Active cooling and power optimization system of plateau turbine generator

Abstract

The invention relates to the technical field of thermal management optimization, and discloses an active cooling and power optimization system of a plateau turbine generator, which comprises a multi-physics field modeling module, a reinforcement learning decision module, a strategy optimization training module and an edge execution control module, and is used for resolving a temperature field / flow field / stress field coupling state in real time, providing environment dynamic characterization, and optimizing the power of the plateau turbine generator. According to the method, multi-source sensing data is processed, a cooling liquid flow rate and cooling fin angle control instruction is generated, strategy iteration optimization is carried out in a digital twin environment, the physical rationality and energy efficiency optimality of a control algorithm are ensured, a lightweight strategy network is deployed, and sub-second-level real-time response and fault safety protection are achieved. Through intelligent control and multi-physical field cooperation, the temperature adjusting precision and the system response speed are improved, efficient power output of the turbine is kept under different altitudes and climate conditions through a self-adaptive strategy, and a reliable clean energy solution is provided for high-altitude areas in combination with multiple safety protection mechanisms.

Description

Technical Field

[0001] The present invention relates to the field of thermal management optimization, and more particularly, to an active cooling and power optimization system for a plateau turbine generator. Background Art

[0002] Turbine generators in plateau areas (elevation ≥3000m) face a unique thermal management crisis. The air density is 30%-40% lower than that in plains, the heat exchange efficiency of traditional air cooling systems is reduced by more than 50%, the temperature difference between day and night can reach 25°C+, and sudden strong winds (≥15m / s) cause transient power fluctuations.

[0003] The defects of existing control strategies are as follows: traditional PID control relies on a fixed parameter set and is difficult to adapt to the nonlinear system characteristics of multi-variable coupling. Therefore, it is proposed to develop a multi-physics field coupled DQN / PPO algorithm to dynamically adjust the coolant flow rate and heat sink angle through real-time environmental data (air pressure, temperature, speed). Summary of the Invention

[0004] The present invention provides an active cooling and power optimization system for a plateau turbine generator, which solves the technical problems in related technologies.

[0005] The present invention provides an active cooling and power optimization system for a plateau turbine generator, comprising:

[0006] Multi-physics modeling module: Establishes a coupled dynamic model covering heat conduction, convection heat transfer, heat sink efficiency, and hydraulic systems. It constructs a unified state space through three-dimensional heat conduction equations, fluid mechanics equations, and equipment constraint equations to achieve normalization of multiple physical quantities.

[0007] Reinforcement Learning Decision Module: Design the observation and action space for reinforcement learning, build a hybrid network architecture that integrates LSTM time series processing and GNN flow field analysis, develop a multi-objective reward function and integrate physical constraints, establish a priority experience replay mechanism, and form an environmental perception-decision-making closed loop;

[0008] Policy Optimization Training Module: This module uses generalized advantage estimation and the clip policy gradient method, combined with value function optimization and entropy regularization techniques, to embed the physical constraints of the Navier-Stokes equation residual into the loss function. This module achieves efficient training of the policy network through adaptive learning rate adjustment.

[0009] Edge execution control module: Quantize and pruned the training model for lightweight processing, designed delay compensation prediction and Lyapunov safety barriers, established a digital twin online calibration mechanism, developed a fault fast switching system, and realized real-time and reliable operation of the strategy on the edge computing node.

[0010] Furthermore, in the Multiphysics Modeling module, perform the following steps:

[0011] S110, Turbine Heat Conduction Modeling: Establishing three-dimensional unsteady heat conduction equations;

[0012] S120, Convective Heat Transfer Modeling: Constructing the coupled coolant-air two-phase flow equations;

[0013] S130, heat sink efficiency modeling: establish the heat sink angle-heat dissipation efficiency function;

[0014] S140, Hydraulic system modeling: Establish the coolant flow rate-pump pressure relationship equation;

[0015] S150, Integration of Multi-Field Coupled Equations: Constructing Unified State-Space Equations;

[0016] S160, parameter normalization processing: realize dimensionless transformation of multiple physical quantities.

[0017] Furthermore, in S150, the state space equation is as follows:

[0018]

[0019] Among them A s represents the effective heat dissipation area, β represents the angle-pressure drop coefficient, J represents the moment of inertia, τ gen represents the generating torque, τ loss represents friction loss torque, V represents system volume, T blade represents the leaf temperature, P cool Indicates the cooling system pressure, ω rotor represents the rotor angular velocity, ΔT represents the temperature difference, represents the temperature Laplace term, h c represents the convective heat transfer coefficient, α pump Indicates the pump speed control coefficient, Q rated represents the rated volume flow rate, θ represents the heat sink inclination angle, c p represents the specific heat capacity of the material, ρ m represents the density of turbine material, k t Represents thermal conductivity.

[0020] Furthermore, the following steps are performed in the reinforcement learning decision module:

[0021] S210, State Space Construction: Defining the reinforcement learning observation space;

[0022] S220, Action Space Design: Defining the continuous control action space;

[0023] S230, Reward Function Construction: Design a multi-objective reward function;

[0024] S240, Network Architecture Design: Building a Physical-Information Fusion Network;

[0025] S250, Experience Replay Design: Constructing Priority Replay Buffer;

[0026] S260, Physical Constraint Embedding: Incorporating Thermodynamic Constraints into Policy Updates.

[0027] Furthermore, the physical information fusion network in S240 is as follows:

[0028] h s =LSTM(s t (1:12)

[0029]

[0030] Among them, LSTM represents processing of 12-dimensional time series state, GNN represents processing of graph structure data, MLP represents processing of dynamic changes in device constraints, and D c represents the device constraint vector, Represents vector concatenation operation, h s Represents the hidden state of LSTM output, h f Represents the features of GNN output, h c represents the constrained features output by MLP, h represents the final fusion features, and LayerNorm represents the layer normalization operation.

[0031] Furthermore, the following steps are performed in the policy optimization training module:

[0032] S310, calculation of odds estimate: using generalized odds estimate;

[0033] S320, strategy loss construction: calculate the clip strategy gradient loss;

[0034] S330, value function optimization: minimize the MSE loss of the value network;

[0035] S340, Entropy Regularized Ensemble: Enhancing Policy Exploration Capabilities;

[0036] S350, Physical Constraint Integration: Integrating multi-physics constraints;

[0037] S360, adaptive learning rate adjustment: implements gradient normalization update;

[0038] S370, comprehensive objective function: combining all loss terms.

[0039] Furthermore, the integration results of multi-physics constraints are as follows:

[0040] L PHY=λ1||NS-Residual||2+λ2max(0,T blade -873) 2

[0041] Where λ1 represents the flow field constraint weight, λ2 represents the temperature constraint penalty coefficient, NS-Residual represents the flow field residual, and L PHY represents the physical constraint loss, T blade Indicates the leaf temperature.

[0042] Furthermore, the following steps are performed in the edge execution control module:

[0043] S410, model lightweight processing: deployment model quantization and pruning;

[0044] S420, delay compensation prediction: sensor-actuator delay compensation;

[0045] S430, Safety Barrier Design: Constructing Lyapunov Constraints;

[0046] S440, online adaptive fine-tuning: real-time policy parameter updates;

[0047] S450, Digital Twin Synchronization: Multiphysics Simulation State Calibration;

[0048] S460, Fault recovery mechanism: safe mode switching strategy.

[0049] Furthermore, in S450, the multi-physics simulation state is calibrated as follows:

[0050]

[0051] Among them F sim represents the simulated flow field data, f sync Indicates the synchronization frequency, F real Represents the actual flow field data, ∈ sync Represents the synchronization error threshold. When the residual exceeds the threshold, the co-simulation is triggered.

[0052] Furthermore, in S460, the security mode switching strategy is as follows:

[0053]

[0054] where σ thermal Indicates the thermal stress coefficient, PID parameters inherit the pump pressure control benchmark, π fallback represents the failure recovery strategy, π DRL represents the DRL control strategy, π PID represents the PID control strategy, represents the normalized temperature deviation.

[0055] The beneficial effects of the present invention are:

[0056] This invention effectively solves the problems of thermal management and power optimization of plateau turbine generators in extreme environments. Through intelligent control and multi-physical field collaboration, it significantly improves the temperature regulation accuracy and system response speed, achieves cooling energy consumption optimization and equipment loss reduction, greatly enhances the environmental adaptability and operational reliability of the unit, and through adaptive strategies, enables the turbine to maintain efficient power output under different altitudes and climatic conditions. Combined with multiple safety protection mechanisms, it provides a stable and reliable clean energy solution for high-altitude areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a structural block diagram of an active cooling and power optimization system for a plateau turbine generator proposed by the present invention;

[0058] Figure 2 is a table of operation results of the example proposed by the present invention;

[0059] In the figure: 101, multi-physics field modeling module; 102, reinforcement learning decision module; 103, strategy optimization training module; 104, edge execution control module. DETAILED DESCRIPTION

[0060] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. In addition, features described with respect to some examples may also be combined in other examples.

[0061] like Figure 1 As shown, an active cooling and power optimization system for a plateau turbine generator includes the following modules:

[0062] Multi-physics modeling module: real-time calculation of temperature field / flow field / stress field coupling state, providing dynamic characterization of the environment;

[0063] The module performs the following steps:

[0064] S100, multi-physics coupling modeling: Establishes a coupled dynamic model covering heat conduction, convection heat transfer, heat sink efficiency, and hydraulic systems. It constructs a unified state space through three-dimensional heat conduction equations, fluid mechanics equations, and equipment constraint equations to achieve normalization of multiple physical quantities.

[0065] In one embodiment of the present invention, the following steps are specifically included:

[0066] S110, Turbine Heat Conduction Modeling: Establishing three-dimensional unsteady heat conduction equations;

[0067]

[0068] where ρ m represents the density of turbine material, c p Indicates the specific heat capacity of the material, k t represents the thermal conductivity, q gen represents the heat production rate per unit volume, T represents temperature, t represents time, x represents the spatial coordinate, represents the Laplace operator;

[0069] S120, Convective Heat Transfer Modeling: Constructing the coupled coolant-air two-phase flow equations;

[0070]

[0071] Nu=C·Re m ·Pr n

[0072]

[0073] where h c represents the convective heat transfer coefficient, Nu represents the Nusselt number, Re represents the Reynolds number, Pr represents the Prandtl number, k f represents the thermal conductivity of the fluid, C, m, n represent the empirical coefficients, ρ f represents the fluid density, v represents the fluid velocity, L represents the characteristic length, μ represents the dynamic viscosity, c p,f represents the specific heat capacity of the fluid;

[0074] S130, heat sink efficiency modeling: establish the heat sink angle-heat dissipation efficiency function;

[0075]

[0076] where m f Indicates the heat sink characteristic parameters, L f Indicates the heat sink length, δ f represents the thickness of the heat sink, θ represents the tilt angle of the heat sink, η fin Indicates the heat sink efficiency, h c represents the convective heat transfer coefficient, k f represents the thermal conductivity of the fluid;

[0077] S140, Hydraulic system modeling: Establish the coolant flow rate-pump pressure relationship equation;

[0078]

[0079] where αpump represents the pump speed control coefficient, α pump ∈[0,1],ρ l Indicates coolant density, Q rated Indicates the rated volume flow rate, A p represents the cross-sectional area of ​​the pump port, f represents the friction coefficient of the pipeline, L p represents the length of the pipe, D represents the diameter of the pipe, v represents the fluid velocity, ΔP pump Indicates pump pressure difference;

[0080] S150, Integration of Multi-Field Coupled Equations: Constructing Unified State-Space Equations;

[0081]

[0082] Among them A s represents the effective heat dissipation area, β represents the angle-pressure drop coefficient, J represents the moment of inertia, τ gen represents the generating torque, τ loss represents friction loss torque, V represents system volume, T blade represents the leaf temperature, P cool Indicates the cooling system pressure, ω rotor represents the rotor angular velocity, ΔT represents the temperature difference, represents the temperature Laplace term;

[0083] S160, parameter normalization processing: realize dimensionless multi-physical quantities;

[0084]

[0085] Where T min Indicates the minimum operating temperature, T min =263K,T max Indicates the maximum operating temperature, T max =923K,P atm represents the local atmospheric pressure, ω rated Indicates the rated speed, represents the normalized temperature, represents the normalized pressure, represents the normalized angular velocity, T represents the temperature, P represents the pressure, and ω represents the angular velocity;

[0086] Reinforcement learning decision module: processes multi-source sensor data and generates coolant flow rate and heat sink angle control instructions;

[0087] This module performs the following steps:

[0088] S200, DRL framework construction: Design the observation and action space for reinforcement learning, build a hybrid network architecture that integrates LSTM time series processing and GNN flow field analysis, develop a multi-objective reward function and integrate physical constraints, establish a priority experience replay mechanism, and form an environmental perception-decision-making closed loop;

[0089] In one embodiment of the present invention, the following steps are specifically included:

[0090] S210, State Space Construction: Defining the reinforcement learning observation space;

[0091]

[0092] in represents the normalized leaf temperature, η fin Indicates the heat sink efficiency, ΔP pump Indicates the pump pressure difference, represents the normalized flow field characteristics, σ thermal represents thermal stress, s t represents the state vector at time t, where t represents the time step;

[0093] S220, Action Space Design: Defining the continuous control action space;

[0094] a t =[α pump ,θ fin ] T

[0095] obey:

[0096]

[0097] θ fin =15°+60°·sigmoid(μ θ )

[0098] where μ α Represents the mean value of the pump speed control quantity, μ α ∈[0,1],σ α represents the standard deviation of the exploration noise, μ θ Logistic parameter representing the heat sink angle, a t represents the action vector at time t, α pump represents the pump speed control coefficient, θ fin represents the heat sink angle, sigmoid represents the sigmoid activation function;

[0099] S230, Reward Function Construction: Design a multi-objective reward function;

[0100]

[0101] in represents the normalized temperature deviation, ΔP pump represents the pump pressure difference, w1 represents the temperature control weight, w1=0.5, w2 represents the energy consumption weight, w2=0.2, w3 represents the safety constraint weight, w3=0.25, w4 represents the action smoothing weight, w4=0.05, k1 represents the temperature attenuation coefficient, k1=6.8, r t represents the reward value at time t, represents the thermal stress constraint indicator function, ||Δθ fin ||2 represents the L2 norm of the heat sink angle change;

[0102] S240, Network Architecture Design: Building a Physical-Information Fusion Network;

[0103] h s =LSTM(s t (1:12)

[0104]

[0105] Among them, LSTM represents processing of 12-dimensional time series state (3 frames of history × 4 parameters), GNN represents processing of graph structure data, and MLP represents processing of dynamic changes in device constraints. c represents the device constraint vector, Represents vector concatenation operation, h s Represents the hidden state of LSTM output, h f Represents the features of GNN output, h c Represents the constrained features output by MLP, h represents the final fusion features, and LayerNorm represents the layer normalization operation;

[0106] S250, Experience Replay Design: Constructing Priority Replay Buffer;

[0107] p i =|δ i |+∈

[0108] in:

[0109] δ i =r i +γV φ (s i+1 )-V φ (s i )

[0110] Sampling probability:

[0111]

[0112] Importance Weight:

[0113]

[0114] Where ∈ represents a small positive number to prevent the priority from being zero, α represents the priority index, β represents the importance sampling index, N represents the playback buffer size, and p i represents the priority of the i-th sample, δ i represents the TD error, r i represents the immediate reward, γ represents the discount factor, V φ represents the value function, s i represents the state, P(i) represents the sampling probability, w i Indicates importance weight;

[0115] S260, Physical Constraint Embedding: Incorporating Thermodynamic Constraints into Policy Updates;

[0116] L total =L CLIP +λ1||NS-Residual||2+λ2max(0,T blade -873) 2

[0117] Where NS-Residual represents the residual of the Navier-Stokes equation, λ1 represents the flow field constraint weight, λ1=0.1, λ2 represents the temperature constraint penalty coefficient, λ2=0.3, L CLIP represents the clip loss of PPO, L total Represents the total loss function, T blade Indicates the leaf temperature;

[0118] Strategy optimization training module: It performs iterative optimization of strategies in a digital twin environment to ensure the physical rationality and energy efficiency optimization of the control algorithm;

[0119] This module performs the following steps:

[0120] S300, PPO policy optimization: uses generalized advantage estimation and the clip policy gradient method, combined with value function optimization and entropy regularization technology, to embed physical constraints such as the Navier-Stokes equation residual into the loss function, and achieves efficient training of the policy network through adaptive learning rate adjustment;

[0121] In one embodiment of the present invention, the following steps are specifically included:

[0122] S310, calculation of odds estimate: using generalized odds estimate (GAE);

[0123] δ t =r t +γ ppo Vφ (s t+1 )-V φ (s t )

[0124]

[0125] where γ ppo represents the discount factor, γ ppo =0.99,λ gae represents the GAE smoothing coefficient, λ gae =0.95,δ t represents the TD error, r t Indicates immediate reward, V φ represents the value function, s t Indicates status, represents the generalized advantage estimate, T represents the trajectory length;

[0126] S320, strategy loss construction: calculate the clip strategy gradient loss;

[0127]

[0128] where ∈ clip represents the strategy update range limit, ∈ clip =0.2,π θ represents the action distribution, θ represents the policy network parameters, θ old represents the old strategy network parameters, a t Indicates action, s t Indicates status, represents the advantage estimate, L CLIP Represents the clip loss function;

[0129] S330, value function optimization: minimize the MSE loss of the value network;

[0130]

[0131] in:

[0132]

[0133] Among them L VF represents the value function loss, φ represents the value network parameter, V φ represents the value function, s t Indicates status, represents the target value, represents the advantage estimate, φ old represents the old value network parameters;

[0134] S340, Entropy Regularized Ensemble: Enhancing Policy Exploration Capabilities;

[0135]

[0136] where β ent represents the entropy coefficient (adaptive adjustment), H represents the Shannon entropy of the strategy distribution, L ENT represents the entropy regularization loss, π θ represents the strategy distribution, s t Indicates status;

[0137] S350, Physical Constraint Integration: Integrating multi-physics constraints;

[0138] L PHY =λ1||NS-Residual||2+λ2max(0,T blade -873) 2

[0139] Where λ1 represents the flow field constraint weight, λ2 represents the temperature constraint penalty coefficient, NS-Residual represents the flow field residual, and L PHY represents the physical constraint loss, T blade Indicates the leaf temperature;

[0140] S360, adaptive learning rate adjustment: implements gradient normalization update;

[0141]

[0142] where η actor represents the policy network learning rate, η critic represents the value network learning rate, ∈ adam represents the Adam optimizer stabilization term, θ new represents the updated policy parameters, φ new Represents the updated value network parameters, L total Represents the total loss function, L VF represents the value function loss, represents the gradient operator;

[0143] S370, comprehensive objective function: combining all loss terms;

[0144] L total =L CLIP +c1L VF -c2L ENT +L PHY

[0145] Where c1 represents the value loss weight, c2 represents the entropy regularization weight, and L total Represents the total loss function, L CLIP represents the strategy clip loss, L VFRepresents the value function loss, L ENT represents the entropy regularization loss, L PHY represents the loss of physical constraints;

[0146] Edge execution control module: deploys a lightweight policy network to achieve sub-second real-time response and fault safety protection;

[0147] This module performs the following steps:

[0148] S400, online deployment optimization: Quantize and prune the training model for lightweight processing, design delay compensation prediction and Lyapunov safety barriers, establish a digital twin online calibration mechanism, develop a fast fault switching system, and achieve real-time and reliable operation of strategies on edge computing nodes;

[0149] In one embodiment of the present invention, the following steps are specifically included:

[0150] S410, model lightweight processing: deployment model quantization and pruning;

[0151] W q =Quantize(W θ ,b)

[0152] ||W θ || prune =p prune ||W θ ||0

[0153] Where W θ represents the policy network weight, b represents the quantization bit width, p prune Indicates the pruning ratio, W q represents the quantized weight, ||W θ ||0 represents the L0 norm of the original weight, ||W θ || prune Indicates the number of weights after pruning;

[0154] S420, delay compensation prediction: sensor-actuator delay compensation;

[0155]

[0156] Where k represents the compensation of 3-step delay, F norm represents the normalized flow field characteristics, represents the k-step prediction state, s t Represents the state, KF-Predict represents the Kalman filter prediction function, Represents the second-order derivative of flow field characteristics;

[0157] S430, Safety Barrier Design: Constructing Lyapunov Constraints;

[0158]

[0159] where Δω rotor Indicates the rotor angular velocity deviation, V(s t ) represents the Lyapunov function, represents the derivative of the Lyapunov function, k v represents the Lyapunov stability coefficient, represents the normalized temperature deviation, α pump Indicates the constraint range of the pump speed control coefficient;

[0160] S440, online adaptive fine-tuning: real-time policy parameter updates;

[0161]

[0162] where η online represents the online learning rate, represents the sliding average reward, Δθ represents the parameter update amount, π θ represents the strategy distribution, a t Indicates action, Represents the predicted state, represents the target value, L represents the sliding window length;

[0163] S450, Digital Twin Synchronization: Multiphysics Simulation State Calibration;

[0164]

[0165] Among them F sim represents the simulated flow field data, f sync Indicates the synchronization frequency, F real Represents the actual flow field data, ∈ sync Represents the synchronization error threshold, which triggers the joint simulation when the residual exceeds the threshold;

[0166] S460, Fault Recovery Mechanism: Safe Mode Switching Strategy;

[0167]

[0168] where σ thermal Indicates the thermal stress coefficient, PID parameters inherit the pump pressure control benchmark, π fallback represents the failure recovery strategy, π DRL represents the DRL control strategy, π PID represents the PID control strategy, represents the normalized temperature deviation.

[0169] In one embodiment of the present invention, in the plateau turbine generator fault recovery strategy, the core definitions of the two control strategies are as follows:

[0170] 1.DRL control strategy: A multivariable adaptive control strategy based on deep reinforcement learning training;

[0171] Control logic:

[0172] Real-time fusion of multi-dimensional sensor data such as temperature, pressure, and speed;

[0173] Dynamically calculate coolant pump speed and radiator angle through neural network;

[0174] Adjust control parameters in advance based on the flow field characteristics prediction results;

[0175] Achieve optimal energy efficiency while ensuring thermal safety;

[0176] 2.PID control strategy: a single variable feedback control strategy with a preset parameter set;

[0177] Control logic:

[0178] Only proportional-integral-differential operations are performed based on the blade temperature deviation;

[0179] Use fixed parameters (e.g., Kp = 2.5, Ki = 0.3, Kd = 0.8);

[0180] Output linear regulation of coolant flow;

[0181] Prioritize ensuring that temperatures do not exceed safety thresholds;

[0182] The switching logic is:

[0183] Enable DRL mode: Activate when temperature fluctuation is ≤10% and thermal stress coefficient is <0.85, giving full play to its multi-objective optimization advantages;

[0184] Forced PID switching: When a severe environmental disturbance (such as a gust of wind speed > 20m / s) or sensor abnormality is detected, the system switches to a reliable but inefficient baseline control.

[0185] The two strategies complement each other: the DRL strategy improves energy efficiency by 15-30% under normal conditions, and the PID strategy serves as a safety redundancy to ensure basic operational reliability under extreme working conditions.

[0186] Based on the above-mentioned active cooling and power optimization system of plateau turbine generator, an operation example of the intelligent cooling system of plateau turbine generator is given:

[0187] The scenario is as follows: At a wind farm at an altitude of 4,800 meters, with an ambient pressure of 58 kPa, a sudden force 8 gust (wind speed 22 m / s) causes the following anomaly in turbine unit #5:

[0188] The blade temperature suddenly increased from 620K to 805K within 30 seconds;

[0189] The coolant pump reaches 92% of its rated power;

[0190] The heat sink angle continues to oscillate at high frequency;

[0191] System response process:

[0192] Phase 1, multi-physics state perception:

[0193] The modeling module detects a sudden 40% increase in the flow field Reynolds number and triggers the emergency sampling mode (1kHz);

[0194] The digital twin predicts that the temperature will exceed the safety threshold (873K) in 5 seconds;

[0195] Phase 2, DRL decision generation:

[0196] The policy network generates control instructions within 28ms:

[0197] Pump speed is reduced from 0.75 to 0.62 to relieve hydraulic system pressure;

[0198] The heat sink angle is adjusted from 55° to 68° to enhance convection;

[0199] Open the standby cooling circuit valve in advance;

[0200] Phase 3, online security protection:

[0201] If the rotor acceleration exceeds the limit, it will automatically inject ±3° angle disturbance to suppress oscillation;

[0202] When the synchronization error of the digital twin exceeds the threshold, the auxiliary laser depth sounder calibration is initiated;

[0203] Phase 4, fault recovery:

[0204] Temperature fluctuations of up to 12% trigger the protection mechanism:

[0205] Keeping the DRL strategy running for the first 30 seconds reduces cooling energy consumption by 19%;

[0206] At the 31st second, the system switches to PID control and the flow rate increases to 95% of the rated value;

[0207] After 45 seconds, the temperature drops back to 798K and automatically switches back to DRL mode;

[0208] In this example, the running results are as follows Figure 2As shown in the figure, unplanned shutdowns caused by gusts can be successfully avoided, economic losses can be reduced compared with traditional control methods, and the effectiveness of intelligent control strategies in extreme environments can be verified.

[0209] The above describes the embodiments of the present invention, but the present invention is not limited to the above specific implementation methods. The above specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make many forms under the guidance of the present invention, all of which are protected by the present invention.

Claims

1. An active cooling and power optimization system for a plateau turbine generator, characterized in that: include: Multi-physics modeling module: Establishes a coupled dynamic model covering heat conduction, convection heat transfer, heat sink efficiency, and hydraulic systems. It constructs a unified state space through three-dimensional heat conduction equations, fluid mechanics equations, and equipment constraint equations to achieve normalization of multiple physical quantities. Reinforcement Learning Decision Module: Design the observation and action space for reinforcement learning, build a hybrid network architecture that integrates LSTM time series processing and GNN flow field analysis, develop a multi-objective reward function and integrate physical constraints, establish a priority experience replay mechanism, and form an environmental perception-decision-making closed loop; Policy Optimization Training Module: This module uses generalized advantage estimation and the clip policy gradient method, combined with value function optimization and entropy regularization techniques, to embed the physical constraints of the Navier-Stokes equation residual into the loss function. This module achieves efficient training of the policy network through adaptive learning rate adjustment. Edge execution control module: Quantize and pruned the training model for lightweight processing, designed delay compensation prediction and Lyapunov safety barriers, established a digital twin online calibration mechanism, developed a fault fast switching system, and realized real-time and reliable operation of the strategy on the edge computing node.

2. The active cooling and power optimization system for a plateau turbine generator according to claim 1, characterized in that: Perform the following steps in the Multiphysics Modeling Module: S110, Turbine Heat Conduction Modeling: Establishing three-dimensional unsteady heat conduction equations; S120, Convective Heat Transfer Modeling: Constructing the coupled coolant-air two-phase flow equations; S130, heat sink efficiency modeling: establish the heat sink angle-heat dissipation efficiency function; S140, Hydraulic system modeling: Establish the coolant flow rate-pump pressure relationship equation; S150, Integration of Multi-Field Coupled Equations: Constructing Unified State-Space Equations; S160, parameter normalization processing: realize dimensionless transformation of multiple physical quantities.

3. The active cooling and power optimization system for a plateau turbine generator according to claim 2, characterized in that: The state space equation in S150 is as follows: Among them A s represents the effective heat dissipation area, β represents the angle-pressure drop coefficient, J represents the moment of inertia, τ gen represents the generating torque, τ loss represents friction loss torque, V represents system volume, T blade represents the leaf temperature, P cool Indicates the cooling system pressure, ω rotor represents the rotor angular velocity, ΔT represents the temperature difference, represents the temperature Laplace term, h c represents the convective heat transfer coefficient, α pump Indicates the pump speed control coefficient, Q rated represents the rated volume flow rate, θ represents the heat sink inclination angle, c p represents the specific heat capacity of the material, ρ m represents the density of turbine material, k t Represents thermal conductivity.

4. The active cooling and power optimization system for a plateau turbine generator according to claim 3, characterized in that: Perform the following steps in the reinforcement learning decision module: S210, State Space Construction: Defining the reinforcement learning observation space; S220, Action Space Design: Defining the continuous control action space; S230, Reward Function Construction: Design a multi-objective reward function; S240, Network Architecture Design: Building a Physical-Information Fusion Network; S250, Experience Replay Design: Constructing Priority Replay Buffer; S260, Physical Constraint Embedding: Incorporating Thermodynamic Constraints into Policy Updates.

5. The active cooling and power optimization system for a plateau turbine generator according to claim 4, characterized in that: The physical information fusion network in S240 is as follows: h s =LSTM(s t (1:12)) Among them, LSTM represents processing of 12-dimensional time series state, GNN represents processing of graph structure data, MLP represents processing of dynamic changes in device constraints, and D c represents the device constraint vector, Represents vector concatenation operation, h s Represents the hidden state of LSTM output, h f Represents the features of GNN output, h c represents the constrained features output by MLP, h represents the final fusion features, and LayerNorm represents the layer normalization operation.

6. The active cooling and power optimization system for a plateau turbine generator according to claim 5, characterized in that: In the policy optimization training module, perform the following steps: S310, calculation of odds estimate: using generalized odds estimate; S320, strategy loss construction: calculate the clip strategy gradient loss; S330, value function optimization: minimize the MSE loss of the value network; S340, Entropy Regularized Ensemble: Enhancing Policy Exploration Capabilities; S350, Physical Constraint Integration: Integrating multi-physics constraints; S360, adaptive learning rate adjustment: implements gradient normalization update; S370, comprehensive objective function: combining all loss terms.

7. The active cooling and power optimization system for a plateau turbine generator according to claim 6, characterized in that: The integration of the multiphysics constraints results in the following: L PHY =λ1∥NS-Residual∥2+λ2max(0,T blade -873) 2 Where λ1 represents the flow field constraint weight, λ2 represents the temperature constraint penalty coefficient, NS-Residual represents the flow field residual, and L PHY represents the physical constraint loss, T blade Indicates the leaf temperature.

8. The active cooling and power optimization system for a plateau turbine generator according to claim 7, characterized in that: Perform the following steps in the Edge Execution Control module: S410, model lightweight processing: deployment model quantization and pruning; S420, delay compensation prediction: sensor-actuator delay compensation; S430, Safety Barrier Design: Constructing Lyapunov Constraints; S440, online adaptive fine-tuning: real-time policy parameter updates; S450, Digital Twin Synchronization: Multiphysics Simulation State Calibration; S460, Fault recovery mechanism: safe mode switching strategy.

9. The active cooling and power optimization system for a plateau turbine generator according to claim 8, characterized in that: The multi-physics simulation state calibration in S450 is as follows: Among them F sim represents the simulated flow field data, f sync Indicates the synchronization frequency, F real Represents the actual flow field data, ∈ sync Represents the synchronization error threshold. When the residual exceeds the threshold, the co-simulation is triggered.

10. The active cooling and power optimization system for a plateau turbine generator according to claim 9, characterized in that: In S460, the security mode switching strategy is as follows: where σ thermal Indicates the thermal stress coefficient, PID parameters inherit the pump pressure control benchmark, π fallback represents the failure recovery strategy, π DRL represents the DRL control strategy, π PID represents the PID control strategy, represents the normalized temperature deviation.

Images