A gas filling whole-process data-driven optimization method and system and a storage medium

Abstract

The application provides a gas filling whole-process data driving optimization method and system and a storage medium, which comprises the following steps: constructing constraint conditions with safety pressure boundary and pressure change rate as cores by monitoring pressure and temperature in real time and calculating state entropy increase gradient; generating an initial filling rate curve by using a generative adversarial network; combining the constraint, efficiency and equipment loss to perform multi-objective evaluation on a discriminator of the network, and adjusting a safety weight in real time according to the entropy increase gradient; and smoothing and optimizing the curve by wavelet transform and nonlinear Kalman filtering, wherein the decomposition layer number and the filtering parameter are both associated with the entropy increase gradient, so that a stable and safe execution rate curve is generated, and intelligent closed-loop control of the filling process is realized.

Description

Technical Field

[0001] This application belongs to the field of optimization, and in particular relates to a data-driven optimization method, system and storage medium for the entire gas filling process. Background Technology

[0002] Common control methods for tank gas filling processes include constant-rate filling, staged filling, or PID control based on pressure and temperature measurements. During the compression filling process, the internal state of the gas undergoes drastic changes; for example, throttling effects can cause temperature increases or decreases, or even phase changes. Traditional control methods, due to their relatively fixed control logic, struggle to adapt to this process in real time. To mitigate risks, they often employ very conservative filling rates, resulting in excessively long filling times and low operational efficiency. Conversely, increasing the rate to pursue efficiency may lead to an inability to predict and respond to drastic fluctuations in the system state, potentially approaching or even exceeding the true safety boundary, causing serious safety hazards. Traditional methods treat the safety boundary as a red line, ignoring the fact that the actual safety margin changes with factors such as real-time tank mass and temperature trends. Regarding the filling rate, it is difficult to achieve a globally optimal solution among the three mutually constraining objectives of filling efficiency, process safety, and wear and tear on valves, pumps, and other actuators. Furthermore, the generated control commands often fail to fully consider the response characteristics of the actuators, potentially containing high-frequency jitter, which not only affects the smoothness and accuracy of control but also accelerates the wear of critical equipment and increases maintenance costs. Therefore, there is an urgent need for a new method that can accurately represent the risks of the filling process in real time, and on this basis, adjust the constraints to generate a smooth optimal control strategy that takes into account multiple objectives. Summary of the Invention

[0003] This invention proposes a data-driven optimization method for the entire gas filling process to address the problem that existing methods struggle to accurately and in real-time represent the risks of the filling process, and that adjusting constraints fails to generate a smooth optimal control strategy that balances multiple objectives. The method includes the following steps:

[0004] The inherent parameters of the tank and the real-time pressure and temperature during the filling process are obtained, and the state entropy increase gradient, which serves as a real-time risk metric, is calculated based on the real-time pressure and temperature.

[0005] Establish constraints including a safe pressure boundary determined based on the current filling mass and temperature change rate, and a maximum permissible pressure change rate determined based on the real-time value of the state entropy increase gradient;

[0006] Generative adversarial networks are used to generate an initial filling rate curve, wherein the discriminator of the generative adversarial network performs a multi-objective comprehensive evaluation based on the constraints, filling efficiency and equipment loss;

[0007] The number of decomposition levels of the wavelet transform is determined based on the historical average value of the state entropy gradient. The initial filling rate curve is decomposed into low-frequency and high-frequency components using the wavelet transform. The process noise covariance of the filter is determined based on the real-time value of the state entropy gradient. The high-frequency component is smoothed using a nonlinear Kalman filter. The smoothed high-frequency component and the low-frequency component are reconstructed to obtain the execution rate curve. The execution mechanism of the filling system is controlled according to the execution rate curve.

[0008] Furthermore, this invention also relates to a data-driven optimization system for the entire gas filling process, comprising the following modules:

[0009] The calculation module is used to obtain the inherent parameters of the tank and the real-time pressure and temperature during the filling process, and to calculate the state entropy increase gradient as a real-time risk measurement indicator based on the real-time pressure and temperature.

[0010] The determination module is used to establish constraints including a safe pressure boundary determined based on the current filling mass and temperature change rate, and a maximum permissible pressure change rate determined based on the real-time value of the state entropy increase gradient.

[0011] The generation module is used to generate an initial filling rate curve using a generative adversarial network, wherein the discriminator of the generative adversarial network performs a multi-objective comprehensive evaluation based on the constraints, filling efficiency and equipment loss;

[0012] The control module is used to determine the number of wavelet transform decomposition levels based on the historical average value of the state entropy gradient, decompose the initial filling rate curve into low-frequency and high-frequency components using wavelet transform, determine the process noise covariance of the filter based on the real-time value of the state entropy gradient, smooth the high-frequency component using a nonlinear Kalman filter, reconstruct the execution rate curve by combining the smoothed high-frequency component with the low-frequency component, and control the actuator of the filling system according to the execution rate curve.

[0013] This invention utilizes the state entropy gradient as a real-time risk metric to achieve a more accurate and scientific assessment of the safety status during the filling process. The established safety boundaries and pressure change rate limits are adjusted based on the actual risks encountered during the process. By comprehensively optimizing multiple objectives—safety, filling efficiency, and equipment wear—and adjusting safety weights according to real-time risks, an optimal filling strategy can be generated, improving filling efficiency while ensuring high safety. Through refined processing of the rate curve, high-frequency jitter and abrupt changes in control commands are filtered out, resulting in a smooth and stable control process. This reduces mechanical impact on valves and other actuators, extends equipment lifespan, and optimizes the overall performance of the entire filling process. Attached Figure Description

[0014] Figure 1 A flowchart of the first embodiment;

[0015] Figure 2 This is a schematic diagram illustrating the relationship between the entropy increase gradient and process parameters;

[0016] Figure 3 This is a schematic diagram of the constraints.

[0017] Figure 4 This is a schematic diagram of the filter optimization for the initial rate curve;

[0018] Figure 5 This is a schematic diagram of closed-loop tracking control for the rate curve. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] In the first embodiment, see Figure 1 The flowchart shown below illustrates the data-driven optimization method for the entire gas filling process, which includes the following steps:

[0021] S1, obtain the inherent parameters of the tank and the real-time pressure and temperature during the filling process, and calculate the state entropy increase gradient as a real-time risk measurement indicator based on the real-time pressure and temperature.

[0022] Specifically, parameters such as tank volume, material, and design pressure are input through a human-machine interface or retrieved from the equipment database. For example, the tank volume is 50 m³ and the design pressure is 2.5 MPa. A pressure transmitter and armored thermocouple are installed on the tank. The PLC collects the real-time pressure P and real-time temperature T inside the tank once per second. The current filling mass m(t) is accumulated by a mass flow meter. Figure 2 .

[0023] In one embodiment, calculating the state entropy increase gradient as a real-time risk metric based on the real-time pressure and real-time temperature includes:

[0024] The pressure and temperature data are collected continuously during the filling process, and the pressure and temperature data are synchronously calibrated and noise-reduced.

[0025] Based on the pressure and temperature data, a state quantity reflecting the characteristics of system state changes is constructed.

[0026] Based on the changing trend of the state variables in adjacent time periods, the rate of change of the system state disorder is calculated, and the rate of change is used as the state entropy increase gradient representing the risk level of the filling process.

[0027] Specifically, the real-time collected pressure P(t) and temperature T(t) are normalized by dividing by the design pressure and critical temperature, respectively, to construct a state vector. Set a sliding time window of length N, calculate the covariance matrix of the state vector sequence within the window, and perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalues. , And calculate the local information entropy. Calculate the derivative of the local information entropy with respect to time, and you will get the state entropy gradient.

[0028] S2, establish constraints including a safe pressure boundary determined based on the current filling mass and temperature change rate, and a maximum allowable pressure change rate determined based on the real-time value of the state entropy increase gradient;

[0029] The system obtains the current filling mass and corresponding temperature change rate; based on the filling mass and temperature change rate, it determines the gas's thermal expansion sensitivity under the current operating conditions; combining the tank structure parameters and design pressure range, it performs a safety mapping on the thermal expansion sensitivity to obtain a dynamic safety pressure boundary corresponding to the current filling state. Specifically, by querying a pre-set gas thermodynamic property database such as NIST REFPROP, it obtains the gas's thermal expansion coefficient and isothermal compressibility under the current pressure and temperature conditions, using them as a physical benchmark for measuring thermal expansion sensitivity; it establishes a multidimensional lookup table to map the above thermal expansion sensitivity indicators to safety margin reduction coefficients; when the gas is in a high thermal expansion sensitivity zone, the lookup table outputs a larger reduction coefficient, and vice versa. Figure 3 .

[0030] The system continuously acquires the current value of the state entropy increase gradient; based on the magnitude and trend of the state entropy increase gradient, it determines the rate of risk change during the filling process; and adaptively sets corresponding upper limits for pressure changes when the rate of risk change is at different levels, thereby obtaining constraints on the maximum allowable rate of pressure change during the filling process. Specifically, the system monitors the state entropy increase gradient in real time. If it is less than or equal to the warning threshold, the rate of risk change is determined to be low, allowing the system to use the baseline maximum rate of pressure change to improve efficiency. If it is greater than or equal to the alarm threshold, it is determined to be a high-risk rate, and an exponential suppression mechanism is forcibly activated. If it is between the warning threshold and the alarm threshold, it is determined to be a medium-risk rate, and a linear suppression mechanism is activated.

[0031] In an optional embodiment, the safety pressure boundary determined based on the current fill mass and the rate of temperature change includes:

[0032] Safety pressure boundary The calculation formula is:

[0033]

[0034] in, The design pressure of the tank is given, and m(t) is the current filling mass. This represents the absolute value of the rate of temperature change. and The preset positive coefficient, The unit is pressure / mass. The unit is pressure / (temperature / time).

[0035] For example, the design pressure of a high-pressure hydrogen storage tank The pressure is set at 25 MPa based on material and manufacturing standards. Safety factor. and Calibration is achieved through finite element analysis and experimental data, for example... It is set to 0.005 MPa / kg to represent the safety margin reduction corresponding to the stress generated on the tank wall per unit filling mass; It is set to 0.2 MPa·s / K to represent the effect of transient thermal stress caused by rapid temperature changes on the safety boundary.

[0036] During the filling process, two variables are monitored in real time: the current filling mass m(t) accumulated by the mass flow meter and the rate of temperature change calculated by the temperature sensor. Suppose that at a certain moment t, the filling mass reaches 1000 kg, and simultaneously, due to the heat released during gas compression, the temperature rises from 300 K to 300.5 K in the past second, i.e., the rate of temperature change. The value is 0.5 K / s. Substituting the real-time data into the formula for calculation, the pressure of the mass term is reduced to... The pressure reduction of the temperature change term is Therefore, the safe pressure boundary at this moment is calculated as follows: The results will serve as the real-time upper limit for the control system, ensuring that the operational pressure at any given time remains below this adjusted safety threshold, which takes into account multiple physical effects.

[0037] To adjust the aggressiveness of the filling process based on real-time risk, in one optional embodiment, the maximum permissible rate of pressure change determined based on the real-time value of the state entropy increase gradient includes:

[0038] Maximum permissible pressure change rate The calculation formula is:

[0039]

[0040] in, The benchmark maximum pressure change rate, This represents the real-time value of the gradient of the state entropy increase. This is a preset positive coefficient, with units of K / kW.

[0041] Reference maximum pressure change rate This is the fastest filling rate allowed under ideal safety conditions, for example, set at 0.25 MPa / s. Risk sensitivity coefficient. This determines the system's sensitivity to risk response, and is set according to the security requirement level, for example... K / kW.

[0042] During the filling process, the gradient of state entropy increase is calculated in real time. Suppose that at a certain moment, due to inlet airflow disturbance, the state becomes unstable, and the calculated... The value is 1.2 kW / K. Substituting this risk value into the formula, the exponential decay factor is calculated, and the value is... The maximum permissible pressure change rate at the current moment is obtained by modulating the reference rate of change using an attenuation factor. The calculation results are transmitted to the actuator controlling the valve, which is then instructed to reduce the valve opening, strictly limiting the actual pressure rise rate to below 0.095 MPa / s, thereby proactively suppressing further risk without interrupting the filling process.

[0043] S3, Generative Adversarial Network (GAN) is used to generate an initial filling rate curve, wherein the discriminator of the GAN performs a multi-objective comprehensive evaluation based on the constraints, filling efficiency and equipment loss;

[0044] The generator uses a Long Short-Term Memory (LSTM) network. It takes the current tank state as input and outputs a sequence of filling rates at multiple future time points as the initial filling rate curve. The discriminator uses a one-dimensional convolutional neural network (CNN). It takes this rate curve as input and outputs a comprehensive evaluation value. This evaluation value is obtained by weighted summation of three parts: safety evaluation, efficiency evaluation, and loss evaluation. The weight coefficient for safety is set as the base weight of 0.5 plus 0.25 multiplied by the real-time entropy increase gradient value. Safety evaluation is achieved through an embedded fast simulation model. This model takes candidate filling rate curves as input, predicts the pressure change trajectory in the future time domain, and compares it with the real-time calculated safe pressure boundary. The sum of squares of the pressure exceeding the limit is calculated by integration as a penalty value, thereby quantifying the safety risk of the strategy. Secondly, efficiency evaluation calculates the total time required to complete the target filling amount by integrating the candidate rate curves, and uses its reciprocal as the efficiency score. The shorter the time, the higher the score. Loss evaluation measures the degree of fluctuation of control commands by calculating the sum of squares of the differences between adjacent data points in the rate curve sequence and taking the negative number as the score. The smaller the fluctuation, the lower the loss to actuators such as valves. The three scores are weighted and summed after the weights are dynamically adjusted according to the real-time entropy increase gradient, which yields the comprehensive evaluation value output by the discriminator.

[0045] In an optional embodiment, the discriminator employs a multi-layer one-dimensional convolutional neural network structure, whose input is time-series data of the filling rate curve. The network comprises three convolutional layers: the first layer uses 32 kernels of length 5 to extract basic fluctuation features; the second layer uses 64 kernels of length 3 to capture more complex temporal patterns; and the third layer uses 128 kernels of length 3 to identify higher-order dynamic characteristics. Each convolutional layer is followed by a ReLU activation function and a max-pooling layer with a stride of 2, progressively compressing the temporal dimension and enhancing feature representation. After convolutional feature extraction, the data is mapped through fully connected layers to three independent branch outputs, corresponding to initial scores for safety, efficiency, and loss, respectively. The scores are then weighted and fused according to dynamic weights, where the safety weight increases linearly with the real-time entropy gradient, ultimately generating a single comprehensive evaluation value. The entire network is optimized through backpropagation and adversarial training to accurately evaluate the merits of filling strategies.

[0046] In one embodiment, the generation of the initial filling rate curve using a generative adversarial network (GAN) is described, wherein the discriminator of the GAN performs a multi-objective comprehensive evaluation based on the constraints, filling efficiency, and equipment loss, including:

[0047] The generative adversarial network is trained using historical safe filling data so that it can generate filling rate curves that meet basic safety requirements.

[0048] The generated filling rate curve is input into the discriminator, and its safety is evaluated based on the maximum allowable pressure change rate constraint.

[0049] Simultaneously, the filling efficiency of the filling rate curve is evaluated from the perspective of filling time and energy utilization, and the equipment loss is evaluated from the perspective of equipment operating load and stress changes;

[0050] Based on the comprehensive evaluation results of safety, filling efficiency and equipment wear, the generated filling rate curves are screened and output as the initial filling rate curves.

[0051] Specifically, during the training phase, W-Distance is used as the loss function. The parameters of the generator and discriminator are optimized iteratively until the discriminator can no longer distinguish the data distribution of the generated curve from the historical real safety curve, marking the completion of model training. In the real-time inference and filtering phase, the generator generates K sets of candidate filling rate curves in parallel based on the current tank state input. Subsequently, the trained discriminator performs forward calculations on these K sets of curves to obtain the comprehensive evaluation value corresponding to each set of curves. The system automatically sorts these K evaluation values ​​and selects the curve with the highest score as the initial filling rate curve output.

[0052] In an optional embodiment, the discriminator of the generative adversarial network performs a multi-objective comprehensive evaluation based on the constraints, filling efficiency, and equipment loss, including:

[0053] The multi-objective comprehensive evaluation of the discriminator is a weighted sum of the scores of evaluation indicators for safety, filling efficiency, and equipment loss;

[0054] Among them, the weight of security The calculation formula is: ;

[0055] in, Based on security weights, This represents the real-time value of the gradient of the state entropy increase. This is a preset positive coefficient, with units of K / kW.

[0056] This is applied to an intelligent control policy generator based on a generative adversarial network (GAN), where the discriminator is responsible for evaluating the merits of alternative control policies. The discriminator's evaluation function is a multi-objective weighted sum, in the form of... Where S represents the score of each objective. Basic weights are preset, for example... , , This represents the highest level of importance placed on safety under normal circumstances. Risk Response Coefficient Set to 0.25.

[0057] During operation, the entropy increase gradient is acquired in real time. Assume that at time t, a moderate risk is detected. At this point, the weight of security is adjusted to... To maintain a total weight of 1, other weights are normalized, for example, by being reduced proportionally. and The adjusted weight set may become , , When the discriminator uses this new set of weights to evaluate an alternative control strategy designed to improve the filling rate, even if the strategy's efficiency score is low... Very high, but only if the security score is high. Even a slight decrease will cause the overall evaluation J(t) to... The generator is penalized for its improved performance. This forces the generator to develop control strategies that prioritize safety in subsequent iterations.

[0058] In another embodiment, the generative adversarial network comprises two neural networks: a generator and a discriminator. The generator's network structure typically employs a Long Short-Term Memory (LSTM) network, containing an input layer, three LSTM hidden layers, and an output layer, processing and generating time-series data. The discriminator's network structure can be a Multilayer Perceptron (MLP), containing an input layer, four fully connected hidden layers, and an output layer. The generator's input is the current system state vector. And a random noise vector z sampled from a Gaussian distribution. The output is a control strategy, specifically a sequence of valve opening control commands over a future time period. The input to the discriminator is a state sequence. and the corresponding control command sequence The sequences can come from a generator or from real historical best-performing operation data. The output is a single scalar evaluation score J(t), representing the overall performance of the input sequences. The training set consists of a large amount of expert operation data collected from historical loading processes or high-performance control sequences verified through simulation, labeled as real samples. The training process employs adversarial learning.

[0059] S4. Determine the decomposition level of the wavelet transform based on the historical average value of the state entropy gradient, and decompose the initial filling rate curve into low-frequency and high-frequency components using the wavelet transform; determine the process noise covariance of the filter based on the real-time value of the state entropy gradient, and smooth the high-frequency component using a nonlinear Kalman filter; reconstruct the execution rate curve by combining the smoothed high-frequency component with the low-frequency component; and control the actuator of the filling system according to the execution rate curve.

[0060] Specifically, the initial filling rate curve was subjected to discrete wavelet transform using the Daubechies4 wavelet basis, with the number of decomposition levels set to 3 + 1.5 × the rounded result of the historical entropy gradient average. The high-frequency coefficients of each decomposed level were filtered using an extended Kalman filter, and the process noise covariance matrix Q was set as a diagonal matrix, with diagonal elements equal to the baseline value of 0.01 × (1 + 8 times the real-time entropy gradient value). Using the processed high-frequency coefficients and the original low-frequency coefficients, the execution rate curve was reconstructed through inverse discrete wavelet transform. (See [reference needed]). Figure 4 .

[0061] The execution rate curve is sent as a time series setpoint to the PLC controller. The PLC has a built-in PID control module, which uses the rate corresponding to the current time point in the curve as the setpoint SP, and the actual filling rate measured by the mass flow meter as the process quantity PV. By calculating the deviation between SP and PV, a 4-20mA current signal is output to control the opening of the electric regulating valve on the filling pipeline, or to adjust the frequency of the compressor or pump inverter, so that the actual filling rate accurately follows the execution rate curve. See [link to relevant documentation]. Figure 5 .

[0062] In an optional embodiment, determining the number of wavelet transform decomposition levels based on the historical average of the state entropy increase gradient includes:

[0063] The formula for calculating the number of decomposition levels L in wavelet transform is: ;

[0064] in, Number of basic decomposition layers The historical average value of the state entropy increase gradient is given. This is a preset positive coefficient, with units of K / kW. This indicates the floor function.

[0065] This embodiment is used to adjust the analysis depth of the signal processing module to better extract fault features from sensor data. It sets basic parameters and the basic decomposition level. Setting it to 4 during stable operation is a default value that strikes a balance between computational efficiency and analytical accuracy. Historical risk impact coefficient. Set to 1.5. Historical average. The calculation window is set to the past 5 minutes, caching all data within that time period. Value and calculate the arithmetic mean.

[0066] Assuming that the filling process experienced a series of small fluctuations over the past 5 minutes, resulting in average The value is 0.9. Substitute this value into the formula to calculate the new decomposition level: The calculation result 5 will be immediately applied to the wavelet transform algorithm for signals such as pressure and temperature. Increasing the decomposition level from 4 to 5 indicates a longer timescale for signal analysis and improved frequency resolution. This allows the algorithm to more precisely detect features related to potential risks, such as low-frequency oscillations or slow drifts, which might be obscured by noise at lower decomposition levels. In this way, the focus of signal analysis is automatically adjusted based on the system's historical risk status, enhancing fault diagnosis.

[0067] In an optional embodiment, determining the process noise covariance of the filter based on the real-time value of the state entropy increase gradient includes:

[0068] The process noise covariance matrix of a nonlinear Kalman filter The calculation formula is: ;

[0069] in, The reference process noise covariance matrix, This represents the real-time value of the gradient of the state entropy increase. This is a preset positive scalar coefficient, with units of K / kW.

[0070] This estimator, for example, is an extended Kalman filter used to estimate the true state of the gas inside the tank from noisy measurements. Baseline parameters are set, including the baseline process noise covariance matrix. Based on prior knowledge of the system model, such as for a system containing two states: pressure and temperature, Set as a diagonal matrix This indicates that the pressure and temperature predicted by the model have a fixed uncertainty at each step. Risk adjustment coefficient. Set to 1.2K / kW.

[0071] During operation, the entropy increase gradient is calculated in real time. At one point, the filling process was very smooth. The value is very low, for example, 0.2. A regulation factor is calculated. The factor will be used to scale the baseline covariance matrix to obtain the current... A smaller Q(t) value indicates high stability and highly reliable predictions. Therefore, the filter relies more on model predictions, filtering out measurement noise and outputting a smoother state estimate. Conversely, if the risk increases, Q(t) will increase, making the filter more reliant on new measurements and responding quickly to real changes in the system state.

[0072] In a second embodiment, the present invention also provides a data-driven optimization system for the entire gas filling process, comprising the following modules:

[0073] The calculation module is used to obtain the inherent parameters of the tank and the real-time pressure and temperature during the filling process, and to calculate the state entropy increase gradient as a real-time risk measurement indicator based on the real-time pressure and temperature.

[0074] The determination module is used to establish constraints including a safe pressure boundary determined based on the current filling mass and temperature change rate, and a maximum permissible pressure change rate determined based on the real-time value of the state entropy increase gradient.

[0075] The generation module is used to generate an initial filling rate curve using a generative adversarial network, wherein the discriminator of the generative adversarial network performs a multi-objective comprehensive evaluation based on the constraints, filling efficiency and equipment loss;

[0076] A control module is used to filter and optimize the initial filling rate curve to generate an execution rate curve. The optimization includes: decomposing the initial filling rate curve into low-frequency and high-frequency components using wavelet transform, wherein the number of decomposition levels of the wavelet transform is determined based on the historical average value of the state entropy gradient; smoothing the high-frequency components using a nonlinear Kalman filter, wherein the process noise covariance of the filter is determined based on the real-time value of the state entropy gradient; reconstructing the smoothed high-frequency components and the low-frequency components; and controlling the actuator of the filling system according to the execution rate curve.

[0077] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0078] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A data-driven optimization method for the entire gas filling process, characterized in that, Includes the following steps: The inherent parameters of the tank and the real-time pressure and temperature during the filling process are obtained, and the state entropy increase gradient, which serves as a real-time risk metric, is calculated based on the real-time pressure and temperature. Establish constraints including a safe pressure boundary determined based on the current filling mass and temperature change rate, and a maximum permissible pressure change rate determined based on the real-time value of the state entropy increase gradient; Generative adversarial networks are used to generate an initial filling rate curve, wherein the discriminator of the generative adversarial network performs a multi-objective comprehensive evaluation based on the constraints, filling efficiency and equipment loss; The number of decomposition levels of the wavelet transform is determined based on the historical average value of the state entropy gradient. The initial filling rate curve is decomposed into low-frequency and high-frequency components using the wavelet transform. The process noise covariance of the filter is determined based on the real-time value of the state entropy gradient. The high-frequency component is smoothed using a nonlinear Kalman filter. The smoothed high-frequency component and the low-frequency component are reconstructed to obtain the execution rate curve. The execution mechanism of the filling system is controlled according to the execution rate curve.

2. The method according to claim 1, characterized in that, The calculation of the state entropy increase gradient, which serves as a real-time risk metric, based on the real-time pressure and temperature includes: The pressure and temperature data are collected continuously during the filling process, and the pressure and temperature data are synchronously calibrated and noise-reduced. Based on the pressure and temperature data, a state quantity reflecting the characteristics of system state changes is constructed. Based on the changing trend of the state variables in adjacent time periods, the rate of change of the system state disorder is calculated, and the rate of change is used as the state entropy increase gradient representing the risk level of the filling process.

3. The method according to claim 1 or 2, characterized in that, The safety pressure boundary determined based on the current filling mass and temperature change rate includes: Obtain the current filling mass and the corresponding temperature change rate; Based on the filling mass and temperature change rate, determine the gas's sensitivity to thermal expansion under the current operating conditions; By combining the tank structure parameters and the design pressure range, the thermal expansion sensitivity is safely mapped to obtain the dynamic safety pressure boundary corresponding to the current filling state.

4. The method according to claim 1 or 2, characterized in that, The constraint condition for the maximum permissible rate of pressure change determined based on the real-time value of the state entropy gradient includes: The current value of the state entropy increase gradient is obtained in real time; Based on the magnitude and trend of the state entropy increase gradient, the rate of risk change during the filling process is determined; When the rate of risk change is at different levels, the corresponding upper limit of pressure change is adaptively set to obtain the constraint condition for limiting the maximum allowable rate of pressure change during the filling process.

5. The method according to claim 1, characterized in that, The initial filling rate curve is generated using a generative adversarial network (GAN), wherein the discriminator of the GAN performs a multi-objective comprehensive evaluation based on the constraints, filling efficiency, and equipment loss, including: The generative adversarial network is trained using historical safe filling data so that it can generate filling rate curves that meet basic safety requirements. The generated filling rate curve is input into the discriminator, and its safety is evaluated based on the maximum allowable pressure change rate constraint. Simultaneously, the filling efficiency of the filling rate curve is evaluated from the perspective of filling time and energy utilization, and the equipment loss is evaluated from the perspective of equipment operating load and stress changes; Based on the comprehensive evaluation results of safety, filling efficiency and equipment wear, the generated filling rate curves are screened and output as the initial filling rate curves.

6. The method according to claim 1, characterized in that, The number of wavelet transform decomposition levels is determined based on the historical average value of the state entropy gradient, including: The formula for calculating the number of decomposition levels L in wavelet transform is: ; in, Number of basic decomposition layers The historical average value of the state entropy increase gradient is given. This is a preset positive coefficient, with units of K / kW. This indicates the floor function.

7. The method according to claim 1, characterized in that, The process noise covariance of the filter determined based on the real-time value of the state entropy gradient includes: The process noise covariance matrix of a nonlinear Kalman filter The calculation formula is: ; in, The reference process noise covariance matrix, This represents the real-time value of the gradient of the state entropy increase. This is a preset positive scalar coefficient, with units of K / kW.

8. A data-driven optimization system for the entire gas filling process, characterized in that, Includes the following modules: The calculation module is used to obtain the inherent parameters of the tank and the real-time pressure and temperature during the filling process, and to calculate the state entropy increase gradient as a real-time risk measurement indicator based on the real-time pressure and temperature. The determination module is used to establish constraints including a safe pressure boundary determined based on the current filling mass and temperature change rate, and a maximum permissible pressure change rate determined based on the real-time value of the state entropy increase gradient. The generation module is used to generate an initial filling rate curve using a generative adversarial network, wherein the discriminator of the generative adversarial network performs a multi-objective comprehensive evaluation based on the constraints, filling efficiency and equipment loss; The control module is used to determine the number of wavelet transform decomposition levels based on the historical average value of the state entropy gradient, decompose the initial filling rate curve into low-frequency and high-frequency components using wavelet transform, determine the process noise covariance of the filter based on the real-time value of the state entropy gradient, smooth the high-frequency component using a nonlinear Kalman filter, reconstruct the execution rate curve by combining the smoothed high-frequency component with the low-frequency component, and control the actuator of the filling system according to the execution rate curve.

9. The system according to claim 8, characterized in that, The calculation of the state entropy increase gradient, which serves as a real-time risk metric, based on the real-time pressure and temperature includes: The pressure and temperature data are collected continuously during the filling process, and the pressure and temperature data are synchronously calibrated and noise-reduced. Based on the pressure and temperature data, a state quantity reflecting the characteristics of system state changes is constructed. Based on the changing trend of the state variables in adjacent time periods, the rate of change of the system state disorder is calculated, and the rate of change is used as the state entropy increase gradient representing the risk level of the filling process.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when executed by a processor, implements the method as described in any one of claims 1-7.

Images