Urban rail electromechanical energy-saving predictive control method and system based on adaptive model

Abstract

This invention discloses a predictive control method and system for energy conservation in urban rail transit electromechanical systems based on an adaptive model, belonging to the field of energy conservation control technology for urban rail transit. Existing control methods for ventilation and air conditioning systems in subway stations struggle to balance environmental comfort and operational energy consumption. Online simulation is time-consuming, online strategy trial-and-error carries significant risks, and real-time adaptive adjustment is difficult. This method deploys an offline-trained neural network proxy model and a control strategy model on edge computing devices in subway stations. The neural network proxy model outputs predicted temperature field distribution data and equipment energy consumption data for multiple areas of the station based on outdoor environmental parameters, passenger flow parameters, and equipment operating parameters. The control strategy model uses this proxy model as an interactive environment to construct a reward function based on thermal comfort and equipment energy consumption indicators for reinforcement learning pre-training. This invention is primarily used for energy-saving optimization control of ventilation and air conditioning systems in subway stations.

Description

Technical Field

[0001] This invention belongs to the field of energy-saving control technology for urban rail transit. More specifically, this invention relates to an energy-saving predictive control method and system for urban rail electromechanical systems based on an adaptive model. Background Technology

[0002] The energy consumption of ventilation and air conditioning systems in subway stations accounts for a significant portion of total operating energy consumption. The system's operation is influenced by multiple factors, including outdoor weather, passenger flow variations, train piston wind, and equipment thermal inertia, resulting in time-varying and nonlinear characteristics. Maintaining thermal comfort in all areas of the station while reducing energy consumption is a problem that needs to be solved in practical operation.

[0003] Currently, common control methods mainly include proportional-integral-derivative (PID) control and rule-based control based on preset thresholds. PID parameters are typically set during the commissioning phase and remain unchanged, resulting in lag in response to significant fluctuations in passenger flow or outdoor weather, making it difficult to coordinate the operation of multiple subsystems. Rule-based control relies on manually set logical conditions, making it unable to respond separately to temperature differences in different areas within the station, nor can it make adjustments in advance based on future load trends.

[0004] If computational fluid dynamics models are used to perform online simulations of the flow and temperature fields within the station to support control decisions, the time consumed in a single calculation is relatively long, which falls short of the response speed required by the control cycle. If a simplified lumped parameter model is used, its ability to describe the non-uniformity of temperature distribution and the impact of airflow organization within the station is limited.

[0005] Training reinforcement learning control strategies requires extensive interaction with the environment. If online trial and error is conducted in a real operational station, inappropriate control commands during this process may cause short-term deviations in thermal comfort or frequent equipment adjustments, adversely affecting operational stability and equipment lifespan. If a simplified simulation environment is used for offline training, the discrepancy between the simulation environment and the actual system's thermal response will cause the control strategy's performance to degrade after actual deployment.

[0006] Sensor data acquisition relies on existing networks, and communication protocols and data formats differ across stations. Combining real-time acquired data with historical parameters to form a usable state representation and mapping control outputs to the equipment's executable range requires specific adaptation. When the predictive model deviates due to equipment aging or changes in operating conditions, there is a lack of clear criteria for switching to backup control modes. The requirements for control accuracy and adjustment frequency differ between operating and non-operating periods; a fixed control cycle and a fixed model accuracy threshold are insufficient to meet the actual needs of both periods. Summary of the Invention

[0007] To achieve these objectives and other advantages of the present invention, according to one aspect of the present invention, an energy-saving predictive control method for urban rail transit electromechanical systems based on an adaptive model is provided, comprising the following steps: The trained neural network proxy model and control strategy model are deployed on the edge computing equipment of the subway station. The neural network proxy model is used to output temperature field distribution prediction data and equipment energy consumption prediction data of multiple areas inside the station based on the input outdoor environmental parameters, passenger flow parameters and equipment operation parameters. Within each control cycle, the edge computing device acquires the current outdoor environmental parameters and current passenger flow parameters collected by the station sensor network, and inputs the current outdoor environmental parameters and current passenger flow parameters into the control strategy model; the control strategy model is obtained through reinforcement learning pre-training based on the temperature field distribution prediction data and equipment energy consumption prediction data provided by the neural network proxy model as the interaction environment, and outputs the equipment operation parameter instructions for the current control cycle according to the input parameters; The device operating parameter command is sent to the controller, which then controls the chiller, chilled water pump, cooling water pump, cooling tower, and air conditioning fan to execute the corresponding operating frequency or valve opening. During system operation, the residual between the predicted temperature distribution data inside the station predicted by the neural network surrogate model and the actual temperature distribution data inside the station collected by the sensors is continuously monitored. When the residual exceeds a preset residual threshold, the edge computing device automatically switches to the proportional-integral-derivative (PID) standby control mode in the next control cycle. The preset residual threshold is set with different values ​​depending on whether the station is in operation or not. In the PID standby control mode, the edge computing device calculates the control output based on the deviation between the set temperature and the actual temperature in each area and maps it to the device operating parameter command. At the same time, it generates a model update trigger signal to indicate that the neural network surrogate model needs to be retrained.

[0008] Preferably, the training steps of the neural network proxy model include: Based on the computational fluid dynamics simulation model of the ventilation and air conditioning system of the subway station, different values ​​of outdoor environmental parameters in the range of -10℃ to 45℃, passenger flow parameters in the range of 0 people / hour to the maximum design passenger flow of the station, and equipment operating parameters in the range of 20% to 100% of the rated operating range of each piece of equipment are selected and combined to form no less than 1,000 sets of boundary conditions. Each set of boundary conditions is input into the computational fluid dynamics simulation model for offline simulation. Simulation data of temperature field distribution in multiple areas inside the station corresponding to each set of boundary conditions, as well as simulation data of equipment energy consumption of chiller units, chilled water pumps, cooling water pumps, cooling towers and air conditioning fans in the ventilation and air conditioning system, are obtained. The simulation dataset is composed of each set of boundary conditions and their corresponding simulation data. Construct a deep neural network consisting of an input layer, three fully connected hidden layers, and an output layer. The number of neurons in the input layer is the same as the total number of outdoor environmental parameters, passenger flow parameters, and equipment operation parameters. The number of neurons in the output layer is the same as the sum of the number of measurement points for temperature field distribution data in the multiple regions and the number of devices for equipment energy consumption data. The deep neural network is trained using the simulation dataset. During the training process, outdoor environmental parameters, passenger flow parameters, and equipment operating parameters are used as network inputs, and temperature field distribution simulation data of multiple areas inside the station and equipment energy consumption simulation data are used as supervision signals for the network output. The network weight parameters are adjusted through the backpropagation algorithm until the mean square error between the network output and the supervision signal is less than 0.001 or the number of training iterations reaches 10,000.

[0009] Preferably, the specific implementation steps of the neural network surrogate model to predict the temperature distribution prediction data inside the station include: The edge computing device combines the current outdoor ambient temperature T1, the current outdoor relative humidity H1, the current passenger flow P, the chiller outlet water temperature setpoint T3 for the current control cycle, the chilled water pump frequency f5, the cooling water pump frequency f6, the cooling tower fan frequency f7, and the air conditioning fan frequency f8 to form an input vector X = [T1, H1, P, T3, f5, f6, f7, f8]. The edge computing device inputs the input vector X into a locally deployed neural network proxy model. The neural network proxy model consists of an input layer, three fully connected hidden layers, and an output layer. The input layer has 8 neurons, the first hidden layer has 64 to 256 neurons, the second hidden layer has 64 to 256 neurons, the third hidden layer has 64 to 256 neurons, and the output layer has N neurons, where N is the total number of areas inside the station. Each hidden layer uses the ReLU activation function. The neural network proxy model performs forward propagation calculation on the input vector X. The output of the first hidden layer is G1 = ReLU(W1·X + b1), the output of the second hidden layer is G2 = ReLU(W2·G1 + b2), the output of the third hidden layer is G3 = ReLU(W3·G2 + b3), and the output of the output layer is Y = W4·G3 + b4. Here, W1, W2, W3, and W4 are weight matrices, and b1, b2, b3, and b4 are bias vectors. The weight matrices and bias vectors are parameter values ​​determined by the backpropagation algorithm during the offline training stage and stored in the edge computing device. The values ​​of each element in the output vector Y represent the predicted temperatures of zones 1 to N within the station. The edge computing device resolves the output vector Y into T. p1 T p2Until T pN This serves as predictive data for the temperature field distribution in multiple areas within the station.

[0010] Preferably, the training steps of the control strategy model include: The trained neural network agent model is used as the interaction environment of the reinforcement learning agent. The neural network agent model receives the equipment operation parameters output by the reinforcement learning agent and returns the corresponding temperature field distribution prediction data and equipment energy consumption prediction data for multiple areas inside the station. Define a state space S, an action space A, and a reward function R for reinforcement learning. The state space S consists of T1, H1, P, the chiller outlet water temperature setpoint T2 from the previous control cycle, the chilled water pump frequency f1 from the previous control cycle, the cooling water pump frequency f2 from the previous control cycle, the cooling tower fan frequency f3 from the previous control cycle, and the air conditioning fan frequency f4 from the previous control cycle. The action space A consists of T3, f5, f6, f7, and f8 from the current control cycle. The reward function R is set as follows: ; Among them, T i Let T4 be the predicted temperature for the i-th region, and P be the set temperature. j Let be the real-time power of the j-th device, α be the thermal comfort weighting coefficient with a value ranging from 0.1 to 10, β be the energy consumption weighting coefficient with a value ranging from 0.01 to 1, N be the total number of areas inside the station, and M be the total number of devices. The reinforcement learning agent is trained offline using a deep deterministic policy gradient algorithm. During training, the agent outputs device operation parameters in action space A based on the current state space S. The neural network agent model receives these parameters and returns the next state space S' and reward function value R. The agent stores the state space S, action space A, reward function value R, and next state space S' generated from each interaction as an experience sample in an experience replay buffer. Every K training steps, B experience samples are randomly sampled from the experience replay buffer to pair the policy network and value... The network is updated using gradient descent, where K ranges from 50 to 200, B ranges from 32 to 512, the learning rate η1 of the policy network ranges from 0.0001 to 0.001, the learning rate η2 of the value network ranges from 0.001 to 0.01, and the discount factor γ ranges from 0.95 to 0.99. Training continues until the standard deviation of the cumulative reward function value over L consecutive training rounds is less than δ, where L ranges from 50 to 200 and δ ranges from 0.01 to 0.1. After training, the control policy model is obtained.

[0011] The values ​​of the hyperparameters mentioned above, such as the empirical replay sampling interval K, batch size B, learning rates η1 and η2, and discount factor γ, are not arbitrarily chosen. Instead, they are optimal engineering ranges determined based on the convergence theory of deep deterministic policy gradient algorithms in continuous control tasks, combined with the computational resource constraints of offline training for subway station ventilation and air conditioning systems and the actual control accuracy requirements. The specific settings for each parameter are as follows: 1. Basis for setting the empirical playback sampling interval K (50 to 200 steps): The value of K determines the frequency of updates to the policy network and the value network. In the DDPG algorithm, if the value of K is too small (e.g., updates every step), the temporal correlation between experience samples will be too strong, the policy network will easily get trapped in local optima, and frequent updates will significantly increase the computational cost of offline training. If the value of K is too large (e.g., more than 500 steps), the policy update will lag, and the convergence speed will drop significantly. In this invention, the length of a single round of training of the policy model is controlled to be 1440 steps (simulating 24 hours of operation), and K is set to 50 to 200 steps, which is equivalent to about 7 to 29 network updates in each round. This frequency ensures that the policy can learn from newly collected experience in a timely manner, and also matches the offline training computing power (usually a single GPU industrial control computer) that edge computing devices can provide in actual deployment, ensuring that the training task can converge within a reasonable time (usually 24 to 48 hours).

[0012] 2. The basis for setting the batch size B (32 to 512): The selection of batch size B needs to balance the accuracy of gradient estimation and memory usage. A lower limit of 32 is the minimum sample size to ensure the gradient descent direction is statistically significant; below this value, the gradient estimation variance is too large, leading to severe oscillations during training. An upper limit of 512 is limited by the hardware constraints of the edge computing device used in this solution (configured with 16GB of memory and a GPU); an excessively large batch size can cause memory overflow during a single forward propagation. Furthermore, in the DDPG algorithm, the batch size is usually set to several times the dimension of the state space. In this invention, the state space dimension is 8 dimensions, and B values ​​from 32 to 512 cover the conventional recommended range of 4 to 64 times. In actual training, 128 can be selected as a balance point, at which a relatively good balance is achieved between gradient estimation variance and computational efficiency.

[0013] 3. Basis for setting the learning rate η1 (0.0001 to 0.001) for the policy network and the learning rate η2 (0.001 to 0.01) for the value network: In the reinforcement learning Actor-Critic architecture, the value network is responsible for fitting the Q-function, and its output is the scalar reward expectation. Convergence is relatively easy, and a higher learning rate η2 (0.001 to 0.01) can be used to accelerate the approximation of the value function. The policy network, on the other hand, directly outputs continuous action vectors. If the learning rate η1 is too large, the policy will fluctuate drastically, leading to frequent changes in the operating frequency of the equipment in actual control, shortening the lifespan of the actuator. Therefore, η1 is taken in a lower range (0.0001 to 0.001). The ratio η1:η2 is approximately between 1:10 and 1:100, which conforms to the recommended settings in the DDPG algorithm principle paper. Simulations have verified that within this range, the training process can stably converge to the termination condition where the cumulative reward standard deviation is less than 0.1.

[0014] 4. Basis for setting the discount factor γ (0.95 to 0.99): The discount factor γ represents the importance of future rewards to current decisions. In the ventilation and air conditioning control of subway stations, equipment operation exhibits significant thermal inertia, and the effects of current control commands will continue to manifest for several minutes to tens of minutes. The closer γ is to 1, the more the strategy tends to consider the comprehensive optimization of long-term energy consumption and comfort. However, if γ equals 1, the value function may diverge in an infinite time domain task. This invention limits γ to between 0.95 and 0.99, ensuring both effective discount weighting of the strategy for the cumulative rewards over the next 15 to 100 steps (corresponding to approximately 15 to 100 minutes) and avoiding the risk of numerical overflow. Offline simulation comparisons show that when γ is 0.98, the strategy performs best in peak load forecasting and pre-adjustment.

[0015] 5. Basis for setting convergence criteria L (50 to 200 rounds) and δ (0.01 to 0.1): L represents the number of consecutive rounds used to calculate the standard deviation of the cumulative reward. If L is too small (e.g., 10 rounds), the convergence determination is easily affected by the randomness of training, leading to misjudgments; if L is too large (e.g., 500 rounds), it will unnecessarily prolong the training time. A L of 50 to 200 rounds corresponds to approximately 72,000 to 288,000 steps of interaction data, which is sufficient to smooth out short-term fluctuations. δ is the convergence threshold. Since the temperature deviation term in the reward function is in °C and the energy consumption term is in kW, the normalized cumulative reward value is usually between -100 and 0. A δ of 0.01 to 0.1 corresponds to a relative fluctuation of the cumulative reward on the order of one-thousandth to one-hundredth. At this point, there is no significant room for improvement in policy performance, and the marginal benefit of continuing training is extremely low.

[0016] The ranges of the aforementioned hyperparameters are all optimized intervals selected for the specific application scenario of urban rail stations, based on extensive offline simulations and hardware-in-the-loop testing, ensuring reliable convergence and efficient execution of the algorithm. In practical implementation, those skilled in the art only need to perform a limited number of routine adjustments within this range to obtain the optimal parameter combination suitable for a specific station; these fine-tunings do not depart from the scope of protection of this invention.

[0017] Preferably, the specific steps for the edge computing device to acquire the current outdoor environmental parameters and current passenger flow parameters collected by the station sensor network include: The edge computing device establishes a data communication connection with the station sensor network through an industrial Ethernet interface. At the beginning of each control cycle, the edge computing device sends a data request command to the station sensor network. In response to the data request command, the station sensor network encapsulates the current T1 and H1 collected by the outdoor temperature and humidity sensor and the current passenger flow P collected by the passenger flow counting sensor into data frames through the Modbus TCP protocol and sends them to the edge computing device. The edge computing device parses the received data frame, extracts T1, H1, and P, and reads T2, f1, f2, f3, and f4 recorded after the execution of the previous control cycle from the local cache of the edge computing device. The above parameters are combined to form the state vector S = [T1, H1, P, T2, f1, f2, f3, f4] of the current control cycle.

[0018] Preferably, the specific implementation steps for the control strategy model to output the device operating parameter instructions for the current control cycle include: The edge computing device inputs the state vector S into a control strategy model deployed locally. The control strategy model is composed of a strategy network. The strategy network performs forward propagation calculation on the input state vector S and outputs the action vector A = [T3, f5, f6, f7, f8] for the current control cycle. The edge computing device performs amplitude limiting processing on the parameter values ​​in the action vector A, limiting T3 to the range of 7°C to 12°C, and limiting f5, f6, f7 and f8 to the range of 20Hz to 50Hz respectively. The amplitude-limited action vector A' is used as the device operation parameter instruction in the current control cycle.

[0019] The setpoint T3 for the chiller outlet water temperature and the frequency limits for each pump and fan are not arbitrarily selected, but rather optimized engineering ranges determined comprehensively based on the operating manuals of existing ventilation and air conditioning equipment in subway stations, the energy efficiency characteristic curves of the chiller units, and the safe operating boundaries of the frequency converters. The specific settings are based on the following: 1. Basis for setting the chiller outlet water temperature setpoint T3 range (7℃ to 12℃): The lower limit of 7℃ is primarily based on the anti-freeze protection and energy efficiency degradation boundaries of chiller units. Centrifugal or screw chiller units used in subway stations face the risk of freezing if the evaporator outlet water temperature falls below 5℃, triggering a protective shutdown. Simultaneously, at a typical outlet water temperature of 7℃, the chiller unit's coefficient of performance (COP) is usually in a high range. Further lowering the outlet water temperature would cause a rapid, non-linear increase in compressor power consumption, while the increase in cooling capacity would be limited, leading to a significant decrease in the system's overall energy efficiency ratio. The upper limit of 12℃ is mainly based on the dehumidification capacity and thermal comfort requirements of the terminal air conditioning units. When the chilled water outlet temperature exceeds 12℃, the surface temperature of the cooler is insufficient to cool the flowing air below the dew point temperature, resulting in excessive relative humidity in the station's public areas and causing passengers to feel stuffy. Furthermore, 7℃ to 12℃ is the recommended range for conventional chilled water supply for air conditioning in the "Subway Design Code" and the "Design Code for Heating, Ventilation and Air Conditioning of Civil Buildings," conforming to industry-standard design practices. This invention constrains the original values ​​output by the policy network to this range after linear mapping, which not only ensures the safe operation of the device, but also ensures that the reinforcement learning exploration process is always in the efficient cooling range.

[0020] 2. Basis for setting the frequency limit range (20Hz to 50Hz) for chilled water pumps, cooling water pumps, cooling tower fans, and air conditioning fans: The lower limit of 20Hz is primarily based on the low-frequency operating characteristics of the inverter and the motor. When the inverter output frequency is below 20Hz, the slip of the asynchronous motor increases significantly, the proportion of harmonic components in the winding current rises, leading to increased motor heating and accelerated insulation aging. Simultaneously, the pump head drops sharply at excessively low speeds, potentially failing to overcome pipeline resistance and causing water shortages or cavitation at the system's end. Furthermore, most general-purpose inverters are factory-set to have a minimum continuous operating frequency of 20Hz.

[0021] The upper limit of 50Hz is determined based on my country's power grid frequency standards and equipment rated parameters. Subway station water pumps and fan motors are typically designed with a rated frequency of 50Hz, at which the motor outputs its rated power and speed. If operation exceeds 50Hz, the motor will enter a constant power, weak magnetic field region, resulting in decreased torque, and the mechanical stress on the bearings and impellers will exceed design margins, posing a risk of rotor rubbing or blade breakage. Strictly limiting the frequency to between 20Hz and 50Hz ensures that any action commands output by the reinforcement learning agent during training and deployment are within the equipment's safe operating range, fundamentally preventing equipment damage or shutdown accidents caused by outlier values ​​output by the algorithm.

[0022] 3. The role of amplitude limiting in reinforcement learning training: During the offline training phase of the control strategy model, the output layer of the policy network uses the Sigmoid activation function to normalize the action values ​​to an interval, and then maps them to the aforementioned limiting range through a linear transformation. This design ensures that the exploration noise of the reinforcement learning agent is always superimposed on the safe action base value, rather than directly outputting unbounded action values ​​and then being forcibly truncated by the environment. This avoids gradient estimation bias caused by action truncation and ensures accurate fitting of the Q function by the value network in the DDPG algorithm. Extensive offline simulations have verified that the policy trained within the aforementioned limiting range outputs action values ​​concentrated in the energy-optimal region (T3 approximately 8°C to 10°C, frequency approximately 35Hz to 45Hz), almost never touching the limiting boundary. This indicates that the range setting effectively covers the optimal operating conditions without imposing unnecessary constraints on policy learning. Preferably, the duration of the control cycle is dynamically adjusted according to the station's operating status. The dynamic adjustment steps include: At the end of each control cycle, the edge computing device calculates the current passenger flow change rate ΔP and the station average temperature deviation rate ΔT. The passenger flow change rate ΔP is calculated using the formula ΔP=|P c -P p | / P max Calculate, where P c For the current passenger flow, P p For the passenger flow in the previous control period, P max The maximum designed passenger flow for the station; The average temperature deviation rate ΔT of the station is calculated according to the formula. Calculate, where T ri T4 is the actual temperature of the i-th area, N is the set temperature, and N is the total number of areas inside the station. The value of i is any integer from 1 to N. The edge computing device compares the passenger flow change rate ΔP with a first preset threshold ε1, and compares the station average temperature deviation rate ΔT with a second preset threshold ε2. The first preset threshold ε1 ranges from 0.03 to 0.10, and the second preset threshold ε2 ranges from 0.02 to 0.08. When the passenger flow change rate ΔP exceeds the first preset threshold ε1 or the station average temperature deviation rate ΔT exceeds the second preset threshold ε2, the edge computing device determines that the station is in a state of rapid heat load change, and adjusts the control cycle duration T in the next control cycle. c Adjust to the smaller value within the first value range, which is 10 seconds to 60 seconds; When the passenger flow change rate ΔP does not exceed the first preset threshold ε1 and the station average temperature deviation rate ΔT does not exceed the second preset threshold ε2, the edge computing device obtains the current system time and determines the station operation status. If the station is in operation, the edge computing device will adjust the control cycle duration T according to the first linear formula. c Adjustments are made within a first value range, and the first linear formula is T. c =T a +(1-ΔP / ε1)×(T b -T a ), where T a T is the lower limit of the first value range. b This is the upper limit of the first value range; If the station is in non-operational hours, the edge computing device will adjust the control cycle duration T according to the second linear formula. c The adjustment is made within a second value range, which is from 120 seconds to 300 seconds, and the second linear formula is T. c =T f -ΔP / ε1×(T f -T e ), where T e T is the lower limit of the second value range. f This is the upper limit of the second range of values.

[0023] The value ranges of the first preset threshold ε1 (0.03 to 0.10) and the second preset threshold ε2 (0.02 to 0.08), as well as the method of smoothing the control cycle based on a linear formula, are all optimized ranges determined by comprehensively considering the statistical laws of passenger flow fluctuations in subway stations and engineering experience in thermal comfort. The specific settings are based on the following: 1. Basis for setting the first preset threshold ε1 (passenger flow change rate threshold): The passenger flow change rate ΔP is defined as the ratio of the absolute difference between the current passenger flow and the passenger flow of the previous cycle to the station's maximum designed passenger flow. Based on the analysis of passenger flow monitoring data from typical stations on multiple subway lines in China, during off-peak operating hours, the fluctuation range of passenger flow between two adjacent control cycles (usually 30 to 60 seconds) generally does not exceed 3% of the maximum passenger flow; however, in scenarios such as the moment a train arrives at the station or the dispersal of a large event, the sudden change in passenger flow between adjacent cycles can reach 8% to 15% of the maximum passenger flow.

[0024] This invention sets the lower limit of ε1 to 0.03 (i.e., 3%). Below this value, the impact of passenger flow changes on the station's heat load can be almost smoothly absorbed by the building's thermal inertia, without needing to shorten the control cycle. The upper limit of ε1 is set to 0.10 (i.e., 10%), corresponding to more significant passenger flow surge events. If ε1 is too high (e.g., exceeding 0.15), the system will be sluggish in responding to moderate passenger flow changes, leading to perceptible overshoot in the regional temperature. If ε1 is too low (e.g., 0.01), the system will frequently trigger the shortest control cycle, causing repeated equipment adjustments. Field testing has verified that setting ε1 in the range of 0.03 to 0.10 is the optimal range for balancing response speed and equipment stability, as it can both accelerate the control response in time when passenger flow changes significantly and avoid frequent jittering of the control cycle due to minor passenger flow disturbances.

[0025] 2. The basis for setting the second preset threshold ε2 (average temperature deviation rate threshold): The average temperature deviation rate ΔT of a station is defined as the absolute value of the difference between the actual temperature and the set temperature in each area, divided by the average of the set temperatures. The set temperature T4 is typically 26℃. According to the "Metro Design Code" and the "Indoor Air Quality Standard," the allowable temperature fluctuation range in the public areas of metro stations is ±1℃ to ±2℃. When ΔT exceeds 0.08 (i.e., an average deviation of approximately 2.08℃), the deviation in some areas has exceeded the upper limit of comfort, and the control cycle must be shortened immediately to accelerate the correction. When ΔT is below 0.02 (i.e., an average deviation of approximately 0.52℃), the system is in a well-regulated state and no additional acceleration is required.

[0026] This invention limits the value of ε2 to 0.02 to 0.08, directly corresponding to the aforementioned comfort boundary. The lower limit of 0.02 is slightly higher than the sensor measurement noise level (typically ±0.3℃ corresponds to a normalized deviation rate of approximately 0.01), which can avoid the rapid adjustment mode being falsely triggered by measurement noise; the upper limit of 0.08 is slightly lower than the critical value, reserving a correction buffer time for the system. This range setting allows the temperature deviation rate index to serve as a reliable substitute signal for changes in heat load, ensuring the robustness of control decisions even when the sensor network fails or the passenger flow counting error is large.

[0027] 3. Basis for setting up the first and second linear formulas: This invention employs a linear formula to dynamically adjust the control cycle under stable operating conditions, rather than a simple threshold-based switching method. The main basis for this is as follows: First linear formula (operating period): The formula is in the form T. c =T a +(1-ΔP / ε1)×(T b -T aThis formula establishes a negative linear correlation between the control cycle duration and the passenger flow change rate ΔP. Its technical significance lies in the following: when ΔP approaches 0 (extremely stable passenger flow), the control cycle automatically extends to its upper limit of 60 seconds to minimize equipment adjustment frequency and reduce wear on the frequency converter and valves; as ΔP gradually increases to near ε1, the control cycle linearly shortens to its lower limit of 10 seconds, achieving a stepless smooth transition in control frequency. This design avoids the step jump in equipment frequency commands caused by abrupt changes in the control cycle near the critical point in traditional threshold switching methods, thereby suppressing system oscillations and overshoot.

[0028] Second linear formula (non-operational periods): The formula is in the form T. c =T f -ΔP / ε1×(T f -T e During non-operational periods, passenger flow is extremely low. While occasional passenger flows (such as night shift cleaning and inspection personnel) may cause ΔP to reach the threshold numerically, their absolute heat load increment is far lower than during operational periods. Therefore, the second linear formula adopts a gentler negative correlation relationship. That is, when ΔP changes from 0 to ε1, the control cycle only shortens from 300 seconds to 120 seconds, instead of decreasing to 10 seconds as during operational periods. This differentiated adjustment rate fully considers the low thermal inertia characteristics of non-operational periods. While ensuring that occasional loads can be responded to, it maximizes the extension of the single cycle length, which is conducive to the full utilization of nighttime natural ventilation cooling strategies and energy-saving optimization of intermittent equipment start-stop.

[0029] Preferably, the preset residual threshold E th During station operating hours, the value ranges from N×0.5℃ to N×1.0℃, and during non-operating hours, the value ranges from N×1.0℃ to N×2.0℃, where N is the total number of areas within the station.

[0030] The total residual E defined in this invention is the sum of the absolute values ​​of the differences between the predicted temperature of each region by the neural network surrogate model and the actual measured temperature of the corresponding region by the sensor, i.e., E = ∑|T pi -T ri In a statistical sense, if the prediction errors of each region are independent and the average absolute error of a single region is δ, then the expected value of the cumulative residual of N regions is approximately N×δ. Therefore, setting the threshold to a multiple of N essentially normalizes the switching criterion to the more universal engineering indicator of "tolerance for average prediction deviation in a single region".

[0031] Another advantage of using the N-multiplication form is the algorithm's portability to stations of different sizes. For example, when a station is divided into 20 temperature measurement zones, the operating time threshold is N×0.8℃, which is 16℃. If another station is divided into 35 zones due to its complex building structure, the same multiplication factor can be used to automatically adapt the corresponding threshold of 28℃. There is no need to calibrate the absolute temperature threshold separately for each station, which significantly reduces the workload of parameter adjustment in engineering deployment.

[0032] During operating hours, a large number of passengers reside in the public areas of subway stations, making thermal comfort a primary control objective. According to the "Metro Design Code" and the "Indoor Air Quality Standard," the permissible temperature fluctuation range in the public areas of subway stations is typically ±1℃ to ±2℃ of the set temperature. This invention sets the lower limit of the average prediction deviation tolerance for a single area at 0.5℃, corresponding to a deviation band of approximately ±0.5℃, slightly stricter than the standard requirements, aiming to provide a safety margin for the pre-adjustment actions of the control strategy; the upper limit of the tolerance is set at 1.0℃, corresponding to a deviation band of approximately ±1.0℃, which is within the comfort boundary that passengers can perceive but find acceptable.

[0033] If the tolerance coefficient is below 0.5℃ (e.g., 0.3℃), the prediction accuracy requirements for the surrogate model are too stringent. In actual operation, due to unmodelable factors such as sensor noise and random fluctuations in passenger flow, the residuals are likely to frequently exceed the threshold, causing the system to repeatedly switch between reinforcement learning control and PID backup mode, disrupting control continuity and increasing equipment wear. If the tolerance coefficient is above 1.0℃ (e.g., 1.5℃), when the surrogate model has already shown significant deviation, the system switching will lag, and passengers will experience noticeable temperature discomfort. Therefore, 0.5℃ to 1.0℃ is the coefficient range determined through extensive field testing, achieving the optimal balance between prediction accuracy tolerance and comfort assurance.

[0034] During non-operational periods (usually nighttime shutdowns), stations have few or no passengers, and only a small number of inspection and cleaning personnel. Thermal comfort requirements are significantly reduced, and the control objective shifts from precise temperature control to basic environmental protection and energy conservation. Relaxing the tolerance for average prediction deviation in a single area to 1.0℃ to 2.0℃ means the system has a higher tolerance for prediction deviations from the surrogate model.

[0035] This relaxed design has dual engineering value: First, during non-operational periods, outdoor temperatures are low and passenger flow is zero, so ventilation and air conditioning systems often operate at low loads or intermittently. The surrogate model's prediction accuracy naturally decreases under low-load conditions with sparse training data. If the strict thresholds for operational periods are still applied, the PID backup mode will be frequently triggered erroneously, preventing the energy-saving advantages of the reinforcement learning strategy from being realized. Second, equipment adjustments during nighttime have no impact on passenger experience. Appropriately relaxing the thresholds avoids unnecessary control mode switching caused by minor prediction errors, ensuring stable system operation under the energy-saving strategy. The lower limit of the coefficient at 1.0℃ aligns with the upper limit during operational periods, while the upper limit of 2.0℃ corresponds to a single-area deviation of approximately ±2.0℃, which is close to the upper limit of permissible temperature drift in equipment management rooms and public areas at night (typically ±3℃). Exceeding this value poses a risk of equipment freezing or condensation; therefore, 2.0℃ constitutes the upper limit of the safety boundary.

[0036] Preferably, the residual between the predicted temperature distribution data inside the station predicted by the neural network surrogate model and the actual temperature distribution data inside the station collected by the sensor is denoted as E. When E≤E th In the next control cycle, the edge computing device continues to use the control strategy model to output device operating parameter instructions; When E>E th At this time, the edge computing device automatically switches the control mode from control strategy model driven to proportional-integral-derivative (PID) standby control mode in the next control cycle. In PID standby control mode, the edge computing device determines the control mode based on the set temperature T4 of the i-th region and the actual temperature T of the i-th region. ri Deviation E between i =T4-T ri According to the formula The control output U is calculated and mapped to the chiller outlet water temperature setpoint, chilled water pump frequency, cooling water pump frequency, cooling tower fan frequency, and air conditioning fan frequency before being sent to the controller. The proportional coefficient K is used in this calculation. p The value range is from 1 to 10, and the integral coefficient K i The value ranges from 0.01 to 0.5, and the differential coefficient K d The value range is from 0 to 2; The edge computing device generates a model retraining trigger signal while switching to the proportional-integral-derivative standby control mode.

[0037] proportionality coefficient K p This determines the instantaneous response strength of the control output to temperature deviation. In subway station ventilation and air conditioning systems, there is a significant thermal inertia delay between the controlled object (such as chiller outlet water temperature and fan frequency) and the temperature response; the time constant of temperature change in a single area is typically between 90 and 300 seconds. If K...p If K is too small (below 1), the system's correction for temperature deviations is insufficient, and the temperature recovers to the set value too slowly, which can easily cause continuous temperature overshoot, especially during peak passenger flow periods; if K p An excessively high value (over 10) will cause frequent and significant fluctuations in the operating frequency of the equipment, which will not only accelerate the wear and tear on the frequency converter and actuator, but may also trigger the chiller unit to shut down due to drastic changes in the outlet water temperature setpoint.

[0038] Based on the Ziegler-Nichols tuning method and field step response test data, the critical proportional gain of a subway station ventilation and air conditioning system in PID control mode is typically between 3 and 8. This invention will use K... p The value range is limited to 1 to 10, covering all working conditions from conservative adjustment to rapid response. In actual deployment, the middle value of 3 to 5 is usually sufficient to obtain a good balance between dynamic response and steady-state accuracy.

[0039] Integral coefficient K i This is used to eliminate steady-state residual error and ensure that the temperature in each area can accurately converge to the set value. Since the ambient temperature in subway stations is affected by multiple disturbances such as outdoor weather, passenger flow heat dissipation, and train piston wind, proportional control alone will have inherent steady-state errors. The introduction of integral action can accumulate historical deviations and apply continuous corrections.

[0040] Integral time constant T i =1 / K i The engineering value of K is usually related to the thermal inertia time constant of the controlled system. For a station thermal response process with a time constant of 90 to 300 seconds, the integral time constant is usually set to 100 to 500 seconds, corresponding to K. i The value is between 0.002 and 0.01. However, considering that the PID backup control mode is only temporarily activated under abnormal operating conditions where the proxy model is inaccurate, and a faster deviation elimination capability is needed to quickly restore thermal comfort, this invention will use K... i The value range is increased to 0.01 to 0.5, corresponding to an integral time constant of 2 to 100 seconds. The lower limit of 0.01 ensures continuous correction of slow drift deviations, while the upper limit of 0.5 prevents control output overshoot and system oscillation caused by integral saturation. Simulation verification shows that within this range, the integral action can effectively eliminate steady-state deviations without causing overshoot.

[0041] Differential coefficient K d To predict the changing trend of temperature deviation, damping is applied in advance to suppress overshoot and oscillation. In the subway station environment, temperature changes are relatively gradual, and high-frequency noise mainly comes from sensor measurement fluctuations and instantaneous disturbances in passenger flow. Excessive derivative action will amplify the noise, causing control output jitter.

[0042] Therefore, the present invention will use Kd The lower limit is set to 0, allowing the differential link to be shut down during non-operational periods when temperature changes are gradual or noise is high. PI control is used to simplify parameter tuning and improve robustness. The upper limit is set to 2, corresponding to a differential time constant of approximately 30 to 60 seconds, which matches the dominant time constant of the station's temperature response. This provides moderate advance adjustment during sudden changes in passenger flow, reducing temperature peak deviation. If K... d If the value exceeds 2, the amplification effect of the differential term on sensor noise will be significantly enhanced, which can easily cause high-frequency jitter in the frequency command of the inverter, resulting in abnormal equipment vibration and increased energy consumption.

[0043] The above K p K i K d The range of values ​​for these parameters collectively constitutes a PID parameter stability domain suitable for subway station ventilation and air conditioning systems. Within this stability domain, the closed-loop system maintains stable convergence regardless of the parameter combination. When switching to the PID standby control mode, this invention employs a preset default parameter combination (e.g., K...). p =3.0, K i =0.05, K d =0.5), this combination, verified through on-site tuning, can quickly restore temperature control under most abnormal operating conditions without real-time intervention from maintenance personnel. Furthermore, the aforementioned range provides clear boundaries for subsequent parameter fine-tuning based on specific station characteristics. Those skilled in the art can obtain the optimal PID parameters for a specific station within this range using conventional trial-and-error methods or self-tuning algorithms.

[0044] The value ranges of the above-mentioned PID parameters are all preferred engineering ranges based on thermodynamic principles and control theory. When applying this invention to a specific station, those skilled in the art only need to make adaptive adjustments within the range according to the actual thermal response test data of the station. These fine adjustments do not depart from the protection scope of this invention.

[0045] This invention also discloses an adaptive model-based predictive control system for energy-saving electromechanical systems in urban rail transit, comprising edge computing devices deployed in subway stations, a station sensor network, a controller, chillers, chilled water pumps, cooling water pumps, cooling towers, and air conditioning fans. The edge computing devices include: The neural network proxy model module is used to output temperature field distribution prediction data and equipment energy consumption prediction data of multiple areas inside the station based on the input outdoor environmental parameters, passenger flow parameters and equipment operating parameters. The neural network proxy model module is a deep neural network model that is deployed on an edge computing device after offline training. The control strategy model module is used to output equipment operation parameter instructions within the current control cycle based on the input current outdoor environmental parameters and current passenger flow parameters. The control strategy model module is a strategy network model that is pre-trained by reinforcement learning using the neural network agent model module as the interaction environment, and is then deployed on the edge computing device. The data acquisition module is used to acquire the current outdoor environmental parameters and current passenger flow parameters collected by the station sensor network in each control cycle, and input the current outdoor environmental parameters and current passenger flow parameters into the control strategy model module; The instruction issuing module is used to issue the device operating parameter instructions output by the control strategy model module to the controller; The anomaly monitoring and switching module is used to continuously monitor the residual between the predicted data of the station's internal temperature distribution by the neural network proxy model module and the actual data of the station's internal temperature distribution collected by the sensors during system operation. When the residual exceeds the preset residual threshold, the control mode is automatically switched from the control strategy model module to the proportional-integral-derivative standby control mode in the next control cycle. The controller is communicatively connected to the edge computing device and is used to receive the device's operating parameter instructions, and control the chiller, chilled water pump, cooling water pump, cooling tower, and air conditioning fan to execute corresponding operating frequencies or valve openings according to the device's operating parameter instructions.

[0046] The present invention offers at least the following advantages: The urban rail electromechanical energy-saving predictive control method and system based on an adaptive model, as described in this invention, replaces online computational fluid dynamics simulation with an offline trained neural network surrogate model, compressing the single prediction time to the millisecond level, enabling environmental state feedback to match the requirements of real-time control cycles. Training the reinforcement learning strategy is completed offline in a virtual environment constructed by the surrogate model, avoiding short-term deviations in thermal comfort and frequent equipment start-ups and shutdowns that may occur during online trial-and-error at real operating stations. In each control cycle, the edge computing device acquires outdoor temperature, humidity, and passenger flow data from the existing sensor network via the Modbus TCP protocol, combining this with locally cached historical control parameters to form state inputs. The strategy model directly outputs operating commands such as the chiller outlet water temperature setpoint, pump frequency, and fan frequency, which, after amplitude limiting, are sent to the controller for execution, forming a closed-loop regulation that requires no manual intervention. At the end of each control cycle, the edge computing device calculates the passenger flow change rate and the station's average temperature deviation rate, and compares them with preset thresholds to determine the heat load change status in real time. When either indicator exceeds the threshold, the system directly adopts the shortest control cycle in the next control cycle to quickly respond to disturbances and reduce temperature overshoot. When neither exceeds the threshold, the control cycle length is continuously adjusted using different linear formulas according to the operating or non-operating periods. During the operating period, the control cycle is shortened as the passenger flow change rate increases, while during the non-operating period, the control cycle is moderately shortened during occasional loads and kept longer during stable periods. This ensures timely control response while reducing equipment adjustment frequency and actuator wear. During operation, the cumulative residual between the surrogate model's predicted temperature and the sensor's measured temperature is continuously compared. The residual threshold is set more strictly during the operating period and moderately relaxed during the non-operating period. When the residual exceeds the corresponding threshold, the system automatically switches to the proportional-integral-derivative (PID) standby control mode and calculates the control output based on the deviation. At the same time, a model retraining signal is generated to ensure the continuity of environmental control during periods of model inaccuracy and to provide clear triggering conditions for model updates. The aforementioned functional modules are integrated into the edge computing device, working collaboratively with the station's existing sensors, controllers, and controlled equipment. This enables adaptive energy-saving control covering all operating conditions without requiring large-scale modifications to the main facilities.

[0047] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0048] Figure 1 This is a flowchart of the urban rail electromechanical energy-saving predictive control method based on an adaptive model as described in this invention.

[0049] Figure 2 This is an architecture diagram of the urban rail electromechanical energy-saving predictive control system based on an adaptive model as described in this invention. Detailed Implementation

[0050] The present invention will now be described in further detail with reference to specific embodiments, so that those skilled in the art can implement it based on the description.

[0051] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0052] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0053] Example 1 like Figure 1 and 2 As shown, this invention provides a predictive control method and system for energy saving in urban rail transit electromechanical systems based on an adaptive model, specifically as follows: The entire control framework revolves around an edge computing device deployed in the station control room. This device can be an edge computing gateway or an industrial PC equipped with a GPU computing unit, with at least 16GB of memory and at least 256GB of storage. The device connects to the station's existing control LAN via an industrial Ethernet interface, establishing communication connections with outdoor temperature and humidity sensors, indoor temperature sensors for various areas, passenger flow counting sensors, and lower-level programmable logic controllers (PLCs). The chiller units, chilled water pumps, cooling water pumps, cooling tower fans, and air conditioning fans are all existing ventilation and air conditioning system equipment in the station and do not require replacement. Only the analog output module of the PLC needs to be interfaced with the local controller of the chiller unit, and the inverters of each pump and fan are connected via their communication interfaces.

[0054] The implementation of this plan consists of two parts: offline training and online operation.

[0055] First, offline training is performed to obtain a neural network agent model and a control policy model. Both models are trained on a backend server or workstation, and then the model files are deployed to edge computing devices.

[0056] The training process of the neural network proxy model is as follows: Based on the CAD drawings of the target subway station, a 3D geometric model of the station is built using SpaceClaim. The model includes the main building structures such as the concourse public area, platform public area, equipment management rooms, air supply vents, return air vents, entrance and exit passages, escalator openings, and platform screen doors. After extracting the fluid computation domain, a polyhedral mesh is generated using Fluent Meshing, with local refinement in the areas of air supply and return air vents. The mesh quality needs to meet the following requirements: the maximum mesh edge length does not exceed one-third of the minimum building structure length, the skewness is not greater than 0.7, and the minimum orthogonality quality is greater than 0.1.

[0057] The mesh model is imported into Fluent, and the air density, specific heat capacity, thermal conductivity, and viscosity are set. A standard k-ε turbulence model is selected, and the boundary condition type and solver parameters are set according to the actual station conditions. Simultaneously, a model of the ventilation and air conditioning system is built in Trnsys, including chiller unit modules, chilled water pump modules, cooling water pump modules, cooling tower modules, air conditioning unit modules, and duct and water pipe modules. Co-simulation between the two software programs is achieved through a dynamic data interface between Trnsys and Fluent. At each time step, Trnsys transmits boundary conditions such as supply air temperature, supply air humidity, and supply air volume to Fluent. After calculating the flow field and temperature field, Fluent returns the return air temperature, return air humidity, and average temperature of each area to Trnsys, which then calculates the equipment energy consumption.

[0058] To generate training data, a sufficient number of boundary condition combinations need to be constructed. Outdoor ambient temperature is taken from -10℃ to 45℃, with values ​​in 5℃ increments; outdoor relative humidity is taken from 20% to 90%, with values ​​in 10% increments; passenger flow ranges from 0 people per hour to the station's maximum designed passenger flow, with values ​​in 10% increments of the maximum passenger flow; the operating frequency of each piece of equipment is taken from 20% to 100% of its rated frequency, also with values ​​in 10% increments. Combining these parameters results in 1200 sets of boundary conditions. Each set of boundary conditions is sequentially input into the co-simulation model for offline simulation, with each simulation lasting 24 hours and a time step of 60 seconds. After the simulation, the temperature values ​​of each region at each time step, as well as the energy consumption values ​​of the chiller, chilled water pump, cooling water pump, cooling tower fan, and air conditioning fan, are recorded. These boundary conditions and their corresponding simulation results together constitute the simulation dataset.

[0059] Next, a deep neural network is constructed. The input layer of the network has 8 neurons, corresponding to 8 input quantities: outdoor ambient temperature, outdoor relative humidity, passenger flow, chiller outlet water temperature setpoint, chilled water pump frequency, cooling water pump frequency, cooling tower fan frequency, and air conditioning fan frequency. This is followed by 3 fully connected hidden layers, each with 128 neurons, using ReLU as the activation function. The number of neurons in the output layer equals the total number of temperature measurement areas within the station plus the total number of devices used for energy consumption calculation. In this implementation, 20 temperature measurement areas are divided within the station, plus the energy consumption values ​​of 5 devices, resulting in a total of 25 neurons in the output layer.

[0060] During training, outdoor environmental parameters, passenger flow parameters, and equipment operating parameters from the simulation dataset are used as input to the network, while the corresponding temperature of each area and energy consumption of each device are used as supervision signals. The loss function is mean squared error, the optimizer is Adam, the initial learning rate is set to 0.001, and 64 samples are taken in each batch. Training continues until the mean squared error between the network output and the supervision signal drops below 0.001, or the number of training iterations reaches 10,000. After training, the network's weight matrices W1, W2, W3, W4 and bias vectors b1, b2, b3, b4 are saved; this is the trained neural network proxy model.

[0061] The control policy model is trained using reinforcement learning. The specific steps are as follows: the control policy model is obtained through reinforcement learning training, and the previously trained neural network agent model acts as the interaction environment for the reinforcement learning agent. In other words, the equipment operation parameter commands issued by the reinforcement learning agent during training are first passed to the agent model, which quickly calculates the corresponding temperature field distribution and energy consumption data, and then returns them to the agent as feedback. This completely avoids trial and error in a real train station.

[0062] The state space S of reinforcement learning consists of 8 components: outdoor ambient temperature T1, outdoor relative humidity H1, passenger flow P, chiller unit outlet water temperature setpoint T2 from the previous control cycle, chilled water pump frequency f1 from the previous control cycle, cooling water pump frequency f2 from the previous control cycle, cooling tower fan frequency f3 from the previous control cycle, and air conditioning fan frequency f4 from the previous control cycle. The action space A consists of 5 components: chiller unit outlet water temperature setpoint T3 from the current control cycle, chilled water pump frequency f5, cooling water pump frequency f6, cooling tower fan frequency f7, and air conditioning fan frequency f8. The setting of the reward function R is a key aspect of this process, and the form adopted... T is here. i P is the temperature of the i-th region predicted by the surrogate model, where T4 is the set temperature, taken as 26℃; jα is the real-time power of the j-th device; α is the thermal comfort weighting coefficient, set to 2.0; β is the energy consumption weighting coefficient, set to 0.1. The ratio of α to β is approximately 20:1. This setting is intended to minimize energy consumption while ensuring that the temperature in each area is close to the set value.

[0063] The training algorithm used is the Deep Deterministic Policy Gradient Algorithm (DDPG). Both the policy network and the value network employ a 3-layer fully connected structure with 256 neurons in each hidden layer and ReLU activation function. A Sigmoid activation function is added to the output layer of the policy network to map the output to the normalized interval of each action component. The experience replay buffer has a capacity of 100,000 records, and every 100 training steps, 128 experience samples are randomly sampled from the buffer to update the policy and value networks. The learning rate for both the policy and value networks is 0.0005, and the discount factor is 0.98. Each training epoch simulates 24 hours of operation, with a control period of 60 seconds, resulting in 1440 steps per epoch. Training continues until the standard deviation of the cumulative reward function value drops below 0.05 after 100 consecutive training epochs. After training, the parameters of the policy network are extracted to obtain the control policy model.

[0064] After both models are trained, the model files are transferred to the edge computing device to complete the deployment.

[0065] During the online operation phase, after the edge computing device is powered on and started, the weight parameters of the neural network agent model and the policy network parameters of the control policy model are loaded first. At the same time, a local cache is opened to store the device operation parameters for each control cycle, which can be called in the next cycle.

[0066] The duration of the control cycle is not fixed, but dynamically adjusted according to the actual operating status of the station. The following example illustrates the online operation process with a control cycle of 30 seconds during operating hours and 180 seconds during non-operating hours.

[0067] At the start of each control cycle, the edge computing device sends a read command to the station's sensor network via the Modbus TCP protocol to retrieve the current T1 and H1 data collected by the outdoor temperature and humidity sensor, as well as the current passenger flow P data collected by the passenger flow counting sensor. After parsing the data frame, it reads T2, f1, f2, f3, and f4, recorded after the previous control cycle, from the local cache. These eight parameters are combined to form the state vector S for the current cycle.

[0068] The state vector S is input into the policy network deployed on the edge computing device. After one forward propagation calculation, the policy network outputs an action vector A, which contains five quantities: T3, f5, f6, f7, and f8. To prevent the model from occasionally outputting values ​​that exceed the device's tolerance, these quantities need to be limited: T3 is limited to between 7°C and 12°C, and f5, f6, f7, and f8 are limited to between 20Hz and 50Hz. The limited action vector A' is the device operating parameter command to be issued in this control cycle.

[0069] These instructions are encapsulated into control data frames via the Modbus TCP protocol and sent to the programmable logic controller (PLC). Upon receiving the PLC, it converts T3 into a 4mA to 20mA current signal and sends it to the chiller unit's local controller. The chiller unit adjusts the compressor's operating frequency and the opening of the electronic expansion valve based on this signal. Simultaneously, the PLC sends f5, f6, f7, and f8 to the corresponding frequency converters, which adjust the speed of their respective motors. After instruction execution is complete, the PLC returns an acknowledgment frame to the edge computing device. The edge computing device then writes the executed T3, f5, f6, f7, and f8 signals of the current cycle into its local cache, which will then be used as T2, f1, f2, f3, and f4 in the state vector for the next cycle.

[0070] At the end of each control cycle, the edge computing device performs two additional tasks. One is to retrieve the predicted temperature T for each region from the neural network agent model at the beginning of the cycle. p1 To T pN Simultaneously, the actual temperature T collected in each area is read through the sensor network. ri Another thing is to calculate the control cycle length for the next control cycle.

[0071] III. Adaptive Adjustment of Control Cycle At the end of each control cycle, the edge computing device calculates two metrics: the passenger flow change rate ΔP and the station average temperature deviation rate ΔT. The formula for ΔP is ΔP = |P| c -P p | / P max , where P c This is the current passenger flow data just collected this period, P. p It's the customer traffic from the previous period's cache, P max This is the station's maximum designed passenger capacity. The formula for calculating ΔT is... T ri T1 is the actual temperature of the i-th zone, T4 is the set temperature, and N is the total number of zones within the station.

[0072] After calculation, the edge computing device compares ΔP with a first preset threshold ε1, and ΔT with a second preset threshold ε2. The value of ε1 ranges from 0.03 to 0.10, and in this embodiment, it is set to 0.05; the value of ε2 ranges from 0.02 to 0.08, and in this embodiment, it is set to 0.04. If ΔP exceeds 0.05, or ΔT exceeds 0.04, it is determined that the station is in a state of rapid change in heat load. In this case, the duration T of the next control cycle... c The system directly takes the lower limit of the first value range. The first value range is 10 seconds to 60 seconds, and the lower limit is 10 seconds. In other words, the system will respond to sudden increases in passenger flow or temperature deviations at the fastest possible pace.

[0073] If ΔP does not exceed 0.05 and ΔT does not exceed 0.04, it indicates that the station's heat load is relatively stable. At this time, the edge computing device determines whether it is currently in an operating or non-operating period based on the current system time. The operating period generally refers to 6:00 to 23:00 every day, and the non-operating period refers to 23:00 to 6:00 the next day.

[0074] During operating hours, T c Calculated according to the first linear formula: T c =T a +(1-ΔP / ε1)×(T b -T a Ta is the lower limit of the first range of values, 10 seconds, and T is... b The upper limit of the first value range is 60 seconds. In this formula, the smaller ΔP is, the longer the control cycle, up to 60 seconds; the larger ΔP is, the shorter the control cycle, down to 10 seconds. This means that when passenger flow is stable, equipment adjustments don't need to be as frequent, reducing the number of inverter operations; when passenger flow changes significantly, the control frequency automatically increases to ensure that temperature regulation keeps pace with load changes.

[0075] During non-operational hours, T c Calculated according to the second linear formula: T c =T f -ΔP / ε1×(T f -T e ). T e It is the lower limit of the second range of values, 120 seconds, T f It is the upper limit of the second value range, 300 seconds. The design idea behind this formula is that during non-operational periods, a relatively long control cycle should be used to reduce equipment wear. Therefore, when ΔP is close to 0, T... c It's close to 300 seconds; however, if there are sporadic changes in passenger flow at night, such as construction or cleaning work, ΔP will also increase, and in this case, T... cThe time will be shortened towards 120 seconds, and the system will still have a certain responsiveness, so that the temperature deviation will not accumulate due to the control cycle being too long.

[0076] Through this dynamic adjustment mechanism, the duration of the control cycle can be smoothly varied between 10 seconds and 300 seconds according to the actual working conditions, which not only ensures the response speed during the operating period, but also reduces unnecessary actions of the equipment during the non-operating period.

[0077] IV. Anomaly Monitoring and Backup Mode Switching Even if the surrogate model has undergone extensive offline training, its prediction accuracy may still decrease during actual operation due to equipment aging, sensor drift, or extreme conditions where training data is not covered. This solution will continuously monitor this situation during operation.

[0078] At the end of each control cycle, the edge computing device outputs the temperature T predicted by the neural network proxy model for each region. pi and the actual temperature T of each area collected by the sensor ri Compare and calculate the total residual E = ∑|T pi -T ri Meanwhile, the system maintains a preset residual threshold E. th E th The value of E also depends on the operating status: during the operating period, E th Let N × 0.8℃, where N is the total number of regions. Here, N is taken as 20, so E th That is 16℃; during non-operational hours, E th Take N × 1.5℃, which is 30℃.

[0079] If E is less than or equal to E th This indicates that the prediction accuracy of the surrogate model is still within an acceptable range, and the next control cycle will continue to be driven by the control strategy model to output equipment operating parameter commands. If E is greater than E... th This indicates that the surrogate model's current predictions have deviated significantly, and continuing to use the control strategy model may affect the control performance. At this point, the edge computing device automatically switches the control mode from control strategy model-driven to proportional-integral-derivative (PID) standby control mode in the next control cycle.

[0080] In the proportional-integral-derivative (PID) backup control mode, edge computing devices no longer rely on agent models and policy networks, but directly adjust based on the deviation between the actual temperature and the set temperature of each region. For the i-th region, the set temperature T4 is 26℃, and the actual temperature T ri Deviation E from T4 i =T4-T ri The control output U is calculated according to the standard PID formula: In this embodiment, K p Take 3.0, K i Take 0.05, K d We set it to 0.5. The calculated U is then mapped to the chiller unit's outlet water temperature setpoint, the frequency of each water pump, and the frequency of each fan, and sent to the controller for execution. Simultaneously with the switch, the edge computing device also generates a model retraining trigger signal, notifying maintenance personnel to retrain the neural network proxy model and control strategy model using recently accumulated operational data.

[0081] V. Long-term system operation and model updates The system operates on a 7-day cycle, generating daily logs that record outdoor environmental parameters, passenger flow parameters, equipment operating parameters, predicted temperature, actual temperature, and energy consumption data for each control cycle. When a model retraining trigger signal is generated, maintenance personnel export the actual operating data from the edge computing device for the past 7 days, add this data to the original simulation dataset, and incrementally train the neural network proxy model following the steps of the offline training phase. Then, they retrain the control strategy model using the updated proxy model. After training is complete, the new model is redeployed to the edge computing device, and the system can resume its normal operation driven by the control strategy model.

[0082] The scheme of Example 1 was deployed in a subway station for a 30-day continuous operation test. The station's ventilation and air conditioning system included 3 chillers, 3 chilled water pumps, 3 cooling water pumps, 2 cooling towers, and 8 air conditioning fans. The station was divided into 20 temperature measurement zones. During the test, the outdoor ambient temperature ranged from -5℃ to 38℃, and the daily passenger flow fluctuated between 8,000 and 58,000 passengers, covering all operating conditions including morning peak, off-peak, evening peak, and non-operating periods. The set temperature T4 was 26℃, the number of neurons in the hidden layer of the neural network surrogate model was 128, and in the control strategy model, α was 2.0, β was 0.1, the first preset threshold ε1 was 0.05, the second preset threshold ε2 was 0.04, the first value range was 10 seconds to 60 seconds, and the second value range was 120 seconds to 300 seconds. The preset residual threshold E during the operating period was... th Take N×0.8℃ (16℃ when N=20), and the preset residual threshold for non-operational periods is N×1.5℃ (30℃). In the PID standby control mode, K... p Take 3.0, K i Take 0.05, K d Take 0.5.

[0083] During the test, the predicted temperature distribution of the 20 regions by the neural network surrogate model was recorded every control cycle and compared with the actual temperature values ​​collected by the sensors in each region at that time. The mean absolute error (MAE) and root mean square error (RMSE) of each region were calculated. The cumulative number of valid samples over 30 days was 86,400, and the statistical results are shown in Table 1.

[0084] Table 1. Error statistics of temperature prediction for different areas of the station using the neural network surrogate model. As shown in Table 1, the average MAE of the neural network surrogate model for predicting temperatures in 20 regions is 0.42℃, and the average RMSE is 0.55℃. These results demonstrate that the deep neural network surrogate model trained based on CFD simulation data can reproduce the temperature field distribution in multiple regions within the station with high accuracy, providing a sufficiently accurate interactive environment for reinforcement learning control strategies.

[0085] Energy consumption comparison tests were conducted on the following seven control schemes at the same test station: Comparative Example 1 (PID Control): The original station's existing PID control scheme, where each PID parameter operates according to the on-site set value, and the control cycle is fixed at 60 seconds.

[0086] Comparative Example 2 (Rule Control): Based on a preset threshold rule control scheme, the equipment operating frequency is set in segments according to outdoor temperature and passenger flow, and the control cycle is fixed at 120 seconds.

[0087] Comparative Example 3 (Existing Reinforcement Learning Scheme): A neural network trained based on measured operating data is used as the reinforcement learning interaction environment (the network structure is a single hidden layer with 32 neurons, and the training data is taken from the station's operating records over the past three months). The same DDPG algorithm and reward function weight coefficients (α=2.0, β=0.1) as this invention are used. The control cycle is fixed at 60 seconds, and no residual monitoring and PID backup switching mechanism are set.

[0088] Comparative Example 4: Based on Comparative Example 3, only the neural network proxy model trained with the measured data is replaced with the deep neural network proxy model trained with the CFD simulation data of this invention, while the rest remains unchanged.

[0089] Comparative Example 5: Based on Comparative Example 3, only the dynamic adaptive adjustment mechanism of the control cycle of the present invention is added (ε1=0.05, ε2=0.04, the first value range is 10~60 seconds, and the second value range is 120~300 seconds), and the rest remain unchanged.

[0090] Comparative Example 6: Based on Comparative Example 3, only the residual monitoring and PID backup switching mechanism of the present invention is added (E th(Temperature is 16℃ during operating hours and 30℃ during non-operating hours), all other temperatures remain unchanged.

[0091] Example 1 (Complete Solution of the Invention): It simultaneously includes a deep neural network surrogate model trained with CFD simulation data, a dynamic adaptive adjustment mechanism for the control cycle, and a residual monitoring PID backup switching mechanism.

[0092] Each scheme was operated sequentially under the same outdoor weather conditions and passenger flow, with each scheme running continuously for 7 days. The stable operating data from the middle 5 days was used as the valid statistical sample to eliminate the impact of the transition period. The power consumption of each device in the station's ventilation and air conditioning system was collected in real time by smart meters and the energy consumption was calculated cumulatively. The comparison results of daily energy consumption and 7-day cumulative energy consumption are shown in Table 2.

[0093] Table 2. Energy consumption comparison of various control schemes (unit: kWh) Based on the cumulative energy consumption data in Table 2, the energy saving rate of each scheme relative to different benchmarks and the individual contribution rate of each distinguishing feature were calculated, and the results are shown in Table 3.

[0094] Table 3 Comparison of energy saving rates of various control schemes As can be seen from Tables 2 and 3: (1) Comparative Example 3 (existing reinforcement learning scheme) achieved an energy saving rate of 18.1% relative to PID control, which is comparable to the level of similar technologies reported in the literature.

[0095] (2) Comparative Example 4 (only replaced with CFD surrogate model) improved the energy saving rate to 20.0%, which is 2.4 percentage points higher than Comparative Example 3. This shows that the high-precision surrogate model enables the reinforcement learning strategy to obtain more accurate temperature field feedback during the training phase, thereby learning a better control strategy.

[0096] (3) Comparative Example 5 (with only dynamic adjustment of control cycle added) achieved an energy saving rate of 22.0%, which is 4.8 percentage points higher than that of Comparative Example 3. It is the one with the largest contribution among all individual features, reflecting the significant effect of fine matching between control cycle and heat load change.

[0097] (4) Comparative Example 6 (with only the addition of residual monitoring switching) only increased the energy saving rate by 0.5 percentage points compared to Comparative Example 3 under steady-state conditions. Its contribution is mainly reflected in the protection of the system's robustness under abnormal conditions.

[0098] (5) Example 1 (the complete solution of the present invention) integrates the above three features simultaneously, achieving an energy saving rate of 23.4%, which is an additional 6.5 percentage points higher than that of Comparative Example 3. This improvement significantly exceeds the range that a person skilled in the art could reasonably expect from conventional parameter tuning of reinforcement learning control strategies (usually 1 to 3 percentage points), demonstrating the unpredictability of the technical effect of the present invention.

[0099] To verify the protective effect of residual monitoring and PID backup switching mechanism under abnormal operating conditions, a measurement deviation fault of +3℃ (simulating sensor drift caused by dust accumulation) was artificially injected into the outdoor temperature sensor from 10:00 to 11:00 AM on the 15th day, lasting for 1 hour. The average station temperature deviation, energy consumption changes, and system switching behavior during this fault period were recorded for Comparative Example 3 (without switching function) and Example 1 (this invention, with switching function), respectively. The results are shown in Table 4.

[0100] Table 4 Comparison of Sensor Deviation Fault Injection Test Results As shown in Table 4, in Comparative Example 3, after the sensor deviation fault occurred, the control strategy model continuously output inappropriate commands based on the erroneous state due to the inaccurate input parameters of the surrogate model, causing the average station temperature to deviate from the set value by as much as 4.2℃, and the energy consumption to reach 298 kWh within one hour of the fault. In contrast, Example 1 detected the residual exceeding the operating period threshold within 10 seconds of the fault occurrence and automatically switched to the PID backup control mode, causing the temperature deviation to quickly converge to within 0.8℃, and the energy consumption within one hour of the fault was 254 kWh, a reduction of approximately 14.8% compared to Comparative Example 3. Simultaneously, Example 1 automatically generated a model retraining trigger signal during the switchover for subsequent model updates. These results demonstrate that the residual monitoring and PID backup switching mechanism can quickly restore temperature control accuracy under abnormal operating conditions, effectively avoiding comfort degradation and energy waste caused by inappropriate control strategies. This is a key safety measure to ensure the long-term stable operation of high-precision surrogate models and dynamic control cycles in practical engineering.

[0101] Based on the above test results, the present invention demonstrates technical effects that exceed the reasonable expectations of those skilled in the art in terms of surrogate model prediction accuracy, energy saving rate improvement, synergistic effect of various features, and protection against abnormal operating conditions. It has significant inventiveness and engineering application value.

[0102] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A predictive control method for energy saving in urban rail transit electromechanical systems based on an adaptive model, characterized in that, Includes the following steps: The trained neural network proxy model and control strategy model are deployed on the edge computing equipment of the subway station. The neural network proxy model is used to output temperature field distribution prediction data and equipment energy consumption prediction data of multiple areas inside the station based on the input outdoor environmental parameters, passenger flow parameters and equipment operation parameters. Within each control cycle, the edge computing device acquires the current outdoor environmental parameters and current passenger flow parameters collected by the station sensor network, and inputs the current outdoor environmental parameters and current passenger flow parameters into the control strategy model; the control strategy model is obtained through reinforcement learning pre-training based on the temperature field distribution prediction data and equipment energy consumption prediction data provided by the neural network proxy model as the interaction environment, and outputs the equipment operation parameter instructions for the current control cycle according to the input parameters; The device operating parameter command is sent to the controller, which then controls the chiller, chilled water pump, cooling water pump, cooling tower, and air conditioning fan to execute the corresponding operating frequency or valve opening. During system operation, the residual between the predicted temperature distribution data inside the station predicted by the neural network surrogate model and the actual temperature distribution data inside the station collected by the sensors is continuously monitored. When the residual exceeds a preset residual threshold, the edge computing device automatically switches to the proportional-integral-derivative (PID) standby control mode in the next control cycle. The preset residual threshold is set with different values ​​depending on whether the station is in operation or not. In the PID standby control mode, the edge computing device calculates the control output based on the deviation between the set temperature and the actual temperature in each area and maps it to the device operating parameter command. At the same time, it generates a model update trigger signal to indicate that the neural network surrogate model needs to be retrained.

2. The urban rail transit electromechanical energy-saving predictive control method based on an adaptive model as described in claim 1, characterized in that, The training steps of the neural network proxy model include: Based on the computational fluid dynamics simulation model of the ventilation and air conditioning system of a subway station, different values ​​are selected for the following: outdoor ambient temperature in the range of -10℃ to 45℃ with a step size of 5℃; outdoor relative humidity in the range of 20% to 90% with a step size of 10%; passenger flow parameters in the range of 0 people / hour to the maximum design passenger flow of the station with a step size of 10% of the maximum passenger flow; and equipment operating parameters in the range of 20% to 100% of the rated operating range of each piece of equipment with a step size of 10%. These values ​​are combined to form no less than 1,000 sets of boundary conditions. Each set of boundary conditions is input into the computational fluid dynamics simulation model for offline simulation. Simulation data of temperature field distribution in multiple areas inside the station corresponding to each set of boundary conditions, as well as simulation data of equipment energy consumption of chiller units, chilled water pumps, cooling water pumps, cooling towers and air conditioning fans in the ventilation and air conditioning system, are obtained. The simulation dataset is composed of each set of boundary conditions and their corresponding simulation data. Construct a deep neural network consisting of an input layer, three fully connected hidden layers, and an output layer. The number of neurons in the input layer is the same as the total number of outdoor environmental parameters, passenger flow parameters, and equipment operation parameters. The number of neurons in the output layer is the same as the sum of the number of measurement points for temperature field distribution data in the multiple regions and the number of devices for equipment energy consumption data. The deep neural network is trained using the simulation dataset. During the training process, outdoor environmental parameters, passenger flow parameters, and equipment operating parameters are used as network inputs, and temperature field distribution simulation data of multiple areas inside the station and equipment energy consumption simulation data are used as supervision signals for the network output. The network weight parameters are adjusted through the backpropagation algorithm until the mean square error between the network output and the supervision signal is less than 0.001 or the number of training iterations reaches 10,000.

3. The urban rail electromechanical energy-saving predictive control method based on an adaptive model as described in claim 1, characterized in that, The specific implementation steps of the neural network proxy model for predicting the internal temperature distribution data of the station include: The edge computing device combines the current outdoor ambient temperature T1, the current outdoor relative humidity H1, the current passenger flow P, the chiller outlet water temperature setpoint T3 for the current control cycle, the chilled water pump frequency f5, the cooling water pump frequency f6, the cooling tower fan frequency f7, and the air conditioning fan frequency f8 to form an input vector X = [T1, H1, P, T3, f5, f6, f7, f8]. The edge computing device inputs the input vector X into a locally deployed neural network proxy model. The neural network proxy model consists of an input layer, three fully connected hidden layers, and an output layer. The input layer has 8 neurons, the first hidden layer has 64 to 256 neurons, the second hidden layer has 64 to 256 neurons, the third hidden layer has 64 to 256 neurons, and the output layer has N neurons, where N is the total number of areas inside the station. Each hidden layer uses the ReLU activation function. The neural network proxy model performs forward propagation calculation on the input vector X. The output of the first hidden layer is G1 = ReLU(W1·X + b1), the output of the second hidden layer is G2 = ReLU(W2·G1 + b2), the output of the third hidden layer is G3 = ReLU(W3·G2 + b3), and the output of the output layer is Y = W4·G3 + b4. Here, W1, W2, W3, and W4 are weight matrices, and b1, b2, b3, and b4 are bias vectors. The weight matrices and bias vectors are parameter values ​​determined by the backpropagation algorithm during the offline training stage and stored in the edge computing device. The values ​​of each element in the output vector Y represent the predicted temperatures of zones 1 to N within the station. The edge computing device resolves the output vector Y into T. p1 T p2 Until T pN This serves as predictive data for the temperature field distribution in multiple areas within the station.

4. The urban rail transit electromechanical energy-saving predictive control method based on an adaptive model as described in claim 3, characterized in that, The training steps for the control strategy model include: The trained neural network agent model is used as the interaction environment of the reinforcement learning agent. The neural network agent model receives the equipment operation parameters output by the reinforcement learning agent and returns the corresponding temperature field distribution prediction data and equipment energy consumption prediction data for multiple areas inside the station. Define a state space S, an action space A, and a reward function R for reinforcement learning. The state space S consists of T1, H1, P, the chiller outlet water temperature setpoint T2 from the previous control cycle, the chilled water pump frequency f1 from the previous control cycle, the cooling water pump frequency f2 from the previous control cycle, the cooling tower fan frequency f3 from the previous control cycle, and the air conditioning fan frequency f4 from the previous control cycle. The action space A consists of T3, f5, f6, f7, and f8 from the current control cycle. The reward function R is set as follows: ； Among them, T i Let T4 be the predicted temperature for the i-th region, and P be the set temperature. j The predicted real-time power of the j-th device is obtained by the neural network proxy model through time difference calculation based on the device energy consumption prediction data output by the input outdoor environmental parameters, passenger flow parameters, and device operating parameters; α is the thermal comfort weight coefficient with a value range of 0.1 to 10, β is the energy consumption weight coefficient with a value range of 0.01 to 1, and the ratio of α to β is between 2:1 and 20:1 to prioritize energy consumption optimization under thermal comfort constraints; N is the total number of areas inside the station; and M is the total number of devices. The reinforcement learning agent is trained offline using a deep deterministic policy gradient algorithm. During training, the agent outputs device operation parameters in action space A based on the current state space S. The neural network agent model receives these parameters and returns the next state space S' and reward function value R. The agent stores the state space S, action space A, reward function value R, and next state space S' generated from each interaction as an experience sample in an experience replay buffer. Every K training steps, B experience samples are randomly sampled from the experience replay buffer to pair the policy network and value... The network is updated using gradient descent, where K ranges from 50 to 200, B ranges from 32 to 512, the learning rate η1 of the policy network ranges from 0.0001 to 0.001, the learning rate η2 of the value network ranges from 0.001 to 0.01, and the discount factor γ ranges from 0.95 to 0.

99. Training continues until the standard deviation of the cumulative reward function value over L consecutive training rounds is less than δ, where L ranges from 50 to 200 and δ ranges from 0.01 to 0.

1. After training, the control policy model is obtained.

5. The urban rail transit electromechanical energy-saving predictive control method based on an adaptive model as described in claim 4, characterized in that, The specific steps for the edge computing device to acquire the current outdoor environmental parameters and current passenger flow parameters collected by the station sensor network include: The edge computing device establishes a data communication connection with the station sensor network through an industrial Ethernet interface. At the beginning of each control cycle, the edge computing device sends a data request command to the station sensor network. In response to the data request command, the station sensor network encapsulates the current T1 and H1 collected by the outdoor temperature and humidity sensor and the current passenger flow P collected by the passenger flow counting sensor into data frames through the Modbus TCP protocol and sends them to the edge computing device. The edge computing device parses the received data frame, extracts T1, H1, and P, and reads T2, f1, f2, f3, and f4 recorded after the execution of the previous control cycle from the local cache of the edge computing device. The parameters are combined to form the state vector S = [T1, H1, P, T2, f1, f2, f3, f4] of the current control cycle.

6. The urban rail transit electromechanical energy-saving predictive control method based on an adaptive model as described in claim 5, characterized in that, The specific implementation steps for the control strategy model to output the equipment operating parameter instructions for the current control cycle include: The edge computing device inputs the state vector S into a control strategy model deployed locally. The control strategy model is composed of a strategy network. The strategy network performs forward propagation calculation on the input state vector S and outputs the action vector A = [T3, f5, f6, f7, f8] for the current control cycle. The edge computing device performs amplitude limiting processing on the parameter values ​​in the action vector A, limiting T3 to the range of 7°C to 12°C, and limiting f5, f6, f7 and f8 to the range of 20Hz to 50Hz respectively. The amplitude-limited action vector A' is used as the device operation parameter instruction in the current control cycle.

7. The urban rail transit electromechanical energy-saving predictive control method based on an adaptive model as described in claim 1, characterized in that, The duration of the control cycle is dynamically adjusted according to the station's operational status, and the dynamic adjustment steps include: At the end of each control cycle, the edge computing device calculates the current passenger flow change rate ΔP and the station average temperature deviation rate ΔT. The passenger flow change rate ΔP is calculated using the formula ΔP=|P c -P p | / P max Calculate, where P c For the current passenger flow, P p For the passenger flow in the previous control period, P max The maximum designed passenger flow for the station; The average temperature deviation rate ΔT of the station is calculated according to the formula. Calculate, where T ri T4 is the actual temperature of the i-th area, N is the set temperature, and N is the total number of areas inside the station. The value of i is any integer from 1 to N. The edge computing device compares the passenger flow change rate ΔP with a first preset threshold ε1, and compares the station average temperature deviation rate ΔT with a second preset threshold ε2. The first preset threshold ε1 ranges from 0.03 to 0.10, and the second preset threshold ε2 ranges from 0.02 to 0.

08. When the passenger flow change rate ΔP exceeds the first preset threshold ε1 or the station average temperature deviation rate ΔT exceeds the second preset threshold ε2, the edge computing device determines that the station is in a state of rapid heat load change, and adjusts the control cycle duration T in the next control cycle. c Adjust to the smaller value within the first value range, which is 10 seconds to 60 seconds; When the passenger flow change rate ΔP does not exceed the first preset threshold ε1 and the station average temperature deviation rate ΔT does not exceed the second preset threshold ε2, the edge computing device obtains the current system time and determines the station operation status. If the station is in operation, the edge computing device will adjust the control cycle duration T according to the first linear formula. c Adjustments are made within a first value range, and the first linear formula is T. c =T a +(1-ΔP / ε1)×(T b -T a ), where T a T is the lower limit of the first value range. b This is the upper limit of the first value range; If the station is in non-operational hours, the edge computing device will adjust the control cycle duration T according to the second linear formula. c The adjustment is made within a second value range, which is from 120 seconds to 300 seconds, and the second linear formula is T. c =T f -ΔP / ε1×(T f -T e ), where T e T is the lower limit of the second value range. f This is the upper limit of the second range of values.

8. The urban rail transit electromechanical energy-saving predictive control method based on an adaptive model as described in claim 7, characterized in that, The preset residual threshold E th During station operating hours, the value ranges from N×0.5℃ to N×1.0℃, and during non-operating hours, the value ranges from N×1.0℃ to N×2.0℃, where N is the total number of areas within the station.

9. The urban rail transit electromechanical energy-saving predictive control method based on an adaptive model as described in claim 8, characterized in that, Let E be the residual between the predicted temperature distribution data inside the station predicted by the neural network surrogate model and the actual temperature distribution data inside the station collected by the sensors. When E≤E th In the next control cycle, the edge computing device continues to use the control strategy model to output device operating parameter instructions; When E>E th At this time, the edge computing device automatically switches the control mode from control strategy model driven to proportional-integral-derivative (PID) standby control mode in the next control cycle. In PID standby control mode, the edge computing device determines the control mode based on the set temperature T4 and the actual temperature T of the i-th region. ri Deviation E between i =T4-T ri According to the formula The control output U is calculated and mapped to the chiller outlet water temperature setpoint, chilled water pump frequency, cooling water pump frequency, cooling tower fan frequency, and air conditioning fan frequency before being sent to the controller. The proportional coefficient K is used in this calculation. p The value range is from 1 to 10, and the integral coefficient K i The value ranges from 0.01 to 0.5, and the differential coefficient K d The value range is from 0 to 2; The edge computing device generates a model retraining trigger signal while switching to the proportional-integral-derivative standby control mode.

10. An urban rail transit electromechanical energy-saving predictive control system based on an adaptive model, characterized in that, This includes edge computing devices deployed in subway stations, station sensor networks, controllers, chillers, chilled water pumps, cooling water pumps, cooling towers, and air conditioning fans. The edge computing devices include: The neural network proxy model module is used to output temperature field distribution prediction data and equipment energy consumption prediction data of multiple areas inside the station based on the input outdoor environmental parameters, passenger flow parameters and equipment operating parameters. The neural network proxy model module is a deep neural network model that is deployed on an edge computing device after offline training. The control strategy model module is used to output equipment operation parameter instructions within the current control cycle based on the input current outdoor environmental parameters and current passenger flow parameters. The control strategy model module is a strategy network model that is pre-trained by reinforcement learning using the neural network agent model module as the interaction environment, and is then deployed on the edge computing device. The data acquisition module is used to acquire the current outdoor environmental parameters and current passenger flow parameters collected by the station sensor network in each control cycle, and input the current outdoor environmental parameters and current passenger flow parameters into the control strategy model module; The instruction issuing module is used to issue the device operating parameter instructions output by the control strategy model module to the controller; The anomaly monitoring and switching module is used to continuously monitor the residual between the predicted data of the station's internal temperature distribution by the neural network proxy model module and the actual data of the station's internal temperature distribution collected by the sensors during system operation. When the residual exceeds the preset residual threshold, the control mode is automatically switched from the control strategy model module to the proportional-integral-derivative standby control mode in the next control cycle. The controller is communicatively connected to the edge computing device and is used to receive the device's operating parameter instructions, and control the chiller, chilled water pump, cooling water pump, cooling tower, and air conditioning fan to execute corresponding operating frequencies or valve openings according to the device's operating parameter instructions.

Images