating an equipment control instruction, and through a reward function, balancing comfort and energy consumption. According to the scheme, on the premise that comfort is guaranteed, energy consumption can be remarkably reduced, the response speed is increased, and the method is suitable for high-requirement scenes such as deep underground laboratories.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent building control, and particularly relates to the multi-parameter collaborative regulation technology for the deep underground space environment. Background Art

[0002] Traditional underground space environment control systems at a depth of 100 - 200m mostly adopt methods such as PID control and fuzzy control. These methods have disadvantages such as slow convergence speed and poor robustness when controlling relatively complex systems or multiple controlled objects.

[0003] Due to the strong temporal dependence of environmental parameters, traditional control schemes (such as PID control) cannot effectively capture the correlation between long-term historical states and current inputs, resulting in control lag.

[0004] Furthermore, there is a contradiction between temperature regulation and energy consumption control. Traditional control schemes lack a dynamic trade-off mechanism and are prone to falling into local optima, that is, there is a problem of multi-objective conflict.

[0005] In recent years, control methods based on machine learning and deep learning have gradually been applied to the field of indoor environment control. Among them, the gated recurrent unit (GRU), as an effective sequence data processing model, has demonstrated excellent performance in multiple fields. However, the traditional GRU model still has certain limitations when processing indoor environment data with multifractal characteristics.

[0006] For example, the underground space environment comfort improvement scheme disclosed in the Chinese patent application with the publication number CN112032982A specifically adopts the On-policy Monte Carlo (OMC) algorithm. This scheme relies on random sampling and policy iteration, resulting in a slow convergence speed and low computational efficiency in a high-dimensional state space, and it is difficult to cope with the dynamic mutations of the underground space environment, such as a sudden increase in CO2 caused by a crowded population.

[0007] It can be seen that for the underground space environment comfort regulation scheme, how to effectively improve the comfort of the deep underground space environment and reduce energy consumption simultaneously is an urgent problem to be solved in this field. Summary of the Invention

[0008] Aiming at the problems existing in the existing deep underground space environment control scheme, the purpose of the present invention is to provide an adaptive regulation scheme for the comfort of the deep underground space environment. Based on a two-stage improved gated recurrent unit, by integrating dynamic fractal feature extraction, spatio-temporal attention mechanism and multi-objective optimization strategy, it realizes real-time adaptive adjustment of temperature, humidity, illuminance and air quality, and significantly improves comfort and energy efficiency ratio.

[0009] To achieve the above object, the present invention provides an adaptive regulation system for the comfort of deep underground space environment based on an improved gated recurrent unit, and the regulation system includes:

[0010] A multi-source data acquisition and preprocessing module that collects multi-source environmental parameters of deep underground space in real time and converts them into a time series window to generate time series window data;

[0011] An environmental parameter prediction module that extracts dynamic fractal features of multi-source environmental data and generates future environmental parameter prediction values through an improved gated recurrent unit (Enhanced GRU, E-GRU) according to the data collected and processed by the multi-source data acquisition and preprocessing module;

[0012] A reinforcement learning policy generation module that generates device control instructions based on the environmental parameter prediction values generated by the environmental parameter prediction module and real-time data (including real-time energy consumption and temperature change trends), and balances comfort and energy consumption through a reward function.

[0013] In some embodiments of the present invention, the improved gated recurrent unit includes an adaptive reset gate (ARG) module, a bidirectional feature fusion (BFF) module, a spatio-temporal attention (STA) module, and a multi-task prediction (MTH) module.

[0014] The adaptive reset gate (ARG) module is used to perceive the mutation characteristics of input data in real time, dynamically adjust the forgetting intensity of the reset gate, and filter abnormal data;

[0015] The bidirectional feature fusion (BFF) module supplements local window features based on the hidden state formed after filtering and purification by the adaptive reset gate (ARG) module;

[0016] The spatio-temporal attention (STA) module performs spatio-temporal weight allocation on the features output by the bidirectional feature fusion (BFF) module to extract key feature information;

[0017] The multi-task prediction (MTH) module generates future state prediction values of each parameter based on the features extracted by the spatio-temporal attention (STA) module.

[0018] In some embodiments of the present invention, the adaptive reset gate (ARG) module is configured such that when the change rate of input data exceeds the historical threshold θ Δx the reset gate weight r t is set to 0, and the historical hidden state h t-1 .

[0019] In some embodiments of the present invention, the Bidirectional Feature Fusion (BFF) module is configured to intercept window data of a total of 7 steps before and after centered on the current moment, flatten the data within the window, and splice it with the hidden state h' output by the Adaptive Reset Gate (ARG) module t ; and input the spliced feature vector into the update gate z t , and the calculation method is:

[0020] z t =σ(W z ·[h', t BFF(x t-3:t+3 )]+b z ),

[0021] where BFF( xt-3:t+3 ) is the feature after window flattening.

[0022] In some embodiments of the present invention, when the Spatio-Temporal Attention (STA) module performs weight allocation, it includes:

[0023] Temporal attention weight calculation: Calculate the correlation weight between the historical moment and the current moment. Through the learnable query matrix Q and key matrix K, map the hidden state to a high-dimensional space and calculate the similarity score:

[0024]

[0025] Spatial attention weight calculation: Evaluate the importance of each sensor parameter and generate weights through a fully connected layer:

[0026]

[0027] Weight fusion: Multiply the temporal and spatial weights and then normalize them, and perform weighted summation to obtain the final hidden state h t , and the formula is:

[0028]

[0029] In some embodiments of the present invention, the Multi-Task Prediction (MTH) module predicts the future state prediction value for each environmental parameter through an independent fully connected layer respectively.

[0030] In some embodiments of the present invention, the reinforcement learning policy generation module includes

[0031] a state space unit that fuses the environmental parameter prediction value generated by the environmental parameter prediction module with real-time data (including real-time energy consumption and temperature change trends) to construct a corresponding feature vector;

[0032] an action space unit that encodes the control instructions of the device into discrete actions to construct a feasible operation set of the device control instructions;

[0033] A reward function unit that calculates an immediate reward value based on the constructed reward function according to the environmental parameters in the state space and the execution results of the action space;

[0034] A deep Q-network (DQN) that calculates and generates a target Q-value based on the high-dimensional feature vector in the state space unit, directly maps the target Q-value to each discrete action in the action space, and selects the action with the largest Q-value as the optimal action for execution; after the action is executed, the deep Q-network (DQN) updates and optimizes the deep Q-network parameters according to the immediate reward value calculated by the reward function unit based on the environmental parameters in the current state space and the execution results of the latest action space.

[0035] To achieve the above object, the present invention provides a deep underground space environment regulation method based on an improved GRU, including:

[0036] Extracting the dynamic fractal features of multi-source environmental data and generating future environmental parameter prediction values through an improved gated recurrent unit (Enhanced GRU, E-GRU);

[0037] Generating device control instructions through a deep Q-network (DQN) based on the environmental parameter prediction values and real-time data (including real-time energy consumption and temperature change trends), and balancing comfort and energy consumption through a reward function.

[0038] In some embodiments of the present invention, the process of generating future environmental parameter prediction values in the deep underground space environment regulation method includes:

[0039] The adaptive reset gate (ARG) module senses the mutation characteristics of the input data in real time, dynamically adjusts the forgetting intensity of the reset gate, and filters abnormal data;

[0040] The bidirectional feature fusion (BFF) module supplements local window features based on the hidden state formed after the adaptive reset gate (ARG) module filters and purifies;

[0041] The spatio-temporal attention (STA) module performs spatio-temporal weight allocation on the features output by the bidirectional feature fusion (BFF) module to extract key feature information;

[0042] The multi-task prediction module (MTH) generates future state prediction values of each parameter based on the features extracted by the spatio-temporal attention (STA) module.

[0043] In some embodiments of the present invention, the process of generating device control instructions in the deep underground space environment regulation method includes:

[0044] In the state space, the predicted environmental parameter values generated by the environmental parameter prediction module are fused with real-time data (including real-time energy consumption and temperature change trends) to construct corresponding feature vectors;

[0045] The deep Q-network (DQN) calculates and generates target Q-values based on the feature vectors in the state space, directly maps the target Q-values to each discrete action in the action space, and selects the action with the largest Q-value as the optimal action for execution;

[0046] After the action is executed, the deep Q-network (DQN) updates and optimizes the deep Q-network parameters according to the immediate reward value calculated by the reward function based on the environmental parameters in the current state space and the execution results of the latest action space.

[0047] The self-adaptive regulation scheme for the environmental comfort of deep underground spaces provided by the present invention can significantly reduce energy consumption and improve the response speed while ensuring comfort, and is applicable to high-demand scenarios such as deep underground laboratories.

[0048] Furthermore, the solution of the present invention extracts spatio-temporal features through an enhanced gated recurrent unit (Enhanced GRU, E-GRU), combines reinforcement learning to achieve multi-objective optimization, reduces energy consumption by 18.7% and improves the response speed by 56% compared with traditional methods. Description of the Drawings

[0049] The present invention will be further described below in conjunction with the drawings and specific embodiments.

[0050] Figure 1 It is a schematic diagram of the composition of the self-adaptive regulation system for the environmental comfort of deep underground spaces in the present invention;

[0051] Figure 2 It is a schematic diagram of the composition of the multi-source data acquisition and preprocessing module in the present invention;

[0052] Figure 3 It is a schematic diagram of the composition of the environmental parameter prediction module in the present invention;

[0053] Figure 4 It is a schematic diagram of the composition of the reinforcement learning policy generation module in the present invention. Detailed Embodiments

[0054] In order to make the technical means, creative features, achieved purposes and effects of the present invention easy to understand, the present invention will be further described below in conjunction with specific drawings.

[0055] The self - adaptive regulation scheme for the comfort of deep underground space environment given by the present invention constructs a closed - loop control process of data perception, time - series modeling, strategy generation, and online optimization based on an improved gated recurrent unit, thereby significantly reducing energy consumption and improving response speed while ensuring comfort.

[0056] Accordingly, the present invention forms a self - adaptive regulation system for the comfort of deep underground space environment based on an improved gated recurrent unit. This system deeply extracts the environmental time - series characteristics through an improved GRU network and integrates a reinforcement learning framework to generate a multi - objective collaborative control strategy, which can significantly reduce energy consumption and improve response speed while ensuring comfort.

[0057] See Figure 1 , which shows the system block diagram of the self - adaptive regulation system for the comfort of deep underground space environment based on an improved gated recurrent unit given by the present invention.

[0058] Based on the illustration, the self - adaptive regulation system 100 for the comfort of deep underground space environment based on an improved gated recurrent unit mainly consists of four functional modules: a multi - source data acquisition and pre - processing module 110, an environmental parameter prediction module 120, a reinforcement learning strategy generation module 130, and an online adaptive parameter - tuning module 140, which cooperate with each other.

[0059] The multi - source data acquisition and pre - processing module 110 in this system can pre - process multi - source environmental parameters of the deep underground space based on a variety of environmental monitoring sensors and can convert them into a time - series window to generate time - series window data.

[0060] The environmental parameter prediction module 120 in this system can form an effective data interaction with the multi - source data acquisition and pre - processing module 110, and can extract the dynamic fractal characteristics of multi - source environmental data and generate future environmental parameter prediction values through an improved gated recurrent unit (Enhanced GRU, E - GRU) according to the data collected and processed by the multi - source data acquisition and pre - processing module.

[0061] The reinforcement learning strategy generation module 130 in this system can form an effective data interaction with the multi - source data acquisition and pre - processing module 110 and the environmental parameter prediction module 120, and can generate device control instructions based on the environmental parameter prediction values generated by the environmental parameter prediction module and real - time data (including real - time energy consumption and temperature change trends), and balance comfort and energy consumption through a reward function.

[0062] The online adaptive parameter tuning module 140 in this system can form effective data interaction with the multi-source data acquisition and preprocessing module 110 and the environmental parameter prediction module 120, and can dynamically update the model parameters of the enhanced gated recurrent unit (E-GRU) in the environmental parameter prediction module 120 by monitoring environmental changes, ensuring long-term control effects.

[0063] For each functional module in this deep underground space environment comfort adaptive control system 100, the present invention further provides a specific implementation solution.

[0064] The multi-source data acquisition and preprocessing module 110 in this system constitutes the data acquisition layer in the system, can acquire the environmental parameters of the environment to be controlled, and can convert the original sensor data into a standardized and denoised time series window, providing high-quality data input for the enhanced gated recurrent unit model in the environmental parameter prediction module 120.

[0065] Specifically, see Figure 2 This multi-source data acquisition and preprocessing module 110 mainly consists of several sensors 111 and a data preprocessing unit 112.

[0066] These several sensors 111 mainly include environmental sensors and equipment status monitoring sensors.

[0067] Among them, the environmental sensors include sensors such as temperature, humidity, CO2 concentration, and light intensity, which are deployed at the monitoring site for real-time sampling to form environmental parameters. Preferably, Kalman filtering is used to eliminate noise for the collected data.

[0068] The equipment status monitoring sensors are used to monitor the working status of environmental control equipment, such as equipment parameters such as the power of the air conditioner compressor, the opening degree of the fresh air damper, and the lighting brightness. Preferably, it is read in real time through the Modbus protocol, and the sampling frequency is synchronized with the environmental sensors.

[0069] The data preprocessing unit 112 establishes a data connection with several sensors 111, and can perform data cleaning, outlier processing, data normalization, and time series window construction on the data collected and uploaded by several sensors 111.

[0070] Furthermore, when this data preprocessing unit 112 performs data cleaning and outlier processing, the 3σ criterion (three standard deviation principle) is used to identify outliers. The historical mean and standard deviation (such as based on the data of the past 7 days) are calculated for each parameter (temperature, humidity, CO2, illuminance) respectively. If a data point at a certain moment meets the following conditions, it is determined to be an outlier. The specific implementation is as follows:

[0071]

[0072] Furthermore, the data preprocessing unit 112 also fills in the outliers for the identified abnormal data points. Here, the sliding median of its previous 3 time steps is used for replacement. For example, if the temperature data is abnormal at time t, then take T t-3 , T t-2 , T t-1 's median as the correction value.

[0073] Furthermore, when the data preprocessing unit 112 performs data normalization, each parameter is mapped to the interval [0, 1] to eliminate the dimension difference.

[0074] The normalization formula here is:

[0075] At the same time, as an example, the parameter range is set as:

[0076] Temperature: 10°C to 40°C;

[0077] CO2 concentration: 400 ppm to 2000 ppm;

[0078] Humidity: 0% to 100%;

[0079] Illuminance: 0 Lux to 2000 Lux.

[0080] Furthermore, when the data preprocessing unit 112 constructs a time series window, centered on the current moment t, it intercepts the past 9 time steps and the current time step to form a time series window with a length of 10. At the same time, the input tensor is:

[0081]

[0082] This multi-source data acquisition and preprocessing module 110 can output the standardized time series window data, which is used as the input of the improved gated recurrent unit (Enhanced GRU, E-GRU) in the environmental parameter prediction module 120 to ensure that the model can capture the time series dependence relationship.

[0083] See Figure 3 , in this system, the improved gated recurrent unit adopted by the environmental parameter prediction module 120 is specifically composed of four functional modules: the adaptive reset gate (ARG) module 121, the bidirectional feature fusion (BFF) module 122, the spatio-temporal attention (STA) module 123, and the multi-task prediction (MTH) module 124, thereby realizing the deep modeling of environmental time series data.

[0084] The adaptive reset gate (ARG) module 121 in this improved gated recurrent unit is used to perceive the mutation characteristics of the input data in real time, dynamically adjust the forgetting intensity of the reset gate, and filter out abnormal data.

[0085] To effectively and dynamically respond to data mutations (such as a sudden increase in temperature), this Adaptive Reset Gate (ARG) module 121 enhances the sensitivity to emergencies by introducing the input change rate as a gating factor. The specific formula is as follows:

[0086] r t =σ(W r ·[h t-1 ,x t +b r +α·△x t )

[0087] Where:

[0088] is the k-step moving average change rate of the input data (default k = 3);

[0089] α is a learnable parameter, initialized to 0.1, and optimized through backpropagation.

[0090] Furthermore, when the Adaptive Reset Gate (ARG) module 121 is running, it first calculates the input change rate and calculates the absolute change mean within the sliding window for each sensor parameter (temperature, humidity, etc.); then it dynamically adjusts the forgetting intensity. When Δ xt > θ (the threshold is set to the 90th percentile of the historical change rate), the reset gate weight r t tends to 0, forcing the forgetting of the historical state to cope with mutations.

[0091] The Bidirectional Feature Fusion (BFF) module 122 in this improved gated recurrent unit can supplement local window features based on the hidden state formed after filtering and purification by the Adaptive Reset Gate (ARG) module 121. This Bidirectional Feature Fusion (BFF) module 122 implicitly introduces backward information through a bidirectional sliding window, enhancing the modeling ability for periodic patterns (such as diurnal temperature changes). The specific mathematical formula is as follows:

[0092] z t =σ(W z ·[h t-1 ,x t-k:t+k ),

[0093] Where, x t-k:t+k =[x t-k ,x t-k+1 ,...,x t,..., x t+k is the symmetric window data centered on the current time step t;

[0094] The window length 2k + 1 is defaulted to 7 (i.e., 3 time steps before and after, a total of 15 minutes of data).

[0095] Furthermore, when the Bidirectional Feature Fusion (BFF) module 122 is running, it first constructs a data window: centered on the current moment, intercepts data for the k steps before and after to form a local window. If future data is unavailable (during real-time prediction), linear extrapolation is used for filling, that is, linear interpolation is adopted; then feature concatenation: flattens the window data into a vector and concatenates it with the previous hidden state h t-1 After concatenation, it is input into the update gate.

[0096] The Spatio-Temporal Attention (STA) module 123 in this improved gated recurrent unit is used to assign spatio-temporal weights to the features output by the Bidirectional Feature Fusion (BFF) module and extract key feature information.

[0097] Since different sensors (spatial dimension) and historical moments (time dimension) have different importance for the current prediction. This Spatio-Temporal Attention (STA) module 123 focuses on key spatio-temporal nodes through two-level attention weights.

[0098] Specifically, the Spatio-Temporal Attention (STA) module 123 assigns spatio-temporal weights as follows:

[0099] Temporal attention calculation:

[0100] Calculate the correlation weights between historical moments and the current moment. Through the learnable query matrix Q and key matrix K, map the hidden state to a high-dimensional space and calculate the similarity score:

[0101]

[0102] Spatial attention calculation:

[0103] Evaluate the importance of each sensor parameter and generate weights through a fully connected layer:

[0104]

[0105] Weight fusion:

[0106] After multiplying the spatio-temporal weights and normalizing, perform weighted summation to obtain the final hidden state h t , and the formula is:

[0107]

[0108] Furthermore, when the Spatio-Temporal Attention (STA) module 123 is running, it first performs temporal attention calculation: for each historical moment i, calculate its correlation score with the current moment t; then performs spatial attention calculation: for each sensor j, evaluate its contribution to the current prediction; finally, weight fusion: after multiplying the spatio-temporal weights and normalizing, perform weighted summation to obtain the final hidden state.

[0109] The multi-task prediction (MTH) module 124 in this improved gated recurrent unit generates predicted future state values of each parameter based on the features extracted by the spatio-temporal attention (STA) module 123.

[0110] Since parameters such as temperature and humidity have different physical characteristics, this multi-task prediction (MTH) module 124 independently predicts the future states of each environmental parameter. That is, for each environmental parameter, the predicted future state value is predicted through an independent fully connected layer, which can avoid interference between tasks.

[0111] Specifically, this multi-task prediction (MTH) module 124 is set to share underlying features, that is, all parameters share the underlying E-GRU unit to extract common temporal features.

[0112] Furthermore, this multi-task prediction (MTH) module 124 is also set to have independent fully connected branches, that is, each output head contains a 2-layer fully connected network, and the structure is as follows:

[0113]

[0114] where k ∈ {temperature (T), humidity (H), CO2 (C), illuminance (I)}.

[0115] For the fused feature h output by the spatio-temporal attention (STA) module 123, based on this multi-task prediction (MTH) module 124, it will be input into four independent fully connected networks for temperature, humidity, CO2, and illuminance, sharing the underlying spatio-temporal features while avoiding interference between tasks.

[0116] This multi-task prediction (MTH) module 124 is set to dynamically allocate loss weights for multiple objectives. That is, when the prediction error of a certain parameter (such as CO2) continues to be high, its loss weight will be automatically increased, forcing the model to prioritize optimizing key parameters.

[0117] The improved gated recurrent unit formed in this way can output predicted values of future environmental parameters as the state input for reinforcement learning.

[0118] See Figure 4 , in this system, the reinforcement learning policy generation module 130 is specifically composed of a state space unit 131, an action space unit 132, a reward function unit 133, and a deep Q-network (DQN) 134 working together.

[0119] Among them, the state space unit 131 fuses the predicted values of environmental parameters generated by the environmental parameter prediction module with real-time data (including real-time energy consumption and temperature change trends) to construct a corresponding feature vector, that is, a state vector.

[0120] Specifically, this state space unit 131 will use the predicted value of the next 5 minutes output by the E-GRU Combined with the current energy consumption \(E_t\) and the temperature change trend \(\Delta T_t\), a 6-dimensional state vector \(S\) is formed. t :

[0121]

[0122] Among them, \(E\) t : The current total energy consumption (kWh), which is obtained in real time through a smart meter.

[0123] \(\Delta T_t=T\) t - \(T\) t-1 : The temperature change trend.

[0124] Furthermore, this state space unit 131 can further dynamically characterize the environmental state: for example, when the prediction shows that the future CO2 concentration will exceed the standard, it will increase significantly in the state vector, triggering subsequent policy adjustments.

[0125] The action space unit 132 in this reinforcement learning policy generation module 130 is used to encode the control instructions of the device into discrete actions and construct a feasible operation set of the device control instructions.

[0126] Specifically, this action space unit 132 can perform multi-device collaborative mapping, such as encoding the control instructions of devices such as air conditioners, fresh air, and lighting into discrete actions (such as "high gear of air conditioner + full open of fresh air + 10% reduction of lighting"), and constructing a feasible operation set of the device control instructions.

[0127] Furthermore, this action space unit 132 can map the abstract decision of the policy network into actual control instructions and ensure the system security through a constraint mechanism. Thus, it can be linked with the deep Q network (DQN), and can use the discrete action Q value output by the DQN network as input to output the physical actions that the device can execute (such as \(\Delta T = -1^{\circ}C\)).

[0128] Furthermore, this action space unit 132 can be linked with the state space: when the state vector shows high energy consumption, the high-energy consumption actions (such as high gear of air conditioner) in the action space will be suppressed by the reward function.

[0129] The reward function unit 133 in this reinforcement learning policy generation module 130 can calculate the immediate reward value based on the constructed reward function according to the environmental parameters in the state space and the execution results of the action space.

[0130] Specifically, the reward function constructed by this reward function unit 133 includes: comfort reward and energy consumption penalty.

[0131] Among them, the comfort reward uses an exponential decay form to strengthen the area near the set value, and the formula is:

[0132]

[0133] Default weights: w1 = 0.4, w2 = 0.3;

[0134] Dynamic adjustment: When ∣T t - T s ∣ > 2 °C, w1 is increased to 0.6 to give priority to ensuring comfort.

[0135] The energy consumption penalty adopts a linear penalty for the current energy consumption ratio:

[0136]

[0137] Default parameters: λ E = 0.3, E max = 100 kWh (set according to the building scale).

[0138] As an example, based on this, the reward function unit realizes the dynamic trade - off between comfort and energy consumption as follows:

[0139] When the temperature deviation from the set value exceeds 2 °C, the exponential comfort reward term will dominate the reward function, prompting the strategy to give priority to cooling.

[0140] In the steady state (when the temperature is close to the set value), the linear energy consumption penalty term (-λ E ·E t ) dominates, guiding the selection of energy - saving actions.

[0141] Furthermore, when the reward function unit 133 is running, it can use the environmental parameters in the state space and the execution results in the action space as inputs, and thus calculate and generate corresponding scalar reward values (combining multiple objectives such as comfort, energy consumption, and safety), thereby quantifying the quality of the strategy and being able to provide a gradient update direction for the DQN network.

[0142] The scalar reward value calculated and generated by this reward function unit will be used for the Q - value update calculation of the DQN network, driving the network to converge to a high - return strategy.

[0143] The deep Q - network (DQN) 134 in this reinforcement learning strategy generation module 130 can calculate and generate target Q - values according to the high - dimensional feature vectors in the state space unit 131, and directly map the target Q - values to each discrete action in the action space, and select the action with the largest Q - value as the optimal action to execute.

[0144] On this basis, after the action is executed, the deep Q - network (DQN) 134 also updates and optimizes the deep Q - network parameters according to the immediate reward value calculated by the reward function unit 133 based on the environmental parameters in the current state space and the execution results of the latest action space. Among them, the reward value participates in the Q - value update through the Bellman equation.

[0145] When the Deep Q-Network (DQN) 134 is running, it can take the eigenvector of the state space, the set of actionable actions in the action space, and the feedback value of the reward function as input values, and thus generate an optimal action selection strategy. At the same time, in this process, the optimal strategy is approximated through Q-value iteration to maximize the long-term cumulative reward.

[0146] When the reinforcement learning policy generation module 130 formed based on the above scheme is running, it can perform forward propagation in the state-action-reward:

[0147] (1) The state space provides the joint features of the environment and the device;

[0148] (2) The DQN selects actions based on the state;

[0149] (3) After the action is executed, the environment feeds back a new state;

[0150] (4) The reward function calculates the immediate reward based on the old and new states and the action;

[0151] (5) The DQN updates the Q-value table using the reward value to optimize the subsequent strategy.

[0152] At the same time, it can also form a feedback loop: high-reward actions are strengthened (Q-value increases), and low-reward actions are suppressed (Q-value decays).

[0153] As an example, when a certain action causes the CO2 concentration to drop rapidly, its Q-value increases significantly, and this action is preferentially selected in subsequent similar states.

[0154] In the present system, the online adaptive parameter tuning module 140 forms effective data interaction with the multi-source data acquisition and preprocessing module 110 and the environmental parameter prediction module 120. Specifically, it realizes the dynamic update of the model parameters of the Enhanced Gated Recurrent Unit (E-GRU) in the environmental parameter prediction module 120 through KL divergence monitoring and incremental training to ensure the long-term control effect.

[0155] Among them, data distribution monitoring: Detect environmental pattern drift through KL divergence, that is, detect changes in the environmental data distribution. When D KL (P old / / P new ) > 0.1, start incremental fine-tuning. For example, when the temperature distribution in the winter heating season is significantly different from that in summer, the KL divergence will exceed the threshold (DKL > 0.1).

[0156] Incremental fine-tuning: During the fine-tuning process, freeze the underlying parameters of the E-GRU module and only update the weights of the fully connected layers of the STA and MTH units. In this way, it can retain the basic feature extraction ability while adapting to the new distribution.

[0157] Furthermore, the incremental fine-tuning implementation solution here uses the Elastic Weight Consolidation (EWC) algorithm to prevent catastrophic forgetting.

[0158] The following specifically describes the specific implementation process of the deep underground space environment comfort adaptive regulation system solution.

[0159] Step 1: Multi-source data collection and preprocessing

[0160] In this step, to ensure the quality of data collection, the signal-to-noise ratio of the input signal is improved through standardization and noise suppression.

[0161] First, data collection is carried out:

[0162] Sensor types: Temperature (DS18B20), humidity (DHT22), CO2 (MH-Z19B), illuminance (BH1750).

[0163] Sampling frequency: Collect once every 3 minutes to generate a four-dimensional vector:

[0164] x t =[T t ,H t ,C t ,I t 。

[0165] Then, abnormal data processing is carried out through the 3σ criterion:

[0166] If is determined to be an outlier, it is filled with the median of the previous 3 steps.

[0167] As an example, if the temperature data suddenly changes to 50 °C (normal range 10 - 40 °C), the replacement logic is triggered.

[0168] Next, normalization processing is carried out.

[0169] Min-Max normalization is independently carried out for each parameter, formula:

[0170]

[0171] Finally, for the processed data, it is converted into a time series window to generate time series window data.

[0172] Here, when constructing the sliding window, with 10 time steps (30 minutes of historical data), an input tensor is formed

[0173] Step 2: E-GRU dynamic feature extraction

[0174] First, the Adaptive Reset Gate (ARG):

[0175] Dynamic forgetting control: introducing input change rate Δ xt As gating factor:

[0176]

[0177] r t =σ(W r ·[h t-1 ,x t ]+b r +α·△x t );

[0178] And when Δ xt >θ (the threshold is set to the 90% quantile of the historical change rate), the old state is forced to be forgotten in response to the mutation.

[0179] Next, bidirectional feature fusion (BFF) is performed.

[0180] Window construction: Take the current moment as the center to intercept 7 steps of data (3 steps before and after), and use linear extrapolation to fill missing values.

[0181] Update gate improvement: flatten the window data and concatenate it with the hidden state:

[0182] z t =σ(W z ·[h t-1 ,Flatten(x t-3:t+3 )]).

[0183] Next, spatiotemporal attention (STA) weights are assigned.

[0184] Temporal attention calculation:

[0185] Calculate the association weight between the historical moment and the current moment. Through the learnable query matrix Q and key matrix K, the hidden state is mapped to a high-dimensional space and the similarity score is calculated:

[0186]

[0187] Spatial attention calculation:

[0188] Evaluate the importance of each sensor parameter and generate weights through the fully connected layer:

[0189]

[0190] Weight Fusion:

[0191] The time and space weights are multiplied and normalized, and the weighted sum is used to obtain the final hidden state h t , the formula is:

[0192]

[0193] Next, the multi-task prediction output.

[0194] In this step, an independent fully-connected branch is adopted, and for each environmental parameter (temperature, humidity, CO2, illuminance), the predicted future state values are predicted through independent fully-connected layers respectively:

[0195]

[0196] On this basis, in this step, the dynamic loss weight adjustment is further carried out:

[0197] The initial weights are set as λ T = 0.4, λ H = 0.3, λ C = 0.2, λ I = 0.1, and during training, they are adjusted proportionally according to the validation set error:

[0198]

[0199] In this step, the predicted future environmental parameter values can be output as the state input for reinforcement learning.

[0200] Step 3: Reinforcement learning policy generation and online optimization.

[0201] Based on the prediction results, this step generates control instructions and adapts to environmental changes through online learning.

[0202] Specifically, in this step, the predicted environmental parameter values generated by the environmental parameter prediction module are fused with the real-time device status to construct a state space.

[0203] The state vector in this constructed state space includes predicted parameters, real-time energy consumption, and temperature trend, and the dimension is 6:

[0204]

[0205] E t : Current total energy consumption (kWh), read in real time from the smart meter;

[0206] ΔT t = Tt - Tt-1: Temperature change trend.

[0207] In this step, a corresponding action space is constructed, and in this space, a feasible operation set of device control instructions is defined, and the device state combinations are encoded into a 60-dimensional discrete action space (4×3×5).

[0208] As an example, the device control instruction set in this example is as follows:

[0209] Air conditioner: Off, low gear (1 kW), medium gear (2 kW), high gear (3 kW).

[0210] Fresh air system: Off, 50% air volume, 100% air volume.

[0211] Lighting adjustment: ±20%, ±10%, maintain.

[0212] In this step, a corresponding reward function is constructed, where the comfort reward uses an exponential decay form to strengthen the area near the set value. The formula is:

[0213]

[0214] Default weights: w1 = 0.4, w2 = 0.3; Dynamic adjustment: When |T t -T s | > 2°C, w1 is increased to 0.6 to give priority to ensuring comfort.

[0215] The energy consumption penalty uses a linear penalty for the current energy consumption ratio:

[0216]

[0217] Default parameters: λ E = 0.3, E max = 100 kWh (set according to the building scale).

[0218] On this basis, DQN training and decision-making are carried out.

[0219] Experience replay: Store recent state transition tuples (st, at, rt, st+1) in the replay buffer with a capacity of 100,000 entries.

[0220] Batch sampling: Randomly sample 512 data entries every 100 steps to calculate the temporal difference error (TD Error) and update the network parameters.

[0221] Target network synchronization: Copy the main network parameters to the target network every 1000 steps to stabilize the Q-value estimation.

[0222] Then, online adaptive parameter tuning is performed.

[0223] In this step, data distribution monitoring is first carried out:

[0224] Calculate the KL divergence between the recent data distribution (Pnew) and the historical distribution (Pold) every 24 hours:

[0225]

[0226] Update trigger condition: If D KL > 0.1, it is determined that a significant change in the environmental mode has occurred.

[0227] Then, incremental training is carried out:

[0228] Data retention: Rolling retention of time series window data for the most recent 30 days.

[0229] Parameter fine-tuning: Gradient descent is performed on the attention layer and output layer of E-GRU with a learning rate of 0.0001, and the underlying GRU parameters are frozen to avoid catastrophic forgetting.

[0230] This reinforcement learning strategy generation and online optimization step can convert the prediction results into control instructions and continuously optimize the strategy through online learning to form a closed-loop control of "perception → prediction → decision-making → feedback".

[0231] For the adaptive regulation scheme of the comfort level of the deep underground space environment given in this example, it will be further illustrated through specific application examples below.

[0232] Application example 1:

[0233] Rapid cooling in summer high temperature

[0234] Initial state: The underground space temperature is 35°C, and the target setting is 26°C.

[0235] Data processing:

[0236] It is detected that the temperature suddenly rises to 35°C (exceeding 3σ of the historical average), which is determined to be abnormal and corrected with the median of the previous 3 steps, 34°C.

[0237] Construct a time series window and input it into E-GRU to predict that the temperature will drop to 32°C in the next 5 minutes.

[0238] Policy generation:

[0239] DQN selects the action {air conditioner: high gear, fresh air: 100%}, and the temperature drops to 27.5°C after 20 minutes, with an energy consumption of 45 kWh.

[0240] Application example 2:

[0241] Response to CO2 concentration exceeding the standard

[0242] Emergency situation: There is a gathering of people in the underground space, and the CO2 concentration suddenly rises from 800 ppm to 1500 ppm.

[0243] Dynamic response:

[0244] (1) ARG detects Δ xt = 300 ppm / 3 min, the reset door closes, and the historical state is forgotten.

[0245] (2) STA allocates a CO2 weight of 0.8 for spatial attention, and decides to start the full-speed mode of the fresh air system. After 10 minutes, the concentration drops to 850 ppm.

[0246] The adaptive regulation scheme for the comfort of the deep underground space environment given in this example has the following beneficial effects compared with the prior art:

[0247] Dynamic fractal drive: The ARG module can sense data mutations in real time, and the response speed is increased by 56% compared with traditional methods.

[0248] Space-time joint modeling: The STA mechanism reduces the temperature prediction error to ±0.3°C, and the CO2 prediction error ≤ 50 ppm.

[0249] Multi-objective optimization: The reward function dynamically balances comfort and energy consumption, and the system PUE (Power Usage Effectiveness) reaches 1.12, which is better than the industry standard of 1.25.

[0250] Online adaptability: KL divergence monitoring ensures that the model remains effective in scenarios such as seasonal changes and equipment aging.

[0251] In the above embodiments, the descriptions of each embodiment have their own emphases. For parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0252] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and modules described above can refer to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0253] Those skilled in the art should understand that the embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0254] The present invention is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products of the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be realized by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, so that the instructions executed by the processors of the computer or other programmable data processing devices generate means for realizing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0255] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to operate in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means that implement the functions specified in one or more of the flow Figure 1 a flow or flows and / or box Figure 1 es or boxes.

[0256] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable apparatus provide steps for implementing the functions specified in one or more of the flow Figure 1 a flow or flows and / or box Figure 1 es or boxes.

[0257] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0258] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM) and / or non-volatile memory such as read-only memory (ROM) or flash memory (flash RAM). Memory is an example of computer-readable media.

[0259] Computer-readable media includes both permanent and non-permanent, removable and non-removable media implemented by any method or technology for storing information. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile discs (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device.

[0260] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, commodity or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, commodity or device comprising the element.

[0261] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code.

[0262] The above method of the present invention, or a specific system unit, or a part of it, is a pure software architecture and can be distributed through program code on a physical medium, such as a hard disk, a CD-ROM, or any electronic device (such as a smart phone, a computer-readable storage medium). When the machine loads the program code and executes it (such as a smart phone loading and executing), the machine becomes a device for implementing the present invention. The above method and device of the present invention can also be in the form of program code and be transmitted through some transmission media, such as cables, optical fibers, or any transmission type. When the program code is received, loaded, and executed by a machine (such as a smart phone), the machine becomes a device for implementing the present invention.

[0263] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and the descriptions in the specification only illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. A deep underground space environmental comfort adaptive control system based on an improved gated circulation unit, characterized in that: The control system comprises: The multi-source data acquisition and preprocessing module collects the environmental parameters of multiple sources in the deep underground space in real time and converts them into time series windows to generate time series window data; An environmental parameter prediction module, which extracts dynamic fractal features of multi-source environmental data and generates future environmental parameter prediction values ​​through an improved gated cycle unit based on the data collected and processed by the multi-source data collection and preprocessing module; The reinforcement learning strategy generation module generates control instructions for control equipment based on the environmental parameter prediction values ​​and real-time data generated by the environmental parameter prediction module, and balances comfort and energy consumption through the reward function.

2. The deep underground space environment comfort adaptive control system according to claim 1 is characterized in that: The improved gated recurrent unit includes an adaptive reset gate (ARG) module, a bidirectional feature fusion (BFF) module, a spatiotemporal attention (STA) module and a multi-task prediction (MTH) module. The adaptive reset gate (ARG) module is used to perceive the mutation characteristics of input data in real time, dynamically adjust the forgetting strength of the reset gate, and filter abnormal data; The bidirectional feature fusion (BFF) module supplements the local window features based on the hidden state formed after filtering and purification by the adaptive reset gate (ARG) module; The spatiotemporal attention (STA) module performs spatiotemporal weight allocation on the features output by the bidirectional feature fusion (BFF) module to extract key feature information; The multi-task prediction (MTH) module generates future state prediction values ​​of each parameter based on the features extracted by the spatiotemporal attention (STA) module.

3. The deep underground space environment comfort adaptive control system according to claim 2 is characterized in that: The adaptive reset gate (ARG) module is configured to reset when the input data change rate exceeds the historical threshold θ Δx When , the gate weight r will be reset t Set to 0 to force forgetting the historical hidden state h t-1 .

4. The deep underground space environment comfort adaptive control system according to claim 2 is characterized in that: The bidirectional feature fusion (BFF) module is configured to intercept the window data of 7 steps before and after the current moment, flatten the data in the window and compare it with the hidden state h′ output by the adaptive reset gate (ARG) module t Splicing; and input the spliced ​​feature vector into the update gate z t , calculated as: With t =σ(W z ·[h′ t ,BFF(x t-3:t+3 )]+b z ), Among them, BFF( xt-3:t+3 ) is the feature after the window is flattened.

5. The deep underground space environment comfort adaptive control system according to claim 2 is characterized in that: The spatiotemporal attention (STA) module includes: Time attention weight calculation: Calculate the association weight between the historical moment and the current moment, map the hidden state to a high-dimensional space through the learnable query matrix Q and key matrix K, and calculate the similarity score: Spatial attention weight calculation: Evaluate the importance of each sensor parameter and generate weights through the fully connected layer: Weight fusion: multiply the temporal and spatial weights and normalize them, and then sum them to get the final hidden state h t , the formula is:

6. The deep underground space environment comfort adaptive control system according to claim 2 is characterized in that: The multi-task prediction (MTH) module predicts the future state prediction value for each environmental parameter through an independent fully connected layer.

7. The deep underground space environment comfort adaptive control system according to claim 1 is characterized in that: The reinforcement learning strategy generation module includes The state space unit fuses the environmental parameter prediction values ​​generated by the environmental parameter prediction module with the real-time data to construct the corresponding feature vector; The action space unit encodes the control instructions of the device into discrete actions and constructs a feasible operation set of the device control instructions; The reward function unit calculates the immediate reward value based on the constructed reward function according to the environmental parameters in the state space and the execution results in the action space; A deep Q network (DQN) generates a target Q value based on a high-dimensional feature vector calculation in a state space unit, and directly maps the target Q value to each discrete action in an action space, and selects the action with the largest Q value as the optimal action to execute; after the action is executed, the deep Q network (DQN) updates and optimizes the deep Q network parameters according to the immediate reward value calculated by the reward function unit based on the environmental parameters in the current state space and the execution results of the latest action space.

8. A deep underground space environment control method based on improved GRU, characterized in that: include: The improved gated recurrent unit is used to extract the dynamic fractal characteristics of multi-source environmental data and generate the predicted values ​​of future environmental parameters. The deep Q network (DQN) generates control instructions for the control equipment based on the predicted values ​​of environmental parameters and real-time data (including real-time energy consumption and temperature change trends), and balances comfort and energy consumption through the reward function.

9. The deep underground space environment control method based on improved GRU according to claim 8 is characterized in that: The process of generating the predicted value of future environmental parameters in the deep underground space environmental control method includes: The adaptive reset gate (ARG) module senses the mutation characteristics of the input data in real time, dynamically adjusts the forgetting strength of the reset gate, and filters abnormal data; The Bidirectional Feature Fusion (BFF) module supplements the local window features based on the hidden state formed after filtering and purification by the Adaptive Reset Gate (ARG) module; The spatiotemporal attention (STA) module allocates spatiotemporal weights to the features output by the bidirectional feature fusion (BFF) module to extract key feature information; The multi-task prediction module (MTH) generates the future state prediction value of each parameter based on the features extracted by the spatiotemporal attention (STA) module.

10. The deep underground space environment control method based on improved GRU according to claim 8, characterized in that: The process of generating control instructions for control equipment in the deep underground space environment control method includes: In the state space, the environmental parameter prediction values ​​generated by the environmental parameter prediction module are fused with the real-time data to construct the corresponding feature vector; The deep Q network (DQN) generates the target Q value based on the feature vector in the state space, and directly maps the target Q value to each discrete action in the action space, and selects the action with the largest Q value as the optimal action to execute; After the action is executed, the deep Q network (DQN) updates and optimizes the deep Q network parameters according to the immediate reward value calculated by the reward function based on the environmental parameters in the current state space and the execution results of the latest action space.

Images