An intelligent control method and device for a liquid natural gas cold energy power generation system

Abstract

This invention discloses an intelligent control method and device for liquefied natural gas (LNG) cold power generation systems. The method includes determining key control variables and controlled variables; constructing an identification model of the LNG cold power generation system based on the key control variables and controlled variables; establishing a composite reward function with cold power generation efficiency, power generation benefits, and system energy consumption as its core; training the constructed SAC (System-Agent Controlled) agent based on the composite reward function to obtain a trained SAC agent; adding a first-order low-pass filter to the output of the trained SAC agent for filtering; the identification model outputting the current state variable based on the filtered action; and the trained SAC agent outputting the optimal control action for the LNG cold power generation system based on the current state variable. This invention can effectively improve the control accuracy and cold power generation efficiency of LNG cold power generation systems.

Description

Technical Field

[0001] This invention relates to the field of industrial process control and artificial intelligence, and in particular to an intelligent control method and device for a liquefied natural gas cold energy power generation system. Background Technology

[0002] Liquefied natural gas (LNG) releases a large amount of high-grade cold energy during its reheating process. Utilizing the cold energy of LNG for power generation is an important way to improve overall energy efficiency. However, LNG cold energy power generation systems have strong coupling, large time lag, and nonlinear characteristics. Traditional LNG cold energy power generation systems often employ static rule control or model-based predictive control, which have the following drawbacks:

[0003] Strong fluctuations in operating conditions: Random changes in LNG feed flow rate, ambient temperature, and seawater temperature cause the system to deviate from the optimal design conditions for a long period of time.

[0004] Strong coupling of multiple variables: Parameters such as working fluid flow rate and pressure are mutually constrained, making it difficult for traditional PID (Proportional-Integral-Derivative) or control based on fixed rules to achieve global optimization.

[0005] Insufficient algorithmic exploration: Existing intelligent control algorithms have limited strategies in random environments and lack the ability to actively explore the optimal action space, which restricts further improvement in power generation efficiency. Summary of the Invention

[0006] This invention addresses the problems of low power generation efficiency and insufficient energy efficiency optimization in LNG cold power generation systems by proposing an intelligent control method and device for liquefied natural gas cold power generation systems. By establishing a system identification model, training a SAC (Soft Actor-Critic) intelligent agent and integrating it into the LNG cold power generation system, and combining it with a first-order low-pass filter system, the key controller settings of the LNG cold power generation system are optimized and adjusted to achieve optimization of power generation efficiency and effectiveness.

[0007] The present invention adopts the following technical solution.

[0008] A first aspect of the present invention provides an intelligent control method for a liquefied natural gas cold power generation system, comprising:

[0009] Identify the key control variables and controlled variables, and construct an identification model for the liquefied natural gas cold energy power generation system based on the key control variables and controlled variables;

[0010] Construct a SAC agent, establish a composite reward function with cold energy power generation efficiency, power generation benefits and system energy consumption as the core, and train the constructed SAC agent based on the composite reward function to obtain a trained SAC agent.

[0011] A first-order low-pass filter is added to the output of the trained SAC agent for filtering. The identification model outputs the current state variable based on the filtered action. The trained SAC agent outputs the optimal control action for the liquefied natural gas cold energy power generation system based on the current state variable.

[0012] Optionally, key control variables include LNG feed flow rate, seawater flow rate, working fluid temperature, and seawater inlet temperature, while controlled variables include power generation, natural gas outlet temperature, working fluid evaporation temperature, working fluid expansion temperature, working fluid condensation temperature, and seawater outlet temperature.

[0013] Optionally, the identification model for the liquefied natural gas (LNG) cold power generation system, based on key control variables and controlled variables, includes:

[0014] A dynamic mechanism model was constructed based on Aspen Dynamics to obtain step data. The dynamic mechanism model was used to simulate the response law of the controlled variable as the key control variable changes step.

[0015] The transfer function between the control variable and the controlled variable is obtained by fitting the step data. The relative gain matrix between the control variable and the controlled variable is calculated based on the gain of the transfer function.

[0016] The variables whose corresponding elements in the relative gain matrix are closest to 1 are selected for pairing. Based on the pairing results, the corresponding transfer function is selected, and a system identification model is constructed based on the selected transfer function.

[0017] Optionally, the relative gain matrix can be calculated based on the gain of the transfer function:

[0018] ,

[0019] in, Represents the relative gain matrix. The gain matrix represents the transfer function. Indicates matrix transpose. The Hadamard product of matrices is represented by the relative gain matrix, which is used to describe the strength of the interaction between input and output variables.

[0020] Optionally, the composite reward function during the SAC agent training process is as follows:

[0021]

[0022]

[0023]

[0024]

[0025] in, This represents the value of the composite reward function. , These represent the weights of power generation efficiency and system energy consumption, respectively. Indicates power generation efficiency. Indicates system energy consumption cost. Indicates a penalty item. Indicates power generation efficiency; The action range constraint of the i-th action output by the SAC agent at time t is the action range constraint term of the reward function, which is used to constrain the range of the actions output by the SAC agent. Let represent the i-th action output by the SAC agent at time t. Indicates the weight of the action penalty. , These represent the lower and upper limits of the given action range, respectively. This represents the state range constraint term of the composite reward function. This represents the j-th state of the system at time t. Indicates the state penalty weight. , Let A and S represent the lower and upper limits of the given state range, respectively. Let A represent the action space and S represent the state space.

[0026] Optionally, the action space of the SAC agent includes LNG feed flow rate, seawater flow rate, and working fluid temperature;

[0027] The state space of the SAC agent includes the natural gas outlet temperature, working fluid evaporation temperature, working fluid expansion temperature, working fluid condensation temperature, and seawater outlet temperature.

[0028] Optionally, the seawater inlet temperature can be used as a perturbation and randomly added during preset training rounds or testing phases in the SAC agent training.

[0029] Optionally, the relative gain matrix between the control variable and the controlled variable is calculated based on the gain of the transfer function, including:

[0030] Construct the gain matrix based on the gain of the transfer function;

[0031] Based on the importance of the controlled variable to power generation, the gain matrix is ​​split into two local linear gain matrices;

[0032] Calculate the relative gain matrix of the two local linear gain matrices respectively.

[0033] Alternatively, the transfer function model is as follows:

[0034] ,

[0035] in, For the transfer function model, For Laplace factor, - These are the pole coefficients. - Zero-point coefficient, For pure time delay, This represents the Laplace transform value of the controlled variable. This represents the Laplace transform value of the key control variable, where m represents the highest order of the denominator polynomial and n represents the highest order of the numerator polynomial.

[0036] A second aspect of the present invention provides an intelligent control device for a liquefied natural gas (LNG) cold power generation system, implementing the aforementioned intelligent control method for an LNG cold power generation system, the device comprising:

[0037] The modeling module is used to determine the key control variables and controlled variables, and to build an identification model of the liquefied natural gas cold energy power generation system based on the key control variables and controlled variables.

[0038] The agent training module is used to construct the SAC agent, establish a composite reward function with cold energy power generation efficiency, power generation benefits and system energy consumption as the core, and train the constructed SAC agent based on the composite reward function to obtain the trained SAC agent.

[0039] The control module is used to add a first-order low-pass filter to the output of the trained SAC agent for filtering. The identification model outputs the current state variable based on the filtered action. The trained SAC agent outputs the optimal control action for the liquefied natural gas cold energy power generation system based on the current state variable.

[0040] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0041] This invention, based on dynamic operating data of an LNG cold power generation system, establishes a system identification model through system identification. Building upon this model, a composite reward function is designed, focusing on cold power generation efficiency, power generation benefits, and system energy consumption. The SAC algorithm is used to train the agent, and a first-order low-pass filter is added after training to smooth control commands. This invention effectively improves the control accuracy and cold power generation efficiency of the LNG cold power generation system, stabilizes power output, reduces system energy consumption, mitigates fluctuations in the agent's output signal, and enhances its adaptability to disturbances such as LNG flow fluctuations and ambient temperature changes, making it suitable for industrial applications.

[0042] The beneficial effects of the present invention also include:

[0043] (1) The present invention uses the dynamic data of the device to build a system identification model, which can reflect the real operation of the LNG cold energy power generation system with large time delay and strong nonlinearity.

[0044] (2) This invention uses the SAC algorithm based on stochastic policy gradient as the training algorithm for the agent. It utilizes the policy entropy maximization to guide the agent to generate more exploratory behaviors. By balancing expected reward and policy entropy, a better balance is achieved between exploration and utilization, thus accelerating the agent training process.

[0045] (3) The intelligent control method of the present invention, based on the real-time status of the LNG cold energy power generation system collected and combined with the trained strategy model, feeds back to the intelligent agent based on the composite reward function to start the next decision cycle.

[0046] (4) The present invention uses a first-order low-pass filter to process the output signal of the intelligent agent, remove fluctuations and noise, and maintain the stable operation of the LNG cold energy power generation system. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0048] Figure 1 This is a schematic diagram of the intelligent control method of applying the SAC intelligent agent of the present invention to an LNG cold energy power generation system;

[0049] Figure 2 This is a schematic diagram of the system identification model structure of the present invention;

[0050] Figure 3 This is a structural diagram of the SAC algorithm of the present invention;

[0051] Figure 4 This is a diagram illustrating the SAC agent training process of the present invention.

[0052] Figure 5 This is a comparison chart of parameters before and after adding intelligent control in this invention;

[0053] Figure 6 This is a comparison diagram before and after adding intelligent control and a first-order low-pass filter to the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are only some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0055] Reference Figure 1 This invention provides an intelligent control method for a liquefied natural gas cold energy power generation system, comprising the following steps:

[0056] Step 1: Determine the key control variables and controlled variables, and construct an identification model of the liquefied natural gas cold energy power generation system based on the key control variables and controlled variables.

[0057] A dynamic mechanism model of the LNG cold power generation system was constructed using Aspen Dynamics software to obtain key parameter variables. The acquired data was then used to obtain a transfer function model using a system identification tool. Based on this transfer function model, an identification model of the LNG cold power generation system was constructed. Step 1 specifically includes:

[0058] Step 1.1: Construct a dynamic mechanism model based on Aspen Dynamics to obtain step data. The dynamic mechanism model is used to simulate the response law of the controlled variable as the key control variable changes step.

[0059] Dynamic Mechanism Modeling: A dynamic mechanism model based on Aspen Dynamics is constructed to simulate disturbances such as LNG feed flow, seawater flow fluctuations, and step changes in working fluid temperature, revealing the system's dynamic response patterns. Multi-dimensional time-series data, including working fluid temperature, pressure, flow rate, and power generation, are collected across all operating conditions. Outliers are then removed, noise interference is eliminated using moving average filtering, and dimensional differences are processed through Min-Max normalization to form a standardized modeling dataset. Finally, a high-precision identification model is constructed using a system identification tool. During runtime, a pre-trained SAC agent is embedded into the system identification model to achieve higher power generation and efficiency.

[0060] In the system identification model modeling, the data used is the sampled data from the dynamic mechanism model of the LNG cold energy power generation system, including the operating data of the key control variables and controlled variables at the corresponding time.

[0061] Specifically, the process flow analysis of the LNG cold energy power generation system is conducted to determine the control variables and controlled variables required for modeling.

[0062] The power generation system in this embodiment adopts the existing combined power generation method, which has two expanders. The specific methods and processes are existing technologies, and the dynamic mechanism model built based on Aspen Dynamics is also an existing technology, which will not be described in detail here.

[0063] The control variables for the LNG cold power generation system are: LNG feed flow rate, seawater flow rate, working fluid temperature, and seawater inlet temperature. The controlled variables are: power generation, natural gas outlet temperature, working fluid expansion temperature, working fluid condensation temperature, and seawater outlet temperature. After determining the variables, the output data of the controlled variables when one control variable experiences a step change while the other key control variables remain unchanged are collected from the dynamic mechanism model as step data. All step data are preprocessed and saved for later use.

[0064] Referring to Table 1, Table 1 shows the model identification results of the LNG cold energy power generation system of the present invention after system identification.

[0065] Table 1

[0066]

[0067] In Table 1, This represents the transfer function with key control variable i and controlled variable j. , G1j, G2j, G3j, and G4j represent the LNG feed flow rate, seawater flow rate, working fluid temperature, and seawater inlet temperature, respectively. Gi1 represents the power generation of expander A (1), Gi2 represents the power generation of expander B (2), Gi3 represents the natural gas outlet temperature, Gi4 represents the working fluid evaporation temperature, Gi5 represents the working fluid expansion temperature, Gi6 represents the working fluid condensation temperature, Gi7 represents the seawater outlet temperature, G18 represents the transfer function of the LNG inlet valve opening on the LNG feed flow rate, G28 represents the transfer function of the seawater inlet valve opening on the seawater flow rate, and G38 represents the transfer function of the working fluid valve opening on the working fluid temperature. These transfer functions constitute a transfer function matrix. - Represents the pole coefficient. - This represents the zero-point coefficient, and Fit Percent represents the percentage of fit.

[0068] A transfer function is established by fitting step data of the control and controlled variables. The gain between the control and controlled variables is then calculated using the transfer function, and the relative gain matrix G is obtained. A and G B This allows us to determine the control variables corresponding to each controlled variable, and ultimately build a complete system identification model.

[0069] Step 1.2: Based on the step data, fit to obtain the transfer function between the control variable and the controlled variable, and calculate the relative gain matrix between the control variable and the controlled variable according to the gain of the transfer function.

[0070] The ratio of the step change of the key control variable to the final steady-state change of the controlled variable is used as the gain of the transfer function. The gains of each transfer function are arranged in a rectangular form to form a gain matrix, which is a 4×7 dimensional matrix. The gain quantifies the static coupling relationship between the key control variable and the controlled variable in steady state, providing key parameters for subsequent relative gain matrix solving and optimal variable pairing.

[0071] The dynamic model of the SAC agent training environment is a transfer function model of the system input and output, as follows:

[0072] ,

[0073] in, For Laplace factor, - These are the pole coefficients. - Zero-point coefficient, For pure time delay, This represents the Laplace transform value of the controlled variable. The Laplace transform value of the key control variable is represented by m, which represents the highest order of the denominator polynomial, and n represents the highest order of the numerator polynomial. The coefficients of the numerator and denominator polynomials together determine the dynamic characteristics of the system.

[0074] The relative gain array (RGA) method is used to quantitatively analyze the coupling characteristics of the system. The formula for calculating the RGA matrix Λ is:

[0075]

[0076] in, Represents the relative gain matrix. The gain matrix represents the transfer function. Λ denotes the matrix transpose, and ⊗ denotes the Hadamard product (one-to-one multiplication) of the matrix. The calculated Λ matrix can clearly reveal the strength of the interaction between the input and output variables.

[0077] Step 1.3: Select the variables whose corresponding elements in the relative gain matrix are closest to 1 for pairing, select the corresponding transfer function based on the pairing results, and construct the identification model of the liquefied natural gas cold energy power generation system.

[0078] The coupling relationship between variables was analyzed using the relative gain matrix method, control structure pairing was completed, and the complex nonlinear model was reduced to a lower order transfer function model for controller design.

[0079] It should be noted that the controllers include LNG flow controllers, seawater flow controllers, and working fluid temperature controllers.

[0080] Since the gain matrix is ​​a 4×7 matrix, in order to apply the RGA method to calculate the relative gain matrix, the non-square matrix is ​​converted into a square matrix. Therefore, the gain matrix is ​​split into two 4×4 square matrices and their respective relative gain matrices are calculated.

[0081] Furthermore, the gain matrix is ​​split according to the importance of the controlled variable to power generation.

[0082] In this embodiment, since the system has two expanders for power generation, the working fluid evaporation temperature directly affects the power generation of the expanders. Therefore, the working fluid evaporation temperature has the highest importance to the power generation. Thus, both local linear gain matrices must include the working fluid evaporation temperature during the decomposition, resulting in the following local linear gain matrices:

[0083]

[0084]

[0085] Among them, G A This describes the local linear gain relationships of four input perturbation variables (LNG feed flow rate, seawater flow rate, working fluid temperature, and seawater inlet temperature) with respect to four key output variables (power generation 1, power generation 2, natural gas outlet temperature, and working fluid evaporation temperature). Each element in the matrix represents the output change caused by a unit change in the corresponding input perturbation. The sign of the gain indicates the direction of action, and the absolute value indicates the sensitivity.

[0086] G B The local linear gain relationships of four input perturbation variables (LNG feed flow rate, seawater flow rate, working fluid temperature, and seawater inlet temperature) with respect to four key output variables (working fluid evaporation temperature, working fluid expansion temperature, working fluid condensation temperature, and seawater outlet temperature) are described.

[0087] Based on the identification results, corresponding controllers were added to the liquefied natural gas (LNG) cold power generation system, and an identification model of the LNG cold power generation system was constructed, such as... Figure 2As shown, PID1 represents the LNG feed flow controller, PID2 represents the seawater flow controller, PID3 represents the working fluid temperature controller, PID4 represents the seawater inlet temperature controller, V01, V02, V03, and V04 represent the valves of the corresponding control channels, M11, M12, and M13 represent the parameters of the transfer function model with LNG feed flow as input (power generation 2), natural gas outlet temperature, and working fluid condensation temperature as output, M21 and M22 represent the parameters of the transfer function model with working fluid temperature as input (working fluid evaporation temperature) and seawater outlet temperature as output, M31 and M32 represent the parameters of the transfer function model with operating temperature as input (power generation 1) and working fluid expansion temperature as output, and M41, M42, and M22 represent the parameters of the transfer function model with seawater inlet temperature as input (power generation 2), natural gas outlet temperature, and working fluid evaporation temperature as output.

[0088] Each control channel is designed based on a precisely established transfer function model. The identified model order covers second-order dynamic characteristics, fully reflecting the system's response behavior and complexity under different operating conditions. Through the above-mentioned reasonable configuration of the control structure, the system can effectively adjust key operating parameters, improving the stability and economy of the LNG cold energy power generation process. The system identification model provides a reliable simulation foundation for subsequent optimization of controller parameters, including but not limited to PID parameters, system performance evaluation, and engineering applications.

[0089] Step 2: Construct the SAC agent and establish a composite reward function with cold energy power generation efficiency, power generation benefits and system energy consumption as the core. Based on the composite reward function, use the deep stochastic policy gradient algorithm to train the constructed SAC agent to obtain the trained SAC agent.

[0090] like Figure 3 As shown, the training of the SAC agent includes the following three steps:

[0091] (a) Data acquisition and establishment of an experience pool: collecting environmental status (S) data of the LNG cold power generation system. t The Actor online policy network outputs an action (A) based on the collected environmental conditions. t The real-time action feedback to the LNG cold energy power generation system forms the next state S in the environment. t+1 And provide feedback rewards r t+1 This forms a complete experience and is stored in the experience pool.

[0092] The state space includes: natural gas outlet temperature, working fluid evaporation temperature, working fluid expansion temperature, working fluid condensation temperature, and seawater outlet temperature.

[0093] Action space: LNG feed flow rate, seawater flow rate, working fluid temperature.

[0094] (b) Learning and Exploration: Batch data is sampled from the experience pool, and two action value networks, critic Q1 and critic Q2, are trained in parallel. The action value networks are used to evaluate the value output Q of the action in the current state. The two networks calculate the temporal difference (TD) error respectively and update it by minimizing the mean square loss. To eliminate the Q-value overestimation bias, the target Q-value is determined by the minimum of the outputs of the two networks, and policy entropy is introduced into the target calculation to fit the maximum entropy framework.

[0095] (c) Policy Update: The Actor network optimizes policy parameters based on the updated Q-value and the principle of maximizing entropy to balance exploration and exploitation. The parameters of each online network are smoothly transferred to the corresponding target network through a soft update mechanism to ensure the stability of the training process.

[0096] When calculating the target Q value, the Actor target network generates action A' based on the next state S', and the Critic target network outputs the corresponding target Q value, which is used as the learning target to update the main Critic network, thereby effectively mitigating the training fluctuation problem caused by bootstrapping.

[0097] Through a three-step cycle, the SAC agent continuously explores and achieves higher power generation efficiency and effectiveness. The trained SAC agent is then deployed in the LNG cold energy power generation system for optimized system control.

[0098] Using the established system identification model as the training environment for the SAC agent, the reward function during the agent training process is as follows:

[0099]

[0100]

[0101]

[0102]

[0103] in, This represents the value of the composite reward function. , The weights for power generation efficiency and system energy consumption are respectively. For power generation efficiency, For system energy consumption cost, As a penalty item, To calculate the power generation efficiency, the power generation benefit is obtained by subtracting the pump power consumption from the generated power and then multiplying the result by the electricity price. The system energy consumption is the pump power consumption. A represents the action space, and S represents the state space.

[0104] It is the action range constraint of the i-th action output by the agent at time t, and is the action range constraint term of the reward function, used to constrain the range of the agent's output actions;

[0105] Let be the i-th action output by the agent at time t. It is the weight of action penalty. , These are the lower and upper limits of the given action range, respectively.

[0106] It is the state range constraint term of the reward function. Let j be the j-th state of the system at time t. It is the state penalty weight. , These represent the lower and upper limits of the given state range, respectively.

[0107] It should be noted that, to eliminate the impact of differences in physical dimensions on the weights of the reward function, this invention normalizes the power generation, energy consumption cost, efficiency, and deviation penalty terms, mapping all terms to the same dimensionless interval. The corrected reward function has a unified physical scale for each component, avoiding the difficulties in parameter tuning and training imbalance caused by differences in dimensions.

[0108] The constraints and penalty coefficients for action and state variables are shown in Table 2.

[0109] Table 2

[0110]

[0111] During the exploration phase, the agent has low basic rewards, and actions and states are prone to triggering constraint penalties, resulting in poor training effects. As the SAC agent continues to explore, its actions tend to stabilize, the basic rewards increase, and it eventually stabilizes.

[0112] The SAC algorithm uses policy entropy maximization to guide the agent to generate more exploratory behavior. By balancing the expected reward (i.e., the expectation of long-term cumulative reward) with policy entropy, it better balances exploration and utilization during the agent's training process, thereby achieving a higher reward function.

[0113] Furthermore, the seawater inlet temperature is used as a perturbation, which is randomly added during preset training rounds or testing phases in the SAC agent training. During SAC agent training, the seawater inlet temperature is selected as the perturbation source to simulate common upstream fluctuations in actual operation. The perturbation is set as a step change in the seawater inlet temperature at a preset temperature threshold. This perturbation is introduced at random time steps in specific training rounds or testing phases to prevent the agent from learning fixed control patterns, thereby enhancing its generalization ability.

[0114] Specifically, based on the established SAC agent training environment and reward function, and referring to Figure 3 The SAC agent was built and trained. Each training round lasted 3 hours. At the 1-hour mark, a 2°C decrease in seawater inlet temperature was added to test the agent's control performance under unsteady conditions. A total of 400 training rounds were conducted.

[0115] SAC agent training process (refer to) Figure 4 As can be seen, the cumulative reward and average reward gradually increase from round 1 to round 80. When the number of rounds reaches 100, the cumulative reward and average reward tend to stabilize, indicating that the learning effect of the agent is continuously improved during the training process and the strategy is continuously optimized.

[0116] Step 3: Add a first-order low-pass filter to the output of the trained SAC agent for filtering, and the identification model outputs the current environmental state variables based on the filtered actions; the trained SAC agent outputs the optimal control actions for the liquefied natural gas cold energy power generation system based on the current environmental state variables.

[0117] The SAC agent output action filtering system addresses the high-frequency noise issue in the output actions of SAC agents based on stochastic policy gradients by designing an action filter based on the principle of first-order low-pass filtering. This filter utilizes an inertial smoothing mechanism, weighting and fusing the new action at the current moment with the output action at the previous moment to suppress high-frequency fluctuations while preserving the trend component of the action. Its core lies in balancing the response speed and smoothness of the action by adjusting the filter coefficients, thereby reducing the influence of process noise and measurement noise while achieving a stable transition in the action output.

[0118] In the design of a first-order low-pass filter, the smoothness and response speed of the control signal are adjusted by the cutoff frequency or the time constant τ. A higher cutoff frequency and a smaller time constant τ allow the filter to pass more high-frequency components, resulting in weaker noise suppression and faster but more volatile output responses. Conversely, a lower cutoff frequency and a larger time constant τ attenuate high-frequency noise more effectively, leading to smoother output responses but slower dynamic responses, potentially lagging behind changes in the system's actual operating conditions. Only by appropriately selecting the cutoff frequency or time constant τ can the agent retain the dynamic characteristics of the control signal while filtering out useless high-frequency noise, thus achieving safe and stable system operation.

[0119] Specifically, after the SAC agent is trained, a first-order low-pass filter is designed with a time constant of 300 and a cutoff frequency of 0.53mHz.

[0120] The SAC agent's performance after training is as follows: Figure 5 As shown, where, as Figure 5 As shown in (a), regarding the reward function, the system reward is 29.6 without the agent, and increases to 31.2 after adding the SAC agent, a relative improvement of approximately 5.4%; Figure 5 As shown in (b), in terms of power generation efficiency, the efficiency is 29.5% without the addition of the agent, and increases to 31.5% after adding the SAC agent; as Figure 5 As shown in Figure (c), in terms of power generation, the power generation was 3475kW without the addition of the intelligent agent, and increased to 3701kW after adding the SAC intelligent agent, an increase of 226kW, which is approximately 6.5%. Overall, the SAC intelligent agent can effectively optimize the control strategy of the LNG cold energy power generation system, achieving stable improvements in reward function, power generation efficiency, and power generation.

[0121] Adding a first-order low-pass filter has the following effect: Figure 6 As shown, Figure 6 As shown in Figure (a), without a filter, the reward value fluctuates between 29.6 and 32.4. After adding a filter, the reward value stabilizes between 31.0 and 31.7, reducing the fluctuation range by 75.0%. Figure 6 As shown in (b), without a filter, the power generation efficiency fluctuated between 30.0% and 32.8%. After adding the filter, the efficiency stabilized between 31.2% and 31.7%, reducing the fluctuation range by 82.1%. Figure 6 As shown in (c), the power generation fluctuates between 3500kW and 3800kW. After adding the filter, the power generation stabilizes between 3700kW and 3755kW, reducing the fluctuation range by 81.7%. In summary, the fluctuation range of key parameters is effectively reduced after adding a first-order low-pass filter.

[0122] This invention establishes a system identification model for an LNG cold energy power generation system, and then sequentially builds a SAC intelligent agent real-time optimization control system and a first-order low-pass filter system. Based on existing PID control, this method improves the control effect of the device and increases the system's power generation efficiency and economic benefits.

[0123] Although the present invention has been illustrated and described with reference to preferred embodiments, those skilled in the art should understand that various changes and modifications can be made to the present invention without departing from the scope defined by the claims.

[0124] The present invention also provides an intelligent control device for a liquefied natural gas (LNG) cold power generation system, which operates the above-described intelligent control method for a LNG cold power generation system. The device includes:

[0125] The modeling module is used to determine the key control variables and controlled variables, and to build an identification model of the liquefied natural gas cold energy power generation system based on the key control variables and controlled variables.

[0126] The agent training module is used to construct the SAC agent, establish a composite reward function with cold energy power generation efficiency, power generation benefits and system energy consumption as the core, and train the constructed SAC agent based on the composite reward function to obtain the trained SAC agent.

[0127] The control module is used to add a first-order low-pass filter to the output of the trained SAC agent for filtering. The identification model outputs the current state variable based on the filtered action. The trained SAC agent outputs the optimal control action for the liquefied natural gas cold energy power generation system based on the current state variable.

[0128] Regarding the system in the above embodiments, the specific manner in which each unit performs operations has been described in detail in the embodiments related to the method, and will not be elaborated here.

[0129] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0130] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0131] All parts not covered in this invention are the same as or can be implemented using existing technologies.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A smart control method for a liquefied natural gas cold energy power generation system, characterized in that, include: Identify the key control variables and controlled variables, and construct an identification model for the liquefied natural gas cold energy power generation system based on the key control variables and controlled variables; Construct a SAC agent, establish a composite reward function with cold energy power generation efficiency, power generation benefits and system energy consumption as the core, and train the constructed SAC agent based on the composite reward function to obtain a trained SAC agent. A first-order low-pass filter is added to the output of the trained SAC agent for filtering. The identification model outputs the current state variable based on the filtered action. The trained SAC agent outputs the optimal control action for the liquefied natural gas cold energy power generation system based on the current state variable.

2. The intelligent control method for a liquefied natural gas cold power generation system according to claim 1, characterized in that, Key control variables include LNG feed flow rate, seawater flow rate, working fluid temperature, and seawater inlet temperature. Controlled variables include power generation, natural gas outlet temperature, working fluid evaporation temperature, working fluid expansion temperature, working fluid condensation temperature, and seawater outlet temperature.

3. The intelligent control method for a liquefied natural gas cold power generation system according to claim 1, characterized in that, The identification model for a liquefied natural gas (LNG) cold power generation system, constructed based on key control variables and controlled variables, includes: A dynamic mechanism model was constructed based on Aspen Dynamics to obtain step data. The dynamic mechanism model was used to simulate the response law of the controlled variable as the key control variable changes step. The transfer function between the control variable and the controlled variable is obtained by fitting the step data. The relative gain matrix between the control variable and the controlled variable is calculated based on the gain of the transfer function. The variables whose corresponding elements in the relative gain matrix are closest to 1 are selected for pairing. Based on the pairing results, the corresponding transfer function is selected, and a system identification model is constructed based on the selected transfer function.

4. The intelligent control method for a liquefied natural gas cold power generation system according to claim 3, characterized in that, Calculate the relative gain matrix based on the gain of the transfer function: ， in, Represents the relative gain matrix. The gain matrix represents the transfer function. Indicates matrix transpose. The Hadamard product of matrices is represented by the relative gain matrix, which is used to describe the strength of the interaction between input and output variables.

5. The intelligent control method for a liquefied natural gas cold power generation system according to claim 1, characterized in that, The composite reward function during the SAC agent training process is as follows: in, This represents the value of the composite reward function. , These represent the weights of power generation efficiency and system energy consumption, respectively. Indicates power generation efficiency. Indicates system energy consumption cost. Indicates a penalty item. Indicates power generation efficiency; The action range constraint of the i-th action output by the SAC agent at time t is the action range constraint term of the reward function, which is used to constrain the range of the actions output by the SAC agent. Let represent the i-th action output by the SAC agent at time t. Indicates the weight of the action penalty. , These represent the lower and upper limits of the given action range, respectively. This represents the state range constraint term of the composite reward function. This represents the j-th state of the system at time t. Indicates the state penalty weight. , Let A and S represent the lower and upper limits of the given state range, respectively. Let A represent the action space and S represent the state space.

6. The intelligent control method for a liquefied natural gas cold power generation system according to claim 1, characterized in that, The action space of the SAC agent includes LNG feed flow rate, seawater flow rate, and working fluid temperature; The state space of the SAC agent includes the natural gas outlet temperature, working fluid evaporation temperature, working fluid expansion temperature, working fluid condensation temperature, and seawater outlet temperature.

7. The intelligent control method for a liquefied natural gas cold power generation system according to claim 2, characterized in that, The seawater inlet temperature is used as a perturbation, which is randomly added during the preset training rounds or testing phases of the SAC agent training.

8. The intelligent control method for a liquefied natural gas cold power generation system according to claim 3, characterized in that, The relative gain matrix between the control variable and the controlled variable is calculated based on the gain of the transfer function, including: Construct the gain matrix based on the gain of the transfer function; Based on the importance of the controlled variable to power generation, the gain matrix is ​​split into two local linear gain matrices; Calculate the relative gain matrix of the two local linear gain matrices respectively.

9. The intelligent control method for a liquefied natural gas cold power generation system according to claim 3, characterized in that, The transfer function model is as follows: ， in, For the transfer function model, For Laplace factor, - These are the pole coefficients. - Zero-point coefficient, For pure time delay, This represents the Laplace transform value of the controlled variable. This represents the Laplace transform value of the key control variable, where m represents the highest order of the denominator polynomial and n represents the highest order of the numerator polynomial.

10. An intelligent control device for a liquefied natural gas (LNG) cold power generation system, implementing the intelligent control method for an LNG cold power generation system as described in any one of claims 1 to 9, characterized in that, The device includes: The modeling module is used to determine the key control variables and controlled variables, and to build an identification model of the liquefied natural gas cold energy power generation system based on the key control variables and controlled variables. The agent training module is used to construct the SAC agent, establish a composite reward function with cold energy power generation efficiency, power generation benefits and system energy consumption as the core, and train the constructed SAC agent based on the composite reward function to obtain the trained SAC agent. The control module is used to add a first-order low-pass filter to the output of the trained SAC agent for filtering. The identification model outputs the current state variable based on the filtered action. The trained SAC agent outputs the optimal control action for the liquefied natural gas cold energy power generation system based on the current state variable.

Images