Power dynamic balance regulation method and system for renewable energy hydrogen-based energy source

Abstract

The application provides a renewable energy hydrogen-based energy power dynamic balance regulation method and system, relates to the technical field of energy system control, and comprises the following steps: acquiring renewable energy power generation power, energy storage state and load demand data, establishing a dynamic balance equation to calculate a surplus power value, and collecting equipment parameters to calculate operation efficiency; a multi-scale space-time attention network is constructed, power and efficiency data are predicted through an adaptive convolution kernel and a position weighted attention mechanism, a power distribution scheme is calculated based on the predicted data, uncertainty is modeled and risk is predicted by using a Gaussian process deep reinforcement learning, control parameters are obtained by combining dynamic programming; and the efficiency is monitored and an optimal power distribution scheme is determined.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy system control, and in particular to a power dynamic balance regulation method and system for a renewable energy hydrogen-based energy source. BACKGROUND

[0002] With the large-scale development of renewable energy, how to efficiently consume renewable energy has become a problem to be solved. Hydrogen production as an important way of renewable energy consumption, the system regulation effect directly affects the energy utilization efficiency;

[0003] At present, the control method of renewable energy hydrogen production system mainly includes rule-based control strategy and traditional optimization algorithm. The rule-based control strategy depends on expert experience, which is difficult to adapt to complex and changeable operating environment, and the traditional optimization algorithm has high computational complexity and poor real-time performance, which cannot meet the regulation and control requirements of large-scale hydrogen production system;

[0004] In recent years, researchers have tried to apply deep learning to renewable energy power generation prediction, but the existing technology still has problems such as not fully considering the dynamic correlation between system components, low prediction accuracy, lack of systematic research on multi-time scale coordinated regulation, and ignoring the system risks brought by environmental uncertainty;

[0005] Therefore, there is an urgent need for a solution to solve the problems in the prior art. SUMMARY

[0006] The embodiments of the present application provide a power dynamic balance regulation method and system for a renewable energy hydrogen-based energy source, which can at least solve some of the problems in the prior art.

[0007] In a first aspect, the present application provides a power dynamic balance regulation method for a renewable energy hydrogen-based energy source, comprising:

[0008] Obtaining real-time power generation data of a renewable energy power generation system, real-time energy storage state data of an energy storage device, and power grid load demand data, establishing a dynamic balance equation and calculating a surplus power value available for hydrogen production, collecting equipment operation parameters of a hydrogen production device to calculate a current operation efficiency value of the hydrogen production device;

[0009] A multi-scale spatio-temporal attention network is constructed to extract multi-dimensional time sequence features through an adaptive convolution kernel, and a location-weighted attention mechanism is used to model the dynamic correlation between system components to calculate power prediction data and efficiency prediction data for the next 24 hours;

[0010] An electrolytic tank power distribution scheme is calculated according to the power prediction data, the efficiency prediction data, the dynamic balance equation, and the surplus power value;

[0011] Based on the electrolytic tank power distribution scheme, a deep reinforcement learning algorithm based on Gaussian process is adopted, the environmental uncertainty is modeled and the system risk is predicted through a Bayesian optimization framework, and the target power parameters and control parameters of each electrolytic tank unit are calculated by combining an online dynamic programming method, and the operating parameters of each electrolytic tank unit are dynamically adjusted according to the control parameters;

[0012] The actual energy utilization efficiency value of each electrolytic tank unit under the control parameters is monitored, and when the actual energy utilization efficiency value is lower than the preset threshold, the electrolytic tank power distribution scheme is updated based on the power prediction data and the efficiency prediction data to obtain an optimal power distribution scheme.

[0013] In an alternative embodiment,

[0014] The dynamic balance equation is established and the surplus power value available for hydrogen production is calculated, and the current operating efficiency value of the hydrogen production device is calculated by collecting the equipment operating parameters of the hydrogen production device, including:

[0015] The real-time power generation data, the real-time energy storage state data and the grid load demand data are substituted into the dynamic balance equation, wherein the dynamic balance equation represents that the real-time power generation data is equal to the sum of the grid load demand data, the charge and discharge power change of the energy storage device and the surplus power value, and the surplus power value available for hydrogen production is calculated according to the dynamic balance equation;

[0016] The temperature parameter, pressure parameter, current parameter and hydrogen production parameter of the hydrogen production device are collected, the product of the hydrogen production parameter and the high heating value of hydrogen is taken as the numerator, and the product of the current parameter and the voltage of the electrolytic tank of the hydrogen production device is taken as the denominator, and the current operating efficiency value of the hydrogen production device is calculated according to the ratio of the numerator to the denominator.

[0017] In an alternative embodiment,

[0018] A multi-scale spatio-temporal attention network is constructed, multi-dimensional time sequence features are extracted through adaptive convolution kernels, and the dynamic correlation between system components is modeled using a position-weighted attention mechanism, and the power prediction data and efficiency prediction data for the next 24 hours are calculated, including:

[0019] A multi-scale spatio-temporal attention network is constructed, multi-dimensional time sequence features are extracted through adaptive convolution kernels, and the dynamic correlation between system components is modeled using a position-weighted attention mechanism, and the power prediction data and efficiency prediction data for the next 24 hours are calculated, including:

[0020] constructing a query matrix and a key matrix based on the multi-dimensional time sequence features, adding a product of the query matrix and the key matrix to a position weighting matrix, and then normalizing to obtain an attention weight matrix, multiplying the attention weight matrix and a value matrix of the multi-dimensional time sequence features to obtain attention output data;

[0021] performing nonlinear transformation on the attention output data by using a multilayer perceptron to model dynamic correlations between system components and obtain a dynamic correlation graph;

[0022] inputting the dynamic correlation graph into a corresponding decoder network together with power historical data and efficiency historical data, calculating to obtain power prediction data and efficiency prediction data for the next 24 hours, constructing a joint loss function based on mean square errors of the power prediction data and actual power data and mean square errors of the efficiency prediction data and actual efficiency data, and optimizing parameters of the decoder network based on the joint loss function.

[0023] In an optional implementation,

[0024] performing nonlinear transformation on the attention output data by using a multilayer perceptron to model dynamic correlations between system components and obtain a dynamic correlation graph includes:

[0025] obtaining attention output data, and constructing a causal graph based on the attention output data, wherein the causal graph includes a node set, an edge set, and a causal probability matrix, features of the attention output data are taken as the node set, correlations between the attention output data are taken as the edge set, and dynamic correlations between the attention output data are taken as the causal probability matrix;

[0026] performing counterfactual intervention by using a multilayer perceptron to impose different intervention conditions on the attention output data, calculating expected outputs under different intervention conditions, and calculating a difference between the expected outputs under different intervention conditions to obtain a causal effect strength;

[0027] performing nonlinear transformation on the attention output data based on the causal effect strength, multiplying a result of the nonlinear transformation and an output value of a sigmoid function to obtain adjusted dynamic correlations;

[0028] constructing physical characteristic parameters and motion data corresponding to system components, constructing kinematic constraint equations, mass conservation equations, and energy conservation equations based on the physical characteristic parameters and the motion data, and combining to obtain a physical constraint equation set, inputting the adjusted dynamic correlations into the physical constraint equation set, calculating a norm value of the physical constraint equation set and a norm value of a gradient of the physical constraint equation set, generating a weighted sum according to the norm value of the physical constraint equation set and the norm value of the gradient of the physical constraint equation set, and taking the weighted sum as a physical consistency loss function;

[0029] The dynamic correlation is adjusted, a nonlinear transformation is performed on the adjusted dynamic correlation by using a multilayer perceptron, the dynamic correlation between system components is modeled, and parameters of the multilayer perceptron are optimized according to the physical consistency loss function, to obtain a dynamic correlation graph.

[0030] In an optional embodiment,

[0031] The electrolytic cell power allocation scheme is calculated according to the power prediction data, the efficiency prediction data, the dynamic balance equation and the surplus power value, and includes:

[0032] The power prediction data and the efficiency prediction data are input as calculation parameters into the dynamic balance equation, and the surplus power value is set as a dynamic balance constraint condition;

[0033] The dynamic balance equation is solved based on a nonlinear programming algorithm, and optimal operation parameters satisfying the system dynamic balance constraint and the surplus power constraint are obtained through iterative calculation;

[0034] The power upper and lower limit constraints of each electrolytic cell are determined according to the optimal operation parameters, the operation efficiency of each electrolytic cell is calculated based on the efficiency prediction data, the electrolytic cells are allocated power in the order from high to low operation efficiency, and an electrolytic cell power allocation scheme is obtained.

[0035] In an optional embodiment,

[0036] A deep reinforcement learning algorithm based on a Gaussian process is used, an environment uncertainty is modeled and a system risk is predicted through a Bayesian optimization framework, and target power parameters and control parameters of each electrolytic cell unit are calculated by combining an online dynamic programming method, and include:

[0037] A Gaussian process regression method is used to model the environment uncertainty, a function value corresponding to any point in the input space is set to follow a Gaussian distribution, a radial basis function is selected as a kernel function, variance parameters and length scale parameters of the kernel function are optimized by maximizing a marginal likelihood function, and a Gaussian process regression model is constructed;

[0038] A Bayesian optimization framework is constructed based on the Gaussian process regression model, a prediction distribution of a new input point is calculated, a system risk function is obtained according to a mean and a variance of the prediction distribution, and the system risk function includes a risk occurrence probability term and an expected loss term;

[0039] A deep reinforcement learning algorithm is constructed for power optimization, operation parameters and environment characteristics of each electrolytic cell unit are set as a state space, a power adjustment instruction is set as an action space, the deep reinforcement learning algorithm is trained based on experience replay and target network technology, and a reward function including a system efficiency term, a power deviation term and a risk loss term is used in the training process;

[0040] The target power parameters and control parameters of each electrolytic cell unit are calculated by using an online dynamic programming method, and the electrolyte concentration, electrolysis temperature and current density of each electrolytic cell unit are dynamically adjusted according to the control parameters.

[0041] In an alternative embodiment,

[0042] The target power parameters and control parameters of each electrolytic cell unit are calculated by using an online dynamic programming method, and the electrolyte concentration, electrolysis temperature and current density of each electrolytic cell unit are dynamically adjusted according to the control parameters.

[0043] The target power parameters and control parameters of each electrolytic cell unit are calculated by using an online dynamic programming method, and the electrolyte concentration, electrolysis temperature and current density of each electrolytic cell unit are dynamically adjusted according to the control parameters.

[0044] The influence of the target power parameters and control parameters on the operating state is obtained by performing a difference operation on the system output under different intervention conditions according to the conditional probability distribution.

[0045] The mapping function is constructed to map the target power parameters and control parameters to the latent space, and it is judged whether the output of the mapping function satisfies the manifold space constraint condition. If not, the target power parameters and control parameters are adjusted and the mapping operation is performed again until the constraint condition is satisfied. Then the Jacobian determinant is calculated using the initial probability distribution in the latent space, and the probability distribution of the system state is calculated according to the Jacobian determinant.

[0046] The key target power parameters and control parameters are selected according to the influence degree, and the probability representation is substituted into the conditional probability distribution for iterative calculation to output the target power parameters and control parameters of each electrolytic cell unit.

[0047] The manifold space constraint conditions include single electrolytic cell power upper limit constraint, single electrolytic cell power lower limit constraint, total power balance constraint, power change rate constraint, adjacent time power difference constraint, current density range constraint, temperature range constraint, pressure range constraint, voltage fluctuation constraint and current fluctuation constraint.

[0048] In an alternative embodiment,

[0049] The optimal power distribution scheme is obtained by updating the electrolytic cell power distribution scheme based on the power prediction data and efficiency prediction data.

[0050] When the actual energy utilization efficiency value is lower than the preset threshold, the historical operation data of the electrolytic cell unit is obtained, including historical power data, historical efficiency data and historical operation condition data. The prediction model is trained by a deep learning algorithm to generate power prediction data and efficiency prediction data for the next time period.

[0051] construct a target optimization function based on the power prediction data and the efficiency prediction data, set the system overall efficiency as an optimization target, take the power distribution amount of each electrolysis cell unit as an optimization variable, combine the power balance constraint and the safe operation constraint as constraint conditions, solve the target optimization function through an iterative optimization algorithm to obtain an optimal power distribution scheme.

[0052] In a second aspect of the embodiment of the present application, a power dynamic balance regulation system of a renewable energy hydrogen production-based energy source is provided, comprising:

[0053] A first unit is configured to acquire real-time power generation data of a renewable energy power generation system, real-time energy storage state data of an energy storage device, and power grid load demand data, establish a dynamic balance equation and calculate a surplus power value available for hydrogen production, and collect equipment operation parameters of a hydrogen production device to calculate a current operation efficiency value of the hydrogen production device;

[0054] A second unit is configured to construct a multi-scale spatiotemporal attention network, extract multi-dimensional time sequence features through an adaptive convolution kernel, model dynamic correlation between system components by using a location-weighted attention mechanism, and calculate power prediction data and efficiency prediction data for the next 24 hours;

[0055] A third unit is configured to calculate an electrolysis cell power distribution scheme according to the power prediction data, the efficiency prediction data, the dynamic balance equation, and the surplus power value;

[0056] A fourth unit is configured to calculate target power parameters and control parameters of each electrolysis cell unit by using a deep reinforcement learning algorithm based on a Gaussian process, modeling environmental uncertainty and predicting system risks through a Bayesian optimization framework, and combining an online dynamic programming method based on the electrolysis cell power distribution scheme;

[0057] A fifth unit is configured to monitor actual energy utilization efficiency values of each electrolysis cell unit under the control parameters, and update the electrolysis cell power distribution scheme based on the power prediction data and the efficiency prediction data to obtain an optimal power distribution scheme when the actual energy utilization efficiency values are lower than a preset threshold.

[0058] In a third aspect of the embodiment of the present application,

[0059] An electronic device is provided, comprising:

[0060] a processor;

[0061] a memory for storing processor-executable instructions;

[0062] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0063] In the present application, by constructing a multi-scale spatio-temporal attention network, the dynamic correlation between system components is accurately modeled, the accuracy of power and efficiency prediction is significantly improved, the Gaussian process is combined with deep reinforcement learning, the quantitative representation of environmental uncertainty is realized, the real-time optimization of power allocation scheme is ensured, the system operation risk is effectively reduced, and the system is always maintained in the optimal operating state through dynamic updating of the power allocation scheme. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 A flowchart of the power dynamic balance regulation method of the renewable energy hydrogen-based energy source of the embodiment of the present application is shown in

[0065] Figure 2 A causal effect analysis accuracy comparison chart of the power dynamic balance regulation method of the renewable energy hydrogen-based energy source of the embodiment of the present application is shown in

[0066] Figure 3 A multi-dimensional performance index radar chart of the power dynamic balance regulation method of the renewable energy hydrogen-based energy source of the embodiment of the present application is shown in

[0067] Figure 4 An electrolytic cell parameter and performance index comparison chart of the power dynamic balance regulation method of the renewable energy hydrogen-based energy source of the embodiment of the present application is shown in DETAILED DESCRIPTION

[0068] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0069] The technical scheme of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0070] Figure 1 A flowchart of the power dynamic balance regulation method of the renewable energy hydrogen-based energy source of the embodiment of the present application is shown in Figure 1 As shown in the figure, the method comprises:

[0071] Real-time power generation data of a renewable energy power generation system, real-time energy storage state data of an energy storage device, and power grid load demand data are acquired, a dynamic balance equation is established, and a surplus power value available for hydrogen production is calculated, equipment operation parameters of a hydrogen production device are collected, and a current operation efficiency value of the hydrogen production device is calculated;

[0072] A multi-scale spatiotemporal attention network is constructed, multi-dimensional time series features are extracted through adaptive convolution kernels, a position-weighted attention mechanism is used to model the dynamic correlation between system components, and power prediction data and efficiency prediction data for the next 24 hours are calculated;

[0073] An electrolyzer power distribution scheme is calculated according to the power prediction data, the efficiency prediction data, the dynamic balance equation, and the surplus power value;

[0074] Based on the electrolyzer power distribution scheme, a deep reinforcement learning algorithm based on a Gaussian process is used, the environmental uncertainty is modeled through a Bayesian optimization framework, and the system risk is predicted, the target power parameters and control parameters of each electrolyzer unit are calculated by combining an online dynamic programming method, and the operation parameters of each electrolyzer unit are dynamically adjusted according to the control parameters;

[0075] The actual energy utilization efficiency value of each electrolyzer unit under the control parameters is monitored, and when the actual energy utilization efficiency value is lower than a preset threshold, the electrolyzer power distribution scheme is updated based on the power prediction data and the efficiency prediction data to obtain an optimal power distribution scheme.

[0076] In an alternative embodiment,

[0077] A dynamic balance equation is established, and a surplus power value available for hydrogen production is calculated, and equipment operation parameters of a hydrogen production device are collected to calculate a current operation efficiency value of the hydrogen production device, including:

[0078] The real-time power generation data, the real-time energy storage state data, and the power grid load demand data are substituted into a dynamic balance equation, wherein the dynamic balance equation represents that the real-time power generation data is equal to the sum of the power grid load demand data, the change in the charging and discharging power of the energy storage device, and the surplus power value, and the surplus power value available for hydrogen production is calculated according to the dynamic balance equation;

[0079] Temperature parameters, pressure parameters, current parameters, and hydrogen production parameters of a hydrogen production device are collected, the product of the hydrogen production parameters and the high heating value of hydrogen is taken as the numerator, the product of the current parameters and the voltage of the electrolyzer of the hydrogen production device is taken as the denominator, and the current operation efficiency value of the hydrogen production device is calculated according to the ratio of the numerator to the denominator.

[0080] Real-time operation data of the power system is acquired by the configured data acquisition module. On the renewable energy power generation side, real-time power generation data of power generation equipment such as wind farms and photovoltaic power stations is acquired; on the energy storage device side, real-time charging and discharging state and power data of equipment such as battery packs and energy storage converters are acquired; on the power grid side, power consumption data of each metering point is acquired.

[0081] To achieve dynamic balance of system power, a dynamic balance equation is established, which reflects the relationship between real-time power generation of renewable energy and each power consumption link, i.e. real-time power generation equals the sum of power grid load demand, energy storage device charging and discharging power variation, and surplus power available for hydrogen production. By substituting the acquired real-time power generation data, energy storage state data and load demand data into the equation, the surplus power value available for hydrogen production at the current time can be obtained through calculation.

[0082] In terms of calculation of the operation efficiency of the hydrogen production device, operation parameters are acquired by sensors installed at key positions of the hydrogen production device, including temperature parameters acquired by a temperature sensor, pressure parameters acquired by a pressure sensor, current parameters acquired by a current sensor, and hydrogen production parameters acquired by a hydrogen flow meter. The acquired hydrogen production parameters are multiplied by a predetermined high heating value of hydrogen to obtain a hydrogen production energy value. The acquired current parameters are multiplied by a real-time operating voltage of an electrolytic cell of the hydrogen production device to obtain an input electric energy value. The operation efficiency value of the hydrogen production device at the current time can be obtained by dividing the hydrogen production energy value by the input electric energy value.

[0083] For example, the operation data of a certain new energy hydrogen production plant at a certain time shows that the total power generation of the wind farm and the photovoltaic power station is 10,000 kW, the actual power consumption load of the power grid user is 6,000 kW, and the energy storage device is charging at a charging power of 2,000 kW. Substituting these data into the dynamic balance equation, it is calculated that the surplus power available for hydrogen production at the current time is 2,000 kW.

[0084] At the same time, the operation parameters of the hydrogen production device show that the equipment operating temperature is 75 degrees, the system pressure is 3 MPa, the electrolytic cell operating current is 2,000 amperes, and the measured hydrogen production is 400 standard cubic meters per hour. The high heating value of hydrogen is 3.54 kW·h per standard cubic meter, and the operating voltage of the electrolytic cell is 2.1 volts. According to these parameters, it is calculated that the product of hydrogen production and high heating value is 1,416 kW·h, the product of current and voltage is 4.2 kW·h, and the final operation efficiency of the hydrogen production device at the current time is 67.4%.

[0085] In this embodiment, by deploying a comprehensive data acquisition module, real-time data acquisition of renewable energy power generation systems, energy storage devices and grid loads is realized, which can timely reflect the system power changes and provide accurate data support for subsequent control. The establishment of the dynamic balance equation organically integrates key elements such as power generation, load demand, energy storage changes and hydrogen production power, and builds a complete power balance system, which can accurately reflect the current power resources available for hydrogen production in the system, avoiding the problem of unreasonable power distribution caused by traditional experience estimation method. Considering the direct relationship between power input and hydrogen output and the actual loss in the energy conversion process, the efficiency evaluation result is more objective and accurate.

[0086] In an alternative embodiment,

[0087] A multi-scale spatio-temporal attention network is constructed, multi-dimensional time series features are extracted through adaptive convolution kernels, and a position-weighted attention mechanism is used to model the dynamic correlation between system components. The power prediction data and efficiency prediction data for the next 24 hours are calculated, including:

[0088] A multi-scale spatio-temporal attention network is constructed, and the adaptive convolution kernel is used to convolve the input time series data to obtain feature data. The feature data corresponding to the adaptive convolution kernel is spliced to obtain multi-dimensional time series features.

[0089] Based on the multi-dimensional time series features, a query matrix and a key matrix are constructed. The product of the query matrix and the key matrix is added to the position weighting matrix, and then normalized to obtain an attention weight matrix. The attention weight matrix is multiplied by the value matrix of the multi-dimensional time series features to obtain attention output data.

[0090] The attention output data is nonlinearly transformed using a multi-layer perceptron to model the dynamic correlation between system components and obtain a dynamic correlation graph.

[0091] The dynamic correlation graph is input into the corresponding decoder network with power historical data and efficiency historical data, respectively. The power prediction data and efficiency prediction data for the next 24 hours are calculated. Based on the mean square error of the power prediction data and the actual power data, and the mean square error of the efficiency prediction data and the actual efficiency data, a joint loss function is constructed. Based on the joint loss function, the parameters of the decoder network are optimized.

[0092] The construction of the multi-scale spatio-temporal attention network starts with feature extraction of the input time series data. By setting adaptive convolution kernels with different observation windows, short-term, medium-term, and long-term time series variation features are captured respectively. Each convolution kernel can automatically adjust its weight parameters according to the characteristics of the data to extract the most representative feature information. The convolution kernel processes the input data simultaneously, and the convolution operation obtains feature representations of different time scales. The features extracted at different scales are spliced and integrated to form a comprehensive feature representation containing multiple time dimensions, which fully reflects the running characteristics at different time scales.

[0093] After obtaining the multi-dimensional time series features, the attention calculation module maps the feature data into a query matrix and a key matrix, which represent the target information and reference information that need to be focused on, respectively. A position weighting matrix is also constructed to represent the importance of different time positions. The query matrix is multiplied by the key matrix to obtain the preliminary correlation strength, which is then added to the position weighting matrix to consider the influence of time position information. After normalization, the attention weight matrix is obtained, which reflects the correlation degree between different time points and features. This attention weight matrix is multiplied by the value matrix of the original features to obtain the attention output data that integrates time position information.

[0094] To further explore the dynamic association between components, a multi-layer perceptron is used to extract deep features from the attention output data. By designing multiple hidden layers, each layer uses a different activation function to perform non-linear transformation on the features, gradually extracting and strengthening the association features between components. Through multiple levels of feature conversion, a dynamic association graph is output, which fully describes the interaction relationships between components at different time scales.

[0095] In the prediction stage, two independent decoder networks perform power prediction and efficiency prediction tasks respectively. These decoders receive the dynamic association graph and historical data as input, perform feature decoding through a multi-layer neural network, and output the prediction results for the next 24 hours. The construction of the joint loss function considers the mean square error of both power prediction and efficiency prediction, and by calculating the error between the predicted value and the actual value, the network parameters are continuously adjusted and optimized to improve the prediction accuracy.

[0096] Exemplarily, in the prediction practice of a new energy hydrogen production plant, a week of historical operation data is collected, including time series data such as power generation, load demand, and energy storage state. With a basic sampling interval of 15 minutes, three observation window convolution kernels of 1 hour, 6 hours, and 24 hours are set for feature extraction. Each convolution kernel outputs 100 feature dimensions, and the features of the three scales are spliced to form a 300-dimensional feature vector. These features are processed by the attention mechanism, successfully identifying the daily cycle change characteristics of photovoltaic power generation and the peak-valley characteristics of electricity demand. The dynamic correlation modeling accurately captures the delay response relationship between the energy storage device and the power generation, as well as the influence law of load changes on hydrogen production power. In the prediction link, the power and efficiency of the next 24 hours are predicted, and the final power prediction error is controlled within 5%, and the efficiency prediction error is controlled within 3%, providing accurate decision-making basis for intelligent regulation.

[0097] In this embodiment, the feature extraction mechanism of the multi-scale spatio-temporal attention network significantly improves the data processing capability. The feature extraction method based on attention weight can adaptively highlight the influence of important time nodes, effectively filter redundant information, and improve the perception ability of the prediction model to key time sequence patterns. The deep feature extraction mechanism breaks through the expression limitation of traditional linear models, can discover and utilize the potential laws in the data, and makes the prediction results more consistent with the dynamic characteristics of the actual system.

[0098] In an alternative embodiment,

[0099] The multi-layer perception is used to perform nonlinear transformation on the attention output data, and the dynamic correlation between system components is modeled to obtain a dynamic correlation graph including:

[0100] The attention output data is obtained, and a causal graph is constructed based on the attention output data. The causal graph includes a node set, an edge set, and a causal probability matrix. The features of the attention output data are taken as the node set, the correlations between the attention output data are taken as the edge set, and the dynamic correlations between the attention output data are taken as the causal probability matrix.

[0101] The multi-layer perception is used to perform counterfactual intervention on the attention output data under different intervention conditions, calculate the expected output under different intervention conditions, and calculate the difference between the expected outputs under different intervention conditions to obtain the causal effect strength.

[0102] Based on the causal effect strength, the attention output data is subjected to nonlinear transformation, and the result of the nonlinear transformation is multiplied by the output value of the sigmoid function to obtain the adjusted dynamic correlation.

[0103] The physical characteristic parameters and motion data corresponding to the system components are constructed, the kinematic constraint equation, the mass conservation equation and the energy conservation equation are constructed based on the physical characteristic parameters and motion data, and the physical constraint equation set is obtained by combination, the adjusted dynamic correlation is input into the physical constraint equation set, the norm value of the physical constraint equation set and the norm value of the gradient of the physical constraint equation set are calculated, the weighted sum is generated according to the norm value of the physical constraint equation set and the norm value of the gradient of the physical constraint equation set, and the weighted sum is used as a physical consistency loss function;

[0104] The adjusted dynamic correlation is subjected to nonlinear transformation by a multilayer perceptron, the dynamic correlation between the system components is modeled, and the parameters of the multilayer perceptron are optimized according to the physical consistency loss function, so as to obtain a dynamic correlation graph.

[0105] After obtaining the attention output data, the construction of the causal graph is performed. Each feature in the attention output data is taken as a node of the causal graph, and the features include various key indicators of system operation. In order to determine the correlation between nodes, the correlation coefficient between any two features needs to be calculated. When the absolute value of the correlation coefficient exceeds a pre-set threshold, it is considered that there is a causal correlation between the two nodes, and the connection relationship of the edge is established accordingly. At the same time, a causal probability matrix is constructed based on the weight values calculated by the attention mechanism. Each element in the matrix represents the causal influence probability between the corresponding nodes. Through multiple iterations of calculation, the values in the causal probability matrix are constantly updated and optimized until the probability distribution tends to be stable or the pre-set iteration number is reached.

[0106] When performing counterfactual intervention analysis, a multilayer perceptron network structure is designed to simulate the intervention effect. For each feature of the attention output data, multiple intervention conditions of different degrees are designed. When each intervention condition is applied, the other features remain unchanged. The feature data after intervention is input into the multilayer perceptron network, and the change of system output is observed. By comparing the output difference before and after intervention, the causal effect intensity of the feature under the intervention condition is calculated. Repeat the analysis of all features under different intervention conditions to obtain a complete causal effect intensity matrix. This matrix comprehensively reflects the causal influence degree between each feature in the system.

[0107] Based on the calculated causal effect intensity matrix, a nonlinear transformation network is designed for feature conversion. The network adopts a multilayer structure, and a nonlinear activation function is used between each layer to enhance the feature expression ability. The causal effect intensity matrix is input into the network to obtain the converted feature matrix. At the same time, the attention output data is subjected to nonlinear transformation to obtain a probability distribution matrix. Multiply the two matrices to obtain an adjusted dynamic correlation matrix. The dynamic correlation matrix not only retains the correlation information between the original features, but also incorporates the results of causal analysis.

[0108] In the physical constraint modeling section, the physical characteristic parameters and motion data of each component of the system are collected. Based on these data, a complete set of physical constraint equations is constructed, including kinematic constraint equations, mass conservation equations, and energy conservation equations. These equations reflect the basic physical laws that the system must follow. The adjusted dynamic correlation is substituted into the equations, and the norm value of the equation set and the norm value of its gradient are calculated. By setting the weight coefficient, the weighted sum of these two norm values is constructed into a physical consistency loss function. This loss function is used to evaluate the degree of conformity of the prediction results with the physical laws.

[0109] A multi-layer perception network is constructed for dynamic correlation modeling. The multi-layer perception network contains multiple hidden layers, each using a different type of activation function to enhance the network's expression ability. The adjusted dynamic correlation matrix is input into the network, and the prediction output is obtained through forward propagation. During the training process, the physical consistency loss function is used as the optimization target, and the network parameters are continuously adjusted using optimization algorithms. The gradient of the loss function with respect to the network parameters is calculated each iteration, and the parameters are updated according to the preset learning rate. When the loss function converges or reaches the maximum number of iterations, the final dynamic correlation graph is output.

[0110] Exemplarily, taking the operation optimization of a certain hydrogen production plant as an example, first, the system operation data is collected, including photovoltaic power generation power (0-100MW), energy storage charging and discharging power (-50MW to 50MW), electrolyzer power (0-80MW), etc. When constructing the causal graph, these features are used as nodes, and the correlation strength calculated by the attention mechanism is used to construct the initial causal probability matrix.

[0111] In the counterfactual intervention analysis, ±20% perturbations are applied to the photovoltaic power generation power, and the response changes of other features are observed. For example, when the photovoltaic power increases by 20%, the energy storage charging power increases by an average of 15MW, and the electrolyzer power increases by 12MW; when the photovoltaic power decreases by 20%, the energy storage discharging power increases by an average of 10MW, and the electrolyzer power decreases by 8MW. The causal effect strength matrix is calculated based on these response data.

[0112] In the physical constraint modeling, the energy balance equation of the electrolyzer is established: input electric power = hydrogen production x unit hydrogen heat value / electrolysis efficiency. At the same time, the charging and discharging efficiency constraints of the energy storage system are considered: charging efficiency 85%, discharging efficiency 90%. These physical constraints are converted into mathematical expressions to construct the physical consistency loss function.

[0113] The optimized dynamic correlation diagram shows that the correlation strength between photovoltaic power generation power and energy storage charging and discharging power is 0.8, and the correlation strength with electrolytic tank power is 0.6. These values not only reflect the statistical law of data, but also meet the physical constraints such as energy conservation. The dynamic correlation diagram successfully guides the system operation optimization, making the hydrogen production efficiency increase by 15%, and the energy storage utilization rate increase by 20%.

[0114] In this embodiment, by constructing a multi-scale spatio-temporal attention network, the causal graph structure is innovatively introduced on the basis of obtaining attention output data, which not only represents the correlation between features, but also quantifies the influence degree between features through a causal probability matrix, making the interaction relationship between system components more clear and interpretable. By introducing a counterfactual intervention analysis mechanism, different degrees of intervention conditions are designed, and the intervention effect is simulated by using a multi-layer perception machine, which can deeply understand the causal effect strength of each feature in the system. A physical constraint mechanism is integrated into the dynamic correlation modeling, a set of physical constraint equations containing kinematic constraints, mass conservation and energy conservation are constructed, and they are converted into loss functions to ensure that the prediction results meet the basic physical laws.

[0115] In the prior art, the modeling of dynamic correlation between system components mainly relies on statistical methods and simple neural network models, which often only focus on the surface features of data and cannot deeply mine the causal relationship between system components. It also fails to fully consider the constraints of physical laws. Traditional methods have problems in capturing long-range dependencies when dealing with long time series data, and are easily affected by data noise, resulting in insufficient reliability and accuracy of the prediction results.

[0116] The embodiment of the present application combines causal analysis, physical constraints and deep learning methods to significantly improve the effect of dynamic correlation modeling between system components, providing reliable technical support for intelligent optimization control of the system. It not only improves the accuracy and reliability of the prediction, but also enhances the interpretability and practicality of the model, providing a new research idea for the development of related fields.

[0117] Figure 2The causal effect analysis accuracy comparison chart of the power dynamic balance regulation method of the renewable energy hydrogen-based energy source of the embodiments of the present application shows the response characteristics of the system under different intervention conditions. The technical solution (square mark) accurately predicts the change of the energy storage charging and discharging power under the condition that the photovoltaic power changes ±20% by combining the multilayer perception with the physical constraint. When the photovoltaic power increases by 20%, the prediction accuracy of the 15MW increase in the energy storage charging power reaches 95.8%, while the prediction accuracy of the traditional VAR model (circular mark) and the TCN model (triangular mark) is 83.2% and 87.5% respectively. Under extreme working conditions (intervention intensity 25%), the scheme still maintains an accuracy of 93.5%, which is 8.7 percentage points higher than the second place, reflecting the robustness of the scheme under large-scale working condition changes.

[0118] In an alternative embodiment,

[0119] The electrolyzer power allocation scheme is calculated according to the power prediction data, the efficiency prediction data, the dynamic balance equation and the surplus power value, which comprises:

[0120] The power prediction data and the efficiency prediction data are inputted into the dynamic balance equation as calculation parameters, and the surplus power value is set as a dynamic balance constraint condition;

[0121] The dynamic balance equation is solved based on a nonlinear programming algorithm, and the optimal operation parameters meeting the system dynamic balance constraint and the surplus power constraint are obtained through iterative calculation;

[0122] The power upper and lower limit constraints of each electrolyzer are determined according to the optimal operation parameters, the operation efficiency of each electrolyzer is calculated based on the efficiency prediction data, the electrolyzers are allocated power in the order from high to low operation efficiency, and the electrolyzer power allocation scheme is obtained.

[0123] First, the power prediction data and the efficiency prediction data are inputted into the dynamic balance equation as calculation parameters. The power prediction data includes time series data such as renewable energy power generation, load demand power, energy storage device charging and discharging power, etc. The efficiency prediction data includes performance parameters such as the operation efficiency of the electrolyzer group and the charging and discharging efficiency of the energy storage device. When inputting these data, the data needs to be preprocessed, including data standardization, outlier processing and time alignment operations, to ensure the quality and consistency of the data. At the same time, the surplus power value of the system is set as the constraint condition of the dynamic balance, which requires that the difference between the power generation and the power consumption of the system at any time does not exceed the preset threshold range.

[0124] The nonlinear programming algorithm is used to solve the dynamic balance equation, and the objective function is constructed, which needs to consider multiple optimization objectives such as maximizing system operation efficiency and minimizing energy loss, and set constraint conditions, including power balance constraint, device operation constraint, and energy storage capacity constraint. In the solving process, an iterative calculation method is used, and the system state variables are updated at each iteration until the optimal solution that satisfies all constraint conditions is found. Initialize system parameters, calculate the value of the objective function and the satisfaction degree of the constraint conditions in each iteration, and adjust the optimization variables according to the calculation results until the algorithm converges or the maximum iteration number is reached.

[0125] Based on the calculated optimal operating parameters, the power allocation scheme of the electrolyzer group is determined, and the upper and lower power constraints of each electrolyzer are determined according to the optimal operating parameters. Technical parameters such as rated power and minimum starting power of the electrolyzer need to be considered, and the operating efficiency of each electrolyzer under different working conditions is calculated based on the efficiency prediction data. The electrolyzers are sorted in descending order of operating efficiency, and the electrolyzers with higher efficiency are preferentially allocated power to maximize the overall efficiency of the system. In the allocation process, the start-stop switching cost and power change rate of the electrolyzer need to be considered to ensure the implementability of the allocation scheme.

[0126] Exemplarily, a certain hydrogen production plant is taken as an example, which includes a photovoltaic power generation system, an energy storage system, and multiple electrolyzers. Receive 24-hour power prediction data, including photovoltaic power prediction values, grid load prediction values, and other time series data. At the same time, obtain efficiency prediction data, including efficiency prediction values of each electrolyzer under different load rates, and charging and discharging efficiency of the energy storage system.

[0127] In the dynamic balance constraint setting, the fluctuation range of the system's surplus power value is specified. When photovoltaic power generation fluctuates, the charging and discharging power of the energy storage system and the operating power of the electrolyzer are adjusted to ensure the dynamic balance of the system. The nonlinear programming algorithm is used for optimization and solution, and through multiple iterations, the optimal operating parameters that satisfy all constraint conditions are obtained.

[0128] Based on the optimal operating parameters, the operating power range of each electrolyzer is determined. By analyzing the efficiency prediction data, the operating efficiency of each electrolyzer under the current working condition is calculated. After sorting the electrolyzers by efficiency, more operating power is allocated to high-efficiency electrolyzers, and the power allocated to low-efficiency electrolyzers is reduced or the electrolyzers are put into standby state. This allocation scheme not only ensures the dynamic balance of the system, but also improves the overall operating efficiency.

[0129] In this embodiment, the power prediction data and the efficiency prediction data are used as calculation parameters, and a complete system dynamic balance model is established by combining the surplus power constraint condition. The mutual influence between the components of the system is fully considered, so that the optimization result can accurately reflect the actual running state of the system. The iterative solving mechanism of the nonlinear programming algorithm ensures the convergence and reliability of the optimization result. Through the synergistic effect of multi-objective optimization and multiple constraint conditions, the system can maximize the running efficiency while ensuring the dynamic balance.

[0130] In an alternative embodiment,

[0131] A deep reinforcement learning algorithm based on a Gaussian process is used to model the environmental uncertainty and predict the system risk through a Bayesian optimization framework. The target power parameters and control parameters of each electrolytic cell unit are calculated by combining an online dynamic programming method, including:

[0132] A Gaussian process regression method is used to model the environmental uncertainty. The function value corresponding to any point in the input space is set to follow a Gaussian distribution. A radial basis function is selected as the kernel function. The variance parameter and length scale parameter of the kernel function are optimized by maximizing the marginal likelihood function, and a Gaussian process regression model is constructed.

[0133] A Bayesian optimization framework is constructed based on the Gaussian process regression model to calculate the prediction distribution of a new input point. The system risk function is obtained according to the mean and variance of the prediction distribution. The system risk function includes a risk occurrence probability term and an expected loss term.

[0134] A deep reinforcement learning algorithm is constructed for power optimization. The operating parameters of each electrolytic cell unit and the environmental characteristics are set as the state space, and the power adjustment instructions are set as the action space. The deep reinforcement learning algorithm is trained based on experience replay and target network technology. A reward function containing a system efficiency term, a power deviation term, and a risk loss term is used during the training process.

[0135] An online dynamic programming method is used to calculate the target power parameters and control parameters of each electrolytic cell unit. The electrolyte concentration, electrolytic temperature, and current density of each electrolytic cell unit are dynamically adjusted according to the control parameters.

[0136] The Gaussian process regression method is used to model the environmental uncertainty, and the input space dimension is determined, including environmental parameters such as temperature, humidity, and light intensity. The environmental parameters are used as input variables. The data of each input point is preprocessed, including data standardization and outlier processing. The radial basis function is selected as the kernel function, which includes two optimization parameters: the variance parameter and the length scale parameter. The marginal likelihood function is constructed as the optimization objective function. The gradient ascent method is used to optimize the marginal likelihood function, and the optimal parameter values of the kernel function are obtained through iterative calculation. Based on the optimized kernel function parameters, the complete Gaussian process regression model is constructed.

[0137] Based on the Gaussian process regression model, a Bayesian optimization framework is constructed to calculate the prediction distribution of any new input point, and the prediction mean and variance are obtained. Based on the prediction distribution, a system risk function is constructed, which considers both the risk occurrence probability and the expected loss. The risk occurrence probability is calculated by the cumulative distribution function of the prediction distribution, and the expected loss is calculated based on the predefined loss function. The two parts are combined to form a complete system risk assessment model.

[0138] The state space is defined, including electrolytic cell operating parameters (power, efficiency, temperature, etc.) and environmental characteristic parameters. The action space is defined as a set of power adjustment commands, including different magnitudes of power adjustment. An experience replay pool is constructed to store training samples. An evaluation network and a target network are established, which have the same network structure but different parameter update frequencies. A reward function is designed, which includes a system efficiency term, a power deviation term, and a risk loss term. During training, the agent interacts with the environment, and the obtained experience is stored in the experience replay pool. Samples are randomly sampled from the experience replay pool for training at regular intervals. The target network is used to calculate the target Q value, and the evaluation network parameters are updated through backpropagation. The evaluation network parameters are periodically copied to the target network.

[0139] The online dynamic programming method is used to calculate the control parameters. According to the current system state and the output of the deep reinforcement learning algorithm, the target power parameters of each electrolytic cell are calculated. Based on the target power parameters, the dynamic characteristic equation of the electrolytic cell is used to calculate the adjustment amount of the electrolyte concentration, the electrolytic temperature, and the current density. The adjustment amount is constrained to ensure that it meets the equipment operation limits. The processed adjustment amount is converted into specific control commands and sent to the actuator to execute the adjustment operation.

[0140] For example, in a certain photovoltaic hydrogen production plant, environmental data is collected, including hourly temperature, humidity, light intensity, and other parameters. After preprocessing these data, they are used to train the Gaussian process regression model. By maximizing the marginal likelihood function, the optimal variance parameter and length scale parameter of the kernel function are obtained.

[0141] Based on the trained Gaussian process regression model, the system predicts the environmental conditions in the future period. For example, predicting the change of light intensity in the next 4 hours. According to the prediction distribution, the probability of risk occurrence is calculated, and the expected loss is calculated combined with the predefined loss function, forming a complete risk assessment result.

[0142] In the deep reinforcement learning algorithm, the operating parameters of the electrolytic cell such as power, efficiency, temperature, and environmental characteristic parameters are combined into a state vector. The action space is defined as [-10%, -5%, 0, +5%, +10%] five power adjustment levels. A deep neural network with multiple hidden layers is constructed as the evaluation network and the target network. The reward function is designed to combine the system efficiency, power deviation, and risk loss according to the preset weight.

[0143] According to the output results of the deep reinforcement learning algorithm, the target power values of each electrolytic cell are calculated. Based on the target power values, combined with the dynamic characteristics of the electrolytic cell, the adjustment instructions of the electrolyte concentration, electrolysis temperature and current density are calculated. For example, when the electrolytic cell power needs to be increased, the system will simultaneously increase the electrolyte concentration and current density, and adjust the electrolysis temperature accordingly to ensure the stable operation of the electrolytic cell under the new working condition.

[0144] In this embodiment, by using radial basis function as kernel function and combining marginal likelihood function optimization, the model can accurately capture the dynamic change characteristics of environmental parameters. The probability-based risk assessment method can effectively identify potential operating risks and provide decision support for safe operation of the system. By calculating the target power parameters and control parameters of each electrolytic cell unit in real time, the system can quickly respond to environmental changes and load demands, ensuring efficient and stable operation of the electrolytic cell under various working conditions.

[0145] In an alternative embodiment,

[0146] The target power parameters and control parameters of each electrolytic cell unit are calculated by using online dynamic programming method, including:

[0147] The target power parameters and control parameters of the electrolytic cell unit are obtained as independent variables, and the operating state is obtained as dependent variables. The intervention operation is performed on the independent variables and dependent variables, and the conditional probability distribution is calculated;

[0148] According to the conditional probability distribution, counterfactual calculation is performed, and the difference operation is performed on the system output under different intervention conditions to obtain the influence degree of the target power parameters and control parameters on the operating state;

[0149] A mapping function is constructed to map the target power parameters and control parameters to a latent space, it is judged whether the output of the mapping function satisfies a manifold space constraint condition, if not, the target power parameters and control parameters are adjusted and the mapping operation is re-executed until the constraint condition is satisfied, then the Jacobian determinant is calculated using the initial probability distribution in the latent space, the probability distribution is calculated according to the Jacobian determinant, and the probability representation of the system state is obtained;

[0150] According to the influence degree, the key target power parameters and control parameters are screened, the probability representation is substituted into the conditional probability distribution for iterative calculation, and the target power parameters and control parameters of each electrolytic cell unit are output;

[0151] The manifold space constraint condition includes single electrolytic cell power upper limit constraint, single electrolytic cell power lower limit constraint, total power balance constraint, power change rate constraint, adjacent time power difference constraint, current density range constraint, temperature range constraint, pressure range constraint, voltage fluctuation constraint and current fluctuation constraint.

[0152] The target power parameters (including power set value, power adjustment range, etc.) and control parameters (including electrolyte concentration, temperature set value, pressure set value, etc.) of the electrolytic cell unit are collected as independent variables, and the operating state data (including actual power, efficiency, temperature, etc.) are collected as dependent variables. Intervention operation is performed on the dependent variables, including: setting different intervention conditions, changing the value of the independent variable, observing the change of the dependent variable, and calculating the conditional probability distribution accordingly. The intervention operation needs to consider the causal relationship between variables to ensure the rationality of the intervention;

[0153] Based on the obtained conditional probability distribution, the system outputs under different intervention conditions are compared and analyzed, a reference scene is set, and a comparison scene is generated by changing one or more parameters. The difference between the system outputs under different scenes is calculated. Through this difference operation, the influence degree of the target power parameters and control parameters on the operating state can be quantified. This process needs to consider the interaction between parameters to ensure the accuracy of the calculation result;

[0154] A mapping function is constructed to map the target power parameters and control parameters to a latent space. The mapping function needs to maintain the essential characteristics of the data while reducing the dimensionality of the data. Then it is judged whether the mapping result satisfies the manifold space constraint condition, which includes power limit, change rate limit, process parameter range limit and other aspects. If the constraint condition is not satisfied, the original parameters need to be adjusted and the mapping operation is re-executed. When the constraint condition is satisfied, the Jacobian determinant is calculated using the initial probability distribution in the latent space, and the probability distribution is calculated based on the determinant to obtain the probability representation of the system state;

[0155] According to the pre-acquired influence degree, the key parameters that significantly affect the system running state are screened out, and these key parameters are combined with the probability representation obtained in the third step to substitute into the conditional probability distribution for iterative calculation. Through multiple iterations, the optimal target power parameters and control parameters of each electrolytic cell unit are finally output.

[0156] Exemplarily, taking 5 electrolytic cells of a certain hydrogen production plant as an example, the running data of each electrolytic cell is collected. The independent variables include: power set value (0-1 MW), electrolyte concentration (25-35%), operating temperature (65-85℃), operating pressure (0.1-3.0 MPa), etc. The dependent variables include: actual output power, system efficiency, hydrogen production, etc.

[0157] In the intervention operation, the parameters such as power set value (±10%), temperature set value (±5℃), pressure set value (±0.2 MPa) are changed respectively, and the system response is recorded. For example, when the power set value is increased from 0.8 MW to 0.88 MW, the efficiency change and hydrogen production change are observed, and the conditional probability distribution is calculated accordingly.

[0158] In the counterfactual calculation, the baseline working condition (power 0.8 MW, temperature 75℃, pressure 2.0 MPa) is set, and multiple groups of comparison working conditions are generated. By calculating the performance difference under different working conditions, the influence degree of each parameter is obtained. For example, it is found that the temperature change has an influence degree of 0.8 on efficiency and an influence degree of 0.6 on hydrogen production.

[0159] In the mapping function construction, a multi-layer neural network is used to map the original parameters to a low-dimensional space. It is checked whether the mapping result meets the constraint conditions, such as the upper limit of power 1 MW, the temperature range 65-85℃, etc. If some constraints are not met, the original parameters are adjusted and remapped. After meeting the constraints, the Jacobian determinant is calculated and the probability distribution is obtained.

[0160] According to the influence degree, the key parameters (such as power set value, temperature set value) are screened out, and the selected key parameters are combined with the probability representation for iterative calculation. After multiple iterations, the optimization parameters of each electrolytic cell are obtained: power set value, temperature set value, pressure set value, etc. These parameters not only meet various constraint conditions, but also ensure the efficient operation of the system.

[0161] In this embodiment, by introducing intervention operation and conditional probability distribution calculation, a causal relationship model between parameters is established, which can accurately identify the real influence of parameter changes on system performance, quantitatively evaluate the influence degree of each parameter by comparing the system response under different intervention conditions, and provide an objective measurement of parameter importance, which provides a scientific basis for subsequent parameter screening and optimization. By constructing a mapping function to map the parameter space to the hidden space and introducing multi-dimensional constraint conditions, it is ensured that the optimization results meet the physical limitations and process requirements of the system, and the implementability of the optimization results is significantly improved.

[0162] In the prior art, parameter optimization of the electrolytic cell system mainly relies on simple statistical analysis and empirical models, often only focuses on the surface correlation between parameters, fails to deeply mine the causal relationship between parameters, and cannot accurately evaluate the actual influence of parameter changes on system performance. When dealing with multi-parameter coupling optimization problems, it is easy to fall into local optimization, and it is difficult to ensure that the optimization results meet multiple constraint conditions.

[0163] The embodiment can effectively handle the complex interaction between parameters and has strong adaptability to system fluctuations, not only improving the accuracy and reliability of optimization, but also enhancing the explainability and practicality of optimization results, providing a new research idea for the development of related fields, and realizing efficient and stable operation of the electrolytic cell system, providing reliable technical support for intelligent control of hydrogen energy production systems.

[0164] Figure 3 The multi-dimensional performance index radar chart of the renewable energy hydrogen-based energy power dynamic balance regulation method of the embodiment of the application shows the comprehensive performance of four different optimization methods on five key performance indicators (system efficiency, stabilization time, hydrogen production, energy consumption and power utilization rate). As can be seen from the figure, the technical scheme (represented by a circular data point) is better than the other three methods in all indicators, forming the outermost pentagonal area. According to the data in Table 1, in terms of system efficiency, the technical scheme reaches 76.5%, which is significantly higher than the 69.2% of PID control, the 72.1% of model predictive control and the 67.8% of fixed parameter method; in terms of stabilization time, the technical scheme only needs 3.2 minutes, which is much lower than the 7.5 minutes of PID control, the 5.1 minutes of model predictive control and the 9.8 minutes of fixed parameter method, reflecting faster system response speed; in terms of hydrogen production, the technical scheme can achieve a production of 1085.3 Nm 3 / h, which is 11.6% higher than PID control, 6.0% higher than model predictive control and 14.8% higher than fixed parameter method; in terms of energy consumption, the technical scheme is 4.32 kWh / Nm 3, 11.3% lower than PID control, 5.9% lower than model predictive control, and 15.6% lower than the fixed parameter method; in terms of power utilization, the technical solution achieves 92.8%, which is much higher than other methods.

[0165] From the shape of the radar chart, the performance curve of the technical solution uniformly expands outward, indicating that there is a balanced advantage in each indicator; the PID control (represented by square data points) performs poorly in stability time and energy consumption, forming an inwardly recessed area; the model predictive control (represented by diamond data points) performs less overall than the technical solution, but is relatively balanced in each indicator; the fixed parameter method (represented by cross-shaped data points) is significantly behind in stability time, hydrogen production, and energy consumption, forming a significantly irregular shape. This all-around performance comparison fully proves the superiority of the online dynamic programming method used in the technical solution, especially the innovative approach of mapping parameters to the latent space through the mapping function and optimizing under manifold space constraints, which enables the system to meet various process constraints while achieving the highest overall performance. Compared to traditional PID control relying on fixed gain parameters and model predictive control relying on accurate mathematical models, the technical solution can more accurately capture the complex relationships between system parameters through conditional probability distribution and counterfactual calculation, thereby achieving more precise and efficient control effects.

[0166] Figure 4 The electrolytic cell parameter and performance index comparison chart of the renewable energy hydrogen-based energy power dynamic balance regulation method of the embodiment of the present application shows the multi-dimensional relationship comparison of electrolytic cell parameters and performance indexes under different load conditions, Figure 4 including seven parallel vertical coordinate axes, from left to right, power load (20%-100%), electrolyte concentration (25%-35%), temperature (65-85℃), pressure (0.5-2.5MPa), current density (2.0-6.0kA / m 2 ), system efficiency (60%-80%), and hydrogen production (600-1200Nm 3 / h). The parameter configuration and performance of the technical solution under three different load conditions (low load 40%, medium load 70%, and high load 90%) are connected by circular data points and solid lines, forming three complete performance curves. Comparing the four optimization methods under the medium load condition, the technical solution (circular data points and connected lines) shows the optimal system efficiency (74.5%) and hydrogen production (1012.7Nm 3 / h); the system efficiency of the PID control (square data points and short dashed lines) is 69.2%, and the hydrogen production is 972.6Nm 3 / h; the efficiency of the model predictive control (diamond data points and long dashed lines) is 72.1%, and the hydrogen production is 1023.4Nm3 / h; the fixed parameter method (cross-shaped data points and dotted line) is the least efficient, with only 67.8% efficiency and only 945.2 Nm 3 / h.

[0167] The technical solution performs well under high load conditions (90%), and by optimizing electrolyte concentration (32.5%), temperature (80°C) and pressure (2.0 MPa) and other parameters, it realizes a system efficiency of 76.5% and a hydrogen production of 1085.3 Nm 3 / h. In contrast, the technical solution can also maintain a high system efficiency (73.5%) under low load conditions (40%) by adjusting the parameters, although the hydrogen production is reduced (895.6 Nm 3 / h). Figure 4 It is clear that the technical solution can intelligently adjust the process parameters according to different load conditions to achieve optimal performance, especially under high load conditions, by increasing the electrolyte concentration and adjusting the operating temperature, a high system efficiency is maintained, while the traditional method has a significant decrease in efficiency under the same conditions. The model predictive control performs well in hydrogen production under medium load, but its system efficiency is lower than that of the technical solution; and the fixed parameter method is significantly inferior in all performance indicators because it cannot dynamically adjust the parameters according to the load changes. This multi-dimensional parameter dynamic optimization capability is the core advantage of the technical solution, which ensures the efficient and stable operation of the electrolytic cell system under various load conditions.

[0168] In an alternative embodiment,

[0169] Updating the electrolytic cell power distribution scheme based on the power prediction data and the efficiency prediction data to obtain an optimal power distribution scheme comprises:

[0170] When the actual energy utilization efficiency value is lower than the preset threshold value, historical operation data of the electrolytic cell unit is obtained, including historical power data, historical efficiency data and historical operation condition data, a prediction model is trained by a deep learning algorithm and power prediction data and efficiency prediction data for the next period are generated;

[0171] Based on the power prediction data and the efficiency prediction data, a target optimization function is constructed, the system overall efficiency is set as the optimization target, the power distribution amount of each electrolytic cell unit is set as the optimization variable, the power balance constraint and the safe operation constraint are combined as the constraint condition, and the target optimization function is solved by an iterative optimization algorithm to obtain an optimal power distribution scheme.

[0172] The energy utilization efficiency value during the system operation process is monitored in real time, and the monitored efficiency value is compared with the preset threshold value. If the actual efficiency value is higher than the pre-set efficiency threshold, the current electrolytic cell power distribution scheme is output as the optimal power distribution scheme. Once it is detected that the actual efficiency value is lower than the threshold value, the data acquisition mechanism is triggered, and the historical operation data of the electrolytic cell unit is extracted from the system database, including the power data, energy conversion efficiency data and operation condition data recorded at each time point, wherein the operation condition data covers the historical records of temperature, pressure, current density and other key parameters.

[0173] In the deep learning algorithm training link, the historical data obtained is reorganized according to the time sequence, the data is normalized, and a deep neural network structure containing multiple hidden layers is established. The network takes historical data as input, extracts data features through multiple nonlinear transformations, and optimizes network parameters using the back propagation algorithm. The network weights are continuously adjusted during the training process until the error between the network output and the actual value converges to a set range. After training, the model performance is evaluated using a validation data set to ensure that the model has good generalization ability.

[0174] Using the trained deep learning model, the power demand and operating efficiency of the next period are predicted, and based on the time sequence characteristics and change rules contained in the historical data, a data set containing predicted power values and predicted efficiency values is generated.

[0175] An optimization function is constructed with the system overall efficiency as the target, and the power distribution of each electrolytic cell unit as the optimization variable. Two types of constraint conditions need to be considered during optimization: power balance constraint requires the sum of power distribution of all electrolytic cells to meet the system total power demand; safety operation constraint ensures that the operating parameters of each electrolytic cell are within the safe range.

[0176] An iterative optimization algorithm is used to solve the optimization problem. In each iteration, the power distribution of each electrolytic cell is dynamically adjusted according to the target function value and the satisfaction of the constraint conditions. Through multiple iterations, the power distribution scheme that can meet all constraint conditions and achieve optimal system efficiency is finally obtained.

[0177] In this embodiment, the data-driven prediction method breaks through the limitations of traditional empirical models, can better capture the dynamic characteristics and change rules of the system, and by integrating power balance constraints and safe operation constraints into the optimization process, both the implementability of the optimization results and the safe and stable operation of the system are ensured, through the application of the iterative optimization algorithm, the efficient solution of the complex optimization problem is realized, and the global optimality of the optimization results is ensured, and finally, the operation efficiency and control precision of the hydrogen production system are significantly improved, the reliability and safety of the operation are enhanced, and a new technical path for intelligent control of the hydrogen production system is provided.

[0178] In a second aspect of the embodiment of the present application, a power dynamic balance regulation and control system of a renewable energy hydrogen production based energy source is provided, comprising:

[0179] The first unit is used for acquiring real-time power generation data of a renewable energy power generation system, real-time energy storage state data of an energy storage device and power grid load demand data, establishing a dynamic balance equation and calculating a surplus power value available for hydrogen production, collecting equipment operation parameters of a hydrogen production device to calculate a current operation efficiency value of the hydrogen production device;

[0180] The second unit is used for constructing a multi-scale spatiotemporal attention network, extracting multi-dimensional time sequence features through an adaptive convolution kernel, modeling dynamic correlation between system components by using a position weighted attention mechanism, and calculating power prediction data and efficiency prediction data for the next 24 hours;

[0181] The third unit is used for calculating an electrolytic cell power distribution scheme according to the power prediction data, the efficiency prediction data, the dynamic balance equation and the surplus power value;

[0182] The fourth unit is used for calculating target power parameters and control parameters of each electrolytic cell unit by using a deep reinforcement learning algorithm based on a Gaussian process through a Bayesian optimization framework to model environmental uncertainty and predict system risks, and combining an online dynamic programming method based on the electrolytic cell power distribution scheme;

[0183] The fifth unit is used for monitoring actual energy utilization efficiency values of each electrolytic cell unit under the control parameters, and updating the electrolytic cell power distribution scheme based on the power prediction data and the efficiency prediction data to obtain an optimal power distribution scheme when the actual energy utilization efficiency value is lower than a preset threshold.

[0184] In a third aspect of the embodiment of the present application,

[0185] An electronic device is provided, comprising:

[0186] a processor;

[0187] a memory for storing processor executable instructions.

[0188] The processor is configured to invoke instructions stored in the memory to perform the method described above.

[0189] The present application can be a method, apparatus, system, and / or computer program product. The computer program product can include a computer-readable storage medium having computer-readable program instructions loaded thereon for executing aspects of the present application.

[0190] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for dynamic power balance regulation of hydrogen-based energy production from renewable energy sources, characterized in that, include: The system acquires real-time power generation data of renewable energy power generation systems, real-time energy storage status data of energy storage devices, and grid load demand data. It establishes dynamic balance equations and calculates the surplus power value that can be used for hydrogen production. It also collects equipment operating parameters of hydrogen production devices and calculates the current operating efficiency value of hydrogen production devices. A multi-scale spatiotemporal attention network is constructed, and multi-dimensional temporal features are extracted through adaptive convolutional kernels. A position-weighted attention mechanism is used to model the dynamic relationships between system components, and power prediction and efficiency prediction data for the next 24 hours are calculated, including: A multi-scale spatiotemporal attention network is constructed. Feature data is obtained by performing convolution operations on the input temporal data through adaptive convolution kernels. The feature data corresponding to the adaptive convolution kernels are concatenated to obtain multi-dimensional temporal features. Based on the multi-dimensional temporal features, a query matrix and a key matrix are constructed. The product of the query matrix and the key matrix is ​​added to the position weighting matrix and then normalized to obtain the attention weight matrix. The attention weight matrix is ​​multiplied by the value matrix of the multi-dimensional temporal features to obtain the attention output data. The attention output data is nonlinearly transformed using a multilayer perceptron to model the dynamic correlation between system components and obtain a dynamic correlation graph. The dynamic correlation graph is input into the corresponding decoder network along with the historical power data and historical efficiency data, respectively, to calculate the power prediction data and efficiency prediction data for the next 24 hours. A joint loss function is constructed based on the mean square error between the power prediction data and the actual power data, and the mean square error between the efficiency prediction data and the actual efficiency data. The parameters of the decoder network are then optimized based on the joint loss function. The power allocation scheme for the electrolytic cell is calculated based on power prediction data, efficiency prediction data, dynamic balance equation, and surplus power value. Based on the electrolytic cell power allocation scheme, a deep reinforcement learning algorithm based on Gaussian process is adopted. The environmental uncertainty is modeled and the system risk is predicted through a Bayesian optimization framework. Combined with the online dynamic programming method, the target power parameters and control parameters of each electrolytic cell unit are calculated. The operating parameters of each electrolytic cell unit are dynamically adjusted according to the control parameters. The actual energy utilization efficiency of each electrolytic cell unit under control parameters is monitored. When the actual energy utilization efficiency is lower than the preset threshold, the electrolytic cell power allocation scheme is updated based on power prediction data and efficiency prediction data to obtain the optimal power allocation scheme.

2. The method according to claim 1, characterized in that, A dynamic equilibrium equation is established and the surplus power available for hydrogen production is calculated. The current operating efficiency of the hydrogen production unit is calculated by collecting equipment operating parameters, including: Substitute the real-time power generation data, the real-time energy storage status data, and the grid load demand data into the dynamic balance equation, wherein the dynamic balance equation represents that the real-time power generation data is equal to the grid load demand data, the change in charging and discharging power of the energy storage device, and the surplus power value. The surplus power value that can be used for hydrogen production is calculated based on the dynamic balance equation. The temperature, pressure, current, and hydrogen production parameters of the hydrogen production unit are collected. The product of the hydrogen production parameter and the higher heating value of hydrogen is used as the numerator, and the product of the current parameter and the voltage of the electrolyzer of the hydrogen production unit is used as the denominator. The current operating efficiency value of the hydrogen production unit is calculated based on the ratio of the numerator to the denominator.

3. The method according to claim 1, characterized in that, The attention output data is nonlinearly transformed using a multilayer perceptron to model the dynamic relationships between system components, resulting in a dynamic relationship graph including: Attention output data is acquired, and a causal graph is constructed based on the attention output data. The causal graph includes a set of nodes, a set of edges, and a causal probability matrix. The features of the attention output data are used as the set of nodes, the associations between the attention output data are used as the set of edges, and the dynamic associations between the attention output data are used as the causal probability matrix. Counterfactual intervention is performed using a multilayer perceptron. Different intervention conditions are applied to the attention output data, the expected output under different intervention conditions is calculated, and the difference between the expected output under different intervention conditions is calculated to obtain the causal effect strength. Based on the intensity of the causal effect, a nonlinear transformation is performed on the attention output data, and the result of the nonlinear transformation is multiplied by the output value of the sigmoid function to obtain the adjusted dynamic correlation. Construct physical characteristic parameters and motion data corresponding to system components. Based on the physical characteristic parameters and motion data, construct kinematic constraint equations, mass conservation equations, and energy conservation equations, and combine them to obtain a set of physical constraint equations. Input the adjusted dynamic correlation into the set of physical constraint equations, calculate the norm value of the set of physical constraint equations and the norm value of the gradient of the set of physical constraint equations, generate a weighted sum based on the norm value of the set of physical constraint equations and the norm value of the gradient of the set of physical constraint equations, and use the weighted sum as the physical consistency loss function. The dynamic correlation between system components is modeled by performing a nonlinear transformation on the adjusted dynamic correlation using a multilayer perceptron, and the parameters of the multilayer perceptron are optimized according to the physical consistency loss function to obtain a dynamic correlation graph.

4. The method according to claim 1, characterized in that, Based on power prediction data, efficiency prediction data, dynamic balance equations, and surplus power values, the electrolytic cell power allocation scheme is calculated as follows: The power prediction data and efficiency prediction data are used as calculation parameters and input into the dynamic equilibrium equation, and the surplus power value is set as a dynamic equilibrium constraint. The dynamic equilibrium equation is solved using a nonlinear programming algorithm, and the optimal operating parameters that satisfy the system's dynamic equilibrium constraints and surplus power constraints are obtained through iterative calculation. The upper and lower power limits of each electrolytic cell are determined based on the optimal operating parameters. The operating efficiency of each electrolytic cell is calculated based on the efficiency prediction data. The power of the electrolytic cells is allocated in descending order of operating efficiency to obtain the electrolytic cell power allocation scheme.

5. The method according to claim 1, characterized in that, A deep reinforcement learning algorithm based on Gaussian processes is employed, using a Bayesian optimization framework to model environmental uncertainties and predict system risks. Combined with online dynamic programming, the target power parameters and control parameters for each electrolyzer unit are calculated, including: The Gaussian process regression method is used to model environmental uncertainty. The function value corresponding to any point in the input space is set to follow a Gaussian distribution. The radial basis function is selected as the kernel function. The variance parameter and length scale parameter of the kernel function are optimized by maximizing the marginal likelihood function, and a Gaussian process regression model is constructed. A Bayesian optimization framework is constructed based on the Gaussian process regression model to calculate the predicted distribution of new input points. The system risk function is obtained based on the mean and variance of the predicted distribution. The system risk function includes a risk occurrence probability term and an expected loss term. A deep reinforcement learning algorithm is constructed for power optimization. The operating parameters and environmental characteristics of each electrolytic cell unit are set as the state space, and the power adjustment command is set as the action space. The deep reinforcement learning algorithm is trained based on experience playback and target network technology. During the training process, a reward function including system efficiency, power deviation and risk loss is used. The target power parameters and control parameters of each electrolytic cell unit are calculated using an online dynamic programming method, and the electrolyte concentration, electrolysis temperature and current density of each electrolytic cell unit are dynamically adjusted according to the control parameters.

6. The method according to claim 5, characterized in that, The target power parameters and control parameters of each electrolytic cell unit were calculated using the online dynamic programming method, including: The target power parameters and control parameters of the electrolytic cell unit are obtained as independent variables and the operating state is obtained as the dependent variable. Intervention operations are performed on the independent and dependent variables to calculate the conditional probability distribution. Based on the conditional probability distribution, counterfactual calculations are performed to perform difference operations on the system output under different intervention conditions, thereby obtaining the degree of influence of the target power parameters and control parameters on the operating state. A mapping function is constructed to map the target power parameters and control parameters to the latent space. It is then determined whether the output of the mapping function satisfies the manifold space constraints. If not, the target power parameters and control parameters are adjusted and the mapping operation is re-executed until the constraints are met. After that, the Jacobian determinant is calculated using the initial probability distribution in the latent space. Based on the Jacobian determinant, the probability distribution is calculated to obtain the probabilistic representation of the system state. Based on the degree of influence, key target power parameters and control parameters are selected, and the probability representation is substituted into the conditional probability distribution for iterative calculation to output the target power parameters and control parameters of each electrolytic cell unit. The manifold space constraints include upper limit constraints on the power of a single electrolytic cell, lower limit constraints on the power of a single electrolytic cell, total power balance constraints, power change rate constraints, power difference constraints between adjacent time points, current density range constraints, temperature range constraints, pressure range constraints, voltage fluctuation constraints, and current fluctuation constraints.

7. The method according to claim 1, characterized in that, The optimal power allocation scheme for the electrolyzer is obtained by updating the power allocation scheme based on power prediction data and efficiency prediction data, including: When the actual energy utilization efficiency value is lower than a preset threshold, the historical operating data of the electrolytic cell unit is obtained, including historical power data, historical efficiency data and historical operating condition data. A prediction model is trained by a deep learning algorithm to generate power prediction data and efficiency prediction data for the next period. Based on the power prediction data and efficiency prediction data, a target optimization function is constructed. The overall system efficiency is set as the optimization target, the power allocation of each electrolytic cell unit is used as the optimization variable, and power balance constraints and safe operation constraints are used as constraints. The optimal power allocation scheme is obtained by solving the target optimization function through an iterative optimization algorithm.

8. A power dynamic balance control system for hydrogen-based energy production from renewable energy sources, used to implement the method described in any one of claims 1-7, characterized in that, include: The first unit is used to acquire real-time power generation data of renewable energy power generation systems, real-time energy storage status data of energy storage devices, and grid load demand data; establish dynamic balance equations and calculate the surplus power value that can be used for hydrogen production; and collect equipment operating parameters of hydrogen production devices to calculate the current operating efficiency value of hydrogen production devices. The second unit is used to construct a multi-scale spatiotemporal attention network. It extracts multi-dimensional temporal features through adaptive convolutional kernels, models the dynamic correlation between system components using a position-weighted attention mechanism, and calculates power prediction data and efficiency prediction data for the next 24 hours. The third unit is used to calculate the power allocation scheme of the electrolytic cell based on power prediction data, efficiency prediction data, dynamic balance equation and surplus power value. The fourth unit is used to calculate the target power parameters and control parameters of each electrolytic cell unit based on the electrolytic cell power allocation scheme, using a deep reinforcement learning algorithm based on Gaussian process, modeling environmental uncertainties and predicting system risks through a Bayesian optimization framework, and combining online dynamic programming methods to dynamically adjust the operating parameters of each electrolytic cell unit according to the control parameters. The fifth unit is used to monitor the actual energy utilization efficiency of each electrolytic cell under the control parameters. When the actual energy utilization efficiency is lower than the preset threshold, the electrolytic cell power allocation scheme is updated based on the power prediction data and efficiency prediction data to obtain the optimal power allocation scheme.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

Images