Power dynamic balance regulation and control method and system for hydrogen-based energy prepared from renewable energy

Abstract

The invention provides a power dynamic balance regulation and control method and system for renewable energy hydrogen-based energy, and relates to the technical field of energy system control, and the method comprises the steps: obtaining renewable energy generation power, an energy storage state and load demand data, building a dynamic balance equation, calculating a surplus power value, collecting equipment parameters, and calculating the operation efficiency; the method comprises the following steps: constructing a multi-scale space-time attention network, predicting power and efficiency data through an adaptive convolution kernel and a position weighted attention mechanism, calculating a power distribution scheme based on the predicted data, carrying out deep reinforcement learning by adopting a Gaussian process to model uncertainty and predict risks, and obtaining control parameters in combination with dynamic planning; and monitoring efficiency and determining an optimal power distribution scheme.

Description

Technical Field

[0001] The present invention relates to the field of energy system control technology, and in particular to a method and system for dynamically balancing power of hydrogen-based energy produced from renewable energy. Background Art

[0002] With the large-scale development of renewable energy, how to efficiently absorb renewable energy has become an urgent problem to be solved. Hydrogen production is an important way to absorb renewable energy, and the system regulation effect directly affects energy utilization efficiency.

[0003] Currently, control methods for renewable energy hydrogen production systems mainly include rule-based control strategies and traditional optimization algorithms. Rule-based control strategies rely on expert experience and are difficult to adapt to complex and changing operating environments. Traditional optimization algorithms, due to their high computational complexity and poor real-time performance, cannot meet the control needs of large-scale hydrogen production systems.

[0004] In recent years, researchers have attempted to apply deep learning to renewable energy generation forecasting. However, existing technologies still have problems such as insufficient consideration of the dynamic correlations between system components, low prediction accuracy, a lack of systematic research on multi-timescale coordinated regulation, and neglect of systemic risks brought about by environmental uncertainty.

[0005] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the Invention

[0006] The embodiments of the present invention provide a method and system for dynamically balancing power of hydrogen-based energy produced from renewable energy, which can at least solve some of the problems existing in the prior art.

[0007] A first aspect of an embodiment of the present invention provides a method for dynamically balancing power of hydrogen-based energy produced from renewable energy, comprising:

[0008] Acquire real-time power generation data from renewable energy power generation systems, real-time energy storage status data from energy storage devices, and grid load demand data, establish a dynamic balance equation, and calculate the surplus power available for hydrogen production. Collect equipment operating parameters from the hydrogen production device to calculate the current operating efficiency of the hydrogen production device.

[0009] A multi-scale spatiotemporal attention network is constructed to extract multi-dimensional temporal features through adaptive convolution kernels. A position-weighted attention mechanism is used to model the dynamic correlations between system components, and power and efficiency prediction data for the next 24 hours is calculated.

[0010] The power distribution plan of the electrolyzer is calculated based on the power prediction data, efficiency prediction data, dynamic balance equation and surplus power value;

[0011] Based on the electrolyzer power allocation scheme, a deep reinforcement learning algorithm based on Gaussian processes is used to model environmental uncertainty and predict system risks through a Bayesian optimization framework. Combined with an online dynamic programming method, the target power parameters and control parameters of each electrolyzer unit are calculated, and the operating parameters of each electrolyzer unit are dynamically adjusted according to the control parameters.

[0012] The actual energy utilization efficiency value of each electrolytic cell unit under the control parameters is monitored. When the actual energy utilization efficiency value is lower than the preset threshold, the electrolytic cell power allocation plan is updated based on the power prediction data and the efficiency prediction data to obtain the optimal power allocation plan.

[0013] In an optional embodiment,

[0014] Establish a dynamic balance equation and calculate the surplus power value available for hydrogen production. Collect the equipment operating parameters of the hydrogen production device and calculate the current operating efficiency value of the hydrogen production device, including:

[0015] Substituting the real-time power generation data, the real-time energy storage status data, and the grid load demand data 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 grid load demand data, the charge and discharge power change of the energy storage device, and the surplus power value, and calculating the surplus power value that can be used for hydrogen production according to the dynamic balance equation;

[0016] The temperature parameters, pressure parameters, current parameters and hydrogen production parameters of the hydrogen production device are collected, the product of the hydrogen production parameter and the higher calorific value of hydrogen is used as the numerator, the product of the current parameter and the electrolyzer voltage of the hydrogen production device is used 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 optional embodiment,

[0018] A multi-scale spatiotemporal attention network is constructed. Adaptive convolution kernels are used to extract multi-dimensional temporal features. A position-weighted attention mechanism is used to model the dynamic correlations between system components. Power and efficiency forecast data for the next 24 hours are calculated, including:

[0019] Construct a multi-scale spatiotemporal attention network, perform convolution operation on the input time series data through the adaptive convolution kernel to obtain feature data, and splice the feature data corresponding to the adaptive convolution kernel to obtain multi-dimensional time series features;

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

[0021] Using a multi-layer perceptron to perform nonlinear transformation on the attention output data, modeling the dynamic correlation between system components to obtain a dynamic correlation graph;

[0022] The dynamic association graph is input into the corresponding decoder network together with the power history data and the efficiency history data, and the power prediction data and efficiency prediction data for the next twenty-four hours are calculated. 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 optimized based on the joint loss function.

[0023] In an optional embodiment,

[0024] The attention output data is nonlinearly transformed using a multi-layer perceptron to model the dynamic correlation between system components to obtain a dynamic correlation graph including:

[0025] Acquire attention output data, and construct 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, wherein features of the attention output data are used as the node set, associations between the attention output data are used as the edge set, and dynamic associations between the attention output data are used as the causal probability matrix;

[0026] Performing counterfactual intervention through a multi-layer perceptron, applying different intervention conditions to the attention output data, calculating the expected output under the different intervention conditions, and calculating the difference between the expected outputs under the different intervention conditions to obtain the causal effect strength;

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

[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 motion data, and combining them to obtain a physical constraint equation group, inputting the adjusted dynamic correlation into the physical constraint equation group, calculating the norm value of the physical constraint equation group and the norm value of the gradient of the physical constraint equation group, generating a weighted sum based on the norm value of the physical constraint equation group and the norm value of the gradient of the physical constraint equation group, and using the weighted sum as a physical consistency loss function;

[0029] The adjusted dynamic correlation is subjected to nonlinear transformation by a multi-layer perceptron, the dynamic correlation between system components is modeled, and the parameters of the multi-layer 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 distribution plan calculated based on power prediction data, efficiency prediction data, dynamic balance equation and surplus power value includes:

[0032] Inputting the power prediction data and the efficiency prediction data as calculation parameters into the dynamic balance equation, and setting the surplus power value as a dynamic balance constraint condition;

[0033] Solving the dynamic balance equation based on a nonlinear programming algorithm, and obtaining the optimal operating parameters that meet the system dynamic balance constraints and surplus power constraints through iterative calculation;

[0034] The upper and lower power constraints of each electrolytic cell are determined according to the optimal operating parameters, the operating efficiency of each electrolytic cell is calculated based on the efficiency prediction data, and the power of the electrolytic cells is allocated in descending order of operating efficiency to obtain an electrolytic cell power allocation plan.

[0035] In an optional embodiment,

[0036] A Gaussian process-based deep reinforcement learning algorithm is used to model environmental uncertainty and predict system risks through a Bayesian optimization framework. Combined with an online dynamic programming method, the target power parameters and control parameters of each electrolyzer unit are calculated, including:

[0037] 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 obey the 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.

[0038] Constructing a Bayesian optimization framework based on the Gaussian process regression model, calculating the predicted distribution of new input points, and obtaining a system risk function based on the mean and variance of the predicted distribution, wherein 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. The operating parameters and environmental characteristics of each electrolyzer unit 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. The training process adopts a reward function that includes system efficiency terms, power deviation terms, and risk loss terms.

[0040] 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.

[0041] In an optional embodiment,

[0042] The target power parameters and control parameters of each electrolytic cell unit are calculated using the online dynamic programming method, including:

[0043] Obtaining target power parameters and control parameters of the electrolyzer unit as independent variables and an operating state as a dependent variable, performing intervention operations on the independent and dependent variables, and calculating a conditional probability distribution;

[0044] Performing counterfactual calculations based on the conditional probability distribution, performing difference operations on system outputs under different intervention conditions, and obtaining the degree of influence of the target power parameters and control parameters on the operating state;

[0045] Constructing a mapping function to map the target power parameter and the control parameter to a latent space, determining whether the output of the mapping function satisfies the constraints of the manifold space; if not, adjusting the target power parameter and the control parameter and re-performing the mapping operation until the constraints are satisfied, calculating the Jacobian determinant using the initial probability distribution in the latent space, performing probability distribution calculation based on the Jacobian determinant, and obtaining a probabilistic representation of the system state;

[0046] Select key target power parameters and control parameters according to the degree of influence, substitute the probability representation into the conditional probability distribution for iterative calculation, and output the target power parameters and control parameters of each electrolytic cell unit;

[0047] Among them, the manifold space constraints include the upper limit constraint of the power of a single electrolytic cell, the lower limit constraint of the power of a single electrolytic cell, the total power balance constraint, the power change rate constraint, the power difference constraint at adjacent moments, the current density range constraint, the temperature range constraint, the pressure range constraint, the voltage fluctuation constraint and the current fluctuation constraint.

[0048] In an optional embodiment,

[0049] The optimal power allocation plan for the electrolyzer is updated based on the power prediction data and the efficiency prediction data, including:

[0050] When the actual energy utilization efficiency value is lower than a preset threshold, obtaining historical operating data of the electrolyzer unit, including historical power data, historical efficiency data, and historical operating condition data, training a prediction model through a deep learning algorithm and generating power prediction data and efficiency prediction data for the next period;

[0051] Based on the power prediction data and efficiency prediction data, a target optimization function is constructed, the overall efficiency of the system is set as the optimization target, the power allocation amount of each electrolytic cell unit is used as the optimization variable, and the power balance constraint and safe operation constraint are combined as constraint conditions. The target optimization function is solved by an iterative optimization algorithm to obtain the optimal power allocation scheme.

[0052] A second aspect of an embodiment of the present invention provides a power dynamic balancing control system for hydrogen-based energy produced from renewable energy, comprising:

[0053] The first unit is used to obtain real-time power generation data of the renewable energy power generation system, real-time energy storage status data of the energy storage device, and grid load demand data, establish a dynamic balance equation and calculate the surplus power value available for hydrogen production, collect equipment operating parameters of the hydrogen production device, and calculate the current operating efficiency value of the hydrogen production device;

[0054] The second unit is used to build a multi-scale spatiotemporal attention network. It extracts multi-dimensional temporal features through adaptive convolution kernels, uses a position-weighted attention mechanism to model the dynamic correlation between system components, and calculates power and efficiency prediction data for the next 24 hours.

[0055] The third unit is used to calculate the electrolytic cell power allocation plan based on the power prediction data, efficiency prediction data, dynamic balance equation and surplus power value;

[0056] A fourth unit is configured to calculate target power parameters and control parameters for each electrolytic cell unit based on the electrolytic cell power allocation scheme, employ a Gaussian process-based deep reinforcement learning algorithm, model environmental uncertainty and predict system risk through a Bayesian optimization framework, and combine this with an online dynamic programming method to dynamically adjust the operating parameters of each electrolytic cell unit according to the control parameters.

[0057] The fifth unit is used to monitor the actual energy utilization efficiency value of each electrolytic cell unit under the control parameters. When the actual energy utilization efficiency value is lower than the preset threshold, the electrolytic cell power allocation plan is updated based on the power prediction data and efficiency prediction data to obtain the optimal power allocation plan.

[0058] According to a third aspect of the embodiments of the present invention,

[0059] An electronic device is provided, comprising:

[0060] processor;

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

[0062] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0063] In the present invention, by constructing a multi-scale spatiotemporal attention network, accurate modeling of the dynamic correlation between system components is achieved, significantly improving the accuracy of power and efficiency predictions. By combining Gaussian processes with deep reinforcement learning, both the quantitative characterization of environmental uncertainty and the real-time optimization of the power allocation scheme are achieved, effectively reducing the system operation risk. By dynamically updating the power allocation scheme, the system is always maintained in the optimal operating state. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a flow chart of a method for dynamically balancing power of hydrogen-based energy produced from renewable energy according to an embodiment of the present invention;

[0065] Figure 2 This is a comparison chart of the accuracy of cause-effect effect analysis of the power dynamic balance control method for producing hydrogen-based energy from renewable energy according to an embodiment of the present invention;

[0066] Figure 3 A radar chart showing the multi-dimensional performance indicators of the method for dynamically balancing power of hydrogen-based energy produced from renewable energy according to an embodiment of the present invention;

[0067] Figure 4 This is a comparison chart of electrolyzer parameters and performance indicators of the power dynamic balance control method for producing hydrogen-based energy from renewable energy according to an embodiment of the present invention. DETAILED DESCRIPTION

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0069] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0070] Figure 1 FIG. 1 is a flow chart of a method for dynamically balancing power of hydrogen-based energy produced from renewable energy according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0071] Acquire real-time power generation data from renewable energy power generation systems, real-time energy storage status data from energy storage devices, and grid load demand data, establish a dynamic balance equation, and calculate the surplus power available for hydrogen production. Collect equipment operating parameters from the hydrogen production device to calculate the current operating efficiency of the hydrogen production device.

[0072] A multi-scale spatiotemporal attention network is constructed to extract multi-dimensional temporal features through adaptive convolution kernels. A position-weighted attention mechanism is used to model the dynamic correlations between system components, and power and efficiency prediction data for the next 24 hours is calculated.

[0073] The power distribution plan of the electrolyzer is calculated based on the power prediction data, efficiency prediction data, dynamic balance equation and surplus power value;

[0074] Based on the electrolyzer power allocation scheme, a deep reinforcement learning algorithm based on Gaussian processes is used to model environmental uncertainty and predict system risks through a Bayesian optimization framework. Combined with an online dynamic programming method, the target power parameters and control parameters of each electrolyzer unit are calculated, and the operating parameters of each electrolyzer unit are dynamically adjusted according to the control parameters.

[0075] The actual energy utilization efficiency value of each electrolytic cell unit under the control parameters is monitored. When the actual energy utilization efficiency value is lower than the preset threshold, the electrolytic cell power allocation plan is updated based on the power prediction data and the efficiency prediction data to obtain the optimal power allocation plan.

[0076] In an optional embodiment,

[0077] Establish a dynamic balance equation and calculate the surplus power value available for hydrogen production. Collect the equipment operating parameters of the hydrogen production device and calculate the current operating efficiency value of the hydrogen production device, including:

[0078] Substituting the real-time power generation data, the real-time energy storage status data, and the grid load demand data 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 grid load demand data, the charge and discharge power change of the energy storage device, and the surplus power value, and calculating the surplus power value that can be used for hydrogen production according to the dynamic balance equation;

[0079] The temperature parameters, pressure parameters, current parameters and hydrogen production parameters of the hydrogen production device are collected, the product of the hydrogen production parameter and the higher calorific value of hydrogen is used as the numerator, the product of the current parameter and the electrolyzer voltage of the hydrogen production device is used 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.

[0080] The configured data acquisition module acquires real-time operational data from the power system. On the renewable energy generation side, real-time power generation data from wind farms, photovoltaic power plants, and other power generation equipment is collected. On the energy storage device side, real-time charge and discharge status and power data from battery packs, energy storage converters, and other equipment are collected. On the grid side, power load data from each metering point is collected.

[0081] To achieve dynamic balance of system power, a dynamic balance equation was established. This equation reflects the relationship between the real-time power generated by renewable energy and each power consumption link. Specifically, real-time power generation equals the sum of the grid load demand, the change in charge and discharge power of the energy storage device, and the surplus power available for hydrogen production. By substituting the collected real-time power generation data, energy storage status data, and load demand data into the equation, the current surplus power available for hydrogen production can be calculated.

[0082] In terms of calculating the operating efficiency of the hydrogen production device, operating parameters are collected by sensors installed at key locations of the hydrogen production device, including: temperature parameters collected by the equipment temperature sensor, pressure parameters collected by the pressure sensor, current parameters collected by the current sensor, and hydrogen production parameters collected by the hydrogen flowmeter. The collected hydrogen production parameters are multiplied by the predetermined high calorific value of hydrogen to obtain the hydrogen production energy value; the collected current parameters are multiplied by the real-time operating voltage of the electrolyzer of the hydrogen production device to obtain the input electrical energy value. The hydrogen production energy value is divided by the input electrical energy value to obtain the current operating efficiency value of the hydrogen production device.

[0083] For example, at a specific moment in time, operating data from a new energy hydrogen production plant shows: the total generating capacity of the wind farm and photovoltaic power station is 10,000 kilowatts, the actual power load of grid users is 6,000 kilowatts, and the energy storage device is charging at a power of 2,000 kilowatts. Substituting this data into the dynamic balance equation, the calculation shows that the available excess power for hydrogen production is 2,000 kilowatts.

[0084] The hydrogen production unit's operating parameters show: an operating temperature of 75°C, a system pressure of 3 MPa, an electrolyzer operating current of 2,000 amperes, and a measured hydrogen production of 400 standard cubic meters per hour. The higher calorific value of hydrogen is 3.54 kilowatt-hours per standard cubic meter, and the electrolyzer operating voltage is 2.1 volts. Based on these parameters, the product of hydrogen production and higher calorific value is calculated to be 1,416 kWh, and the product of current and voltage is 4.2 kWh, resulting in a current efficiency of 67.4%.

[0085] In this embodiment, by deploying a comprehensive data acquisition module, real-time data acquisition of the renewable energy power generation system, energy storage device and grid load is achieved, 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 factors such as power generation power, load demand, energy storage changes and hydrogen production power, and constructs a complete power balance system. It can accurately reflect the current power resources available for hydrogen production in the system, avoid the unreasonable power allocation problem caused by traditional empirical estimation methods, take into account the direct relationship between power input and hydrogen output and the actual loss in the energy conversion process, and make the efficiency evaluation results more objective and accurate.

[0086] In an optional embodiment,

[0087] A multi-scale spatiotemporal attention network is constructed. Adaptive convolution kernels are used to extract multi-dimensional temporal features. A position-weighted attention mechanism is used to model the dynamic correlations between system components. Power and efficiency forecast data for the next 24 hours are calculated, including:

[0088] Construct a multi-scale spatiotemporal attention network, perform convolution operation on the input time series data through the adaptive convolution kernel to obtain feature data, and splice the feature data corresponding to the adaptive convolution kernel to obtain multi-dimensional time series features;

[0089] Constructing a query matrix and a key matrix based on the multi-dimensional time series features, adding the product of the query matrix and the key matrix to a position weight matrix and then normalizing the result to obtain an attention weight matrix, and multiplying the attention weight matrix by the value matrix of the multi-dimensional time series features to obtain attention output data;

[0090] Using a multi-layer perceptron to perform nonlinear transformation on the attention output data, modeling the dynamic correlation between system components to obtain a dynamic correlation graph;

[0091] The dynamic association graph is input into the corresponding decoder network together with the power history data and the efficiency history data, and the power prediction data and efficiency prediction data for the next twenty-four hours are calculated. 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 optimized based on the joint loss function.

[0092] The construction of a multi-scale spatiotemporal attention network begins with feature extraction from the input time series data. By setting adaptive convolution kernels with different observation windows, it captures short-term, medium-term, and long-term temporal variation characteristics. Each convolution kernel automatically adjusts its weight parameters based on the data characteristics to extract the most representative feature information. The convolution kernels process the input data simultaneously, generating feature representations at different time scales through convolution operations. The features extracted at different scales are concatenated and integrated to form a comprehensive feature representation encompassing multiple time dimensions. This multi-dimensional feature representation comprehensively reflects the operational characteristics at different time scales.

[0093] After obtaining multi-dimensional temporal 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 of interest, respectively. A position weighting matrix is also constructed to characterize the importance of different temporal positions. The query matrix is multiplied by the key matrix to obtain a preliminary correlation strength, which is then added to the position weighting matrix to comprehensively consider the influence of temporal position information. After normalization, the resulting attention weight matrix reflects the degree of correlation 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 incorporates temporal position information.

[0094] To further explore the dynamic connections between components, a multi-layer perceptron performs deep feature extraction on the attention output data. By designing multiple hidden layers, each using a different activation function, nonlinear features are transformed, gradually extracting and strengthening the correlation features between components. This multi-layer feature transformation outputs a dynamic correlation graph that fully describes the interactions between components at different time scales.

[0095] During the prediction phase, two independent decoder networks perform power and efficiency prediction tasks, respectively. These decoders receive a dynamic correlation graph and historical data as input, decode features through a multi-layer neural network, and output predictions for the next 24 hours. The joint loss function considers the mean squared error (MSE) of both power and efficiency predictions. By calculating the error between the predicted and actual values, network parameters are continuously adjusted and optimized to improve prediction accuracy.

[0096] For example, in a forecasting practice at a new energy hydrogen production plant, a week of historical operating data was collected, including time-series data such as power generation, load demand, and energy storage status. With a basic sampling interval of 15 minutes, convolution kernels with observation windows of 1, 6, and 24 hours were used for feature extraction. Each convolution kernel output 100 feature dimensions, and the features at the three scales were concatenated to form a 300-dimensional feature vector. These features were processed using an attention mechanism, successfully identifying the diurnal variations in photovoltaic power generation and the peaks and valleys in electricity load. Dynamic correlation modeling accurately captured the delayed response relationship between energy storage devices and power generation, as well as the impact of load variations on hydrogen production power. During the forecasting phase, power and efficiency were predicted for the next 24 hours, with the final power prediction error within 5% and the efficiency prediction error within 3%, providing a precise basis for intelligent control decisions.

[0097] In this embodiment, the feature extraction mechanism of the multi-scale spatiotemporal attention network significantly improves the data processing capability. The feature extraction method based on attention weights can adaptively highlight the influence of important time nodes, effectively filter out redundant information, and improve the prediction model's ability to perceive key timing patterns. The deep feature extraction mechanism breaks through the expression limitations of traditional linear models, can discover and utilize the potential laws in the data, and make the prediction results more in line with the dynamic characteristics of the actual system.

[0098] In an optional embodiment,

[0099] The attention output data is nonlinearly transformed using a multi-layer perceptron to model the dynamic correlation between system components to obtain a dynamic correlation graph including:

[0100] Acquire attention output data, and construct 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, wherein features of the attention output data are used as the node set, associations between the attention output data are used as the edge set, and dynamic associations between the attention output data are used as the causal probability matrix;

[0101] Performing counterfactual intervention through a multi-layer perceptron, applying different intervention conditions to the attention output data, calculating the expected output under the different intervention conditions, and calculating the difference between the expected outputs under the different intervention conditions to obtain the causal effect strength;

[0102] Performing a nonlinear transformation on the attention output data based on the causal effect strength, and multiplying the result of the nonlinear transformation by the output value of the sigmoid function to obtain an adjusted dynamic correlation;

[0103] 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 motion data, and combining them to obtain a physical constraint equation group, inputting the adjusted dynamic correlation into the physical constraint equation group, calculating the norm value of the physical constraint equation group and the norm value of the gradient of the physical constraint equation group, generating a weighted sum based on the norm value of the physical constraint equation group and the norm value of the gradient of the physical constraint equation group, and using the weighted sum as a physical consistency loss function;

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

[0105] After obtaining the attention output data, a causal graph is constructed. Each feature in the attention output data is used as a node in the causal graph. The features include various key indicators of system operation. To determine the association between nodes, it is necessary to calculate the correlation coefficient between any two features. When the absolute value of the correlation coefficient exceeds a preset threshold, it can be considered that there is a causal relationship between the two nodes, and an edge connection relationship 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 probability of causal influence between the corresponding nodes. Through multiple iterative calculations, the values in the causal probability matrix are continuously updated and optimized until the probability distribution stabilizes or the preset number of iterations is reached.

[0106] When performing counterfactual intervention analysis, a multilayer perceptron network architecture is designed to simulate intervention effects. For each feature of the attention output data, multiple intervention conditions of varying degrees are designed. While applying each intervention condition, other features remain unchanged. The post-intervention feature data is fed into the multilayer perceptron network, and the changes in the system output are observed. By comparing the output differences before and after the intervention, the causal effect strength of the feature under each intervention condition is calculated. Repeating this analysis for all features under different intervention conditions yields a complete causal effect strength matrix. This matrix comprehensively reflects the degree of causal influence between each feature in the system.

[0107] Based on the calculated causal effect strength matrix, a nonlinear transformation network is designed to perform feature conversion. This network employs a multi-layered structure, with nonlinear activation functions used between each layer to enhance feature representation. The causal effect strength matrix is input into this network to produce a transformed feature matrix. Simultaneously, a nonlinear transformation is performed on the attention output data to produce a probability distribution matrix. These two matrices are multiplied together to produce an adjusted dynamic correlation matrix, which retains the correlation information between the original features while incorporating the results of causal analysis.

[0108] During the physical constraint modeling phase, the physical characteristic parameters and motion data of each system component are collected. Based on this 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 fundamental physical laws that the system must adhere to. The adjusted dynamic correlations are substituted into the equations, and the norm of the system and the norm of its gradient are calculated. By setting weight coefficients, the weighted sum of these two norms is constructed into a physical consistency loss function. This loss function is used to assess the degree to which the prediction results conform to physical laws.

[0109] A multilayer perceptron (MLP) is constructed for dynamic association modeling. The MLP consists of multiple hidden layers, each using a different activation function to enhance the network's expressive power. The adjusted dynamic association matrix is input into the network, and the predicted output is obtained through forward propagation. During training, a physical consistency loss function is used as the optimization objective, and an optimization algorithm is used to continuously adjust the network parameters. At each iteration, the gradient of the loss function with respect to the network parameters is calculated, and the parameters are updated according to a preset learning rate. The final dynamic association map is output when the loss function converges or the maximum number of iterations is reached.

[0110] For example, in the case of optimizing the operation of a hydrogen production plant, system operational data is first collected, including features such as photovoltaic power generation (0-100MW), energy storage charging and discharging power (-50MW to 50MW), and electrolyzer power (0-80MW). When constructing the causal graph, these features are used as nodes, and the initial causal probability matrix is constructed using the association strength calculated by the attention mechanism.

[0111] In the counterfactual intervention analysis, we applied ±20% perturbations to PV power generation and observed the responses of other characteristics. For example, when PV power increased by 20%, energy storage charging power increased by an average of 15 MW, and electrolyzer power increased by 12 MW. When PV power decreased by 20%, energy storage discharging power increased by an average of 10 MW, and electrolyzer power decreased by 8 MW. Using these response data, we calculated a causal effect strength matrix.

[0112] In physical constraint modeling, the energy balance equation for the electrolyzer is established: input power = hydrogen production × unit hydrogen calorific value / electrolysis efficiency. The energy storage system's charge and discharge efficiency constraints are also considered: 85% charging efficiency and 90% discharge efficiency. These physical constraints are converted into mathematical expressions to construct a physical consistency loss function.

[0113] The optimized dynamic correlation diagram shows a correlation strength of 0.8 between photovoltaic power generation and energy storage charging and discharging power, and a correlation strength of 0.6 between photovoltaic power generation and electrolyzer power. These values reflect both statistical regularity and physical constraints such as energy conservation. This dynamic correlation diagram successfully guided system operation optimization, resulting in a 15% increase in hydrogen production efficiency and a 20% improvement in energy storage utilization.

[0114] In this embodiment, a multi-scale spatiotemporal attention network is constructed. Based on the acquisition of attention output data, a causal graph structure is innovatively introduced. This not only characterizes the correlation between features, but also quantifies the degree of influence between features through a causal probability matrix, making the interaction between system components clearer and more explainable. A counterfactual intervention analysis mechanism is introduced. By designing intervention conditions of varying degrees and simulating the intervention effects using a multi-layer perceptron, the causal effect strength of each feature in the system can be deeply understood. A physical constraint mechanism is incorporated into dynamic correlation modeling. By constructing a set of physical constraint equations including kinematic constraints, mass conservation, and energy conservation, and converting them into a loss function, the prediction results are ensured to meet basic physical laws.

[0115] Existing technologies for modeling dynamic dependencies between system components primarily rely on statistical methods and simple neural network models. These methods often focus only on surface data features, failing to delve deeply into the causal relationships between system components and failing to fully consider the constraints of physical laws. Traditional methods struggle to capture long-range dependencies when processing long-term time series data and are susceptible to data noise, resulting in insufficient reliability and accuracy in prediction results.

[0116] This embodiment significantly improves the effect of modeling dynamic correlations between system components by innovatively combining causal analysis, physical constraints, and deep learning methods, providing reliable technical support for the system's intelligent optimization control. It not only improves the accuracy and reliability of predictions, but also enhances the interpretability and practicality of the model, providing new research ideas for technological development in related fields.

[0117] Figure 2This is a comparison chart of the accuracy of the causal effect analysis of the power dynamic balance control method of renewable energy hydrogen-based energy in the embodiment of the present invention, showing the response characteristics of the system under different intervention conditions. This technical solution (square mark) accurately predicts the changes in energy storage charging and discharging power under conditions where the photovoltaic power changes by ±20% by combining a multi-layer perceptron with physical constraints. When the photovoltaic power increases by 20%, the prediction accuracy of the energy storage charging power increase by 15MW reaches 95.8%, while the prediction accuracy of the traditional VAR model (circular mark) and TCN model (triangle mark) are 83.2% and 87.5% respectively. Under extreme conditions (intervention intensity of 25%), this solution still maintains an accuracy of 93.5%, which is 8.7 percentage points higher than the second place, reflecting the robustness of the solution under large changes in working conditions.

[0118] In an optional embodiment,

[0119] The electrolytic cell power distribution plan calculated based on power prediction data, efficiency prediction data, dynamic balance equation and surplus power value includes:

[0120] Inputting the power prediction data and the efficiency prediction data as calculation parameters into the dynamic balance equation, and setting the surplus power value as a dynamic balance constraint condition;

[0121] Solving the dynamic balance equation based on a nonlinear programming algorithm, and obtaining the optimal operating parameters that meet the system dynamic balance constraints and surplus power constraints through iterative calculation;

[0122] The upper and lower power constraints of each electrolytic cell are determined according to the optimal operating parameters, the operating efficiency of each electrolytic cell is calculated based on the efficiency prediction data, and the power of the electrolytic cells is allocated in descending order of operating efficiency to obtain an electrolytic cell power allocation plan.

[0123] First, the power forecast data and efficiency forecast data are input as calculation parameters into the dynamic balance equation. The power forecast data includes time series data such as renewable energy generation power, load demand power, and energy storage device charge and discharge power. The efficiency forecast data includes performance parameters such as the operating efficiency of the electrolyzer group and the charge and discharge efficiency of the energy storage device. When inputting this data, it is necessary to preprocess the data, including data standardization, outlier processing, and time alignment, to ensure data quality and consistency. At the same time, the system's surplus power value is set as a constraint condition for dynamic balance. This constraint condition requires that the difference between the system's generated power and power consumption at any time does not exceed a preset threshold range.

[0124] Using a nonlinear programming algorithm to solve the dynamic equilibrium equation and construct the objective function requires comprehensive consideration of multiple optimization objectives, such as maximizing system operating efficiency and minimizing energy loss, as well as setting constraints, including power balance constraints, equipment operation constraints, and energy storage capacity constraints. During the solution process, an iterative calculation method is used, updating the system state variables with each iteration until an optimal solution that satisfies all constraints is found. System parameters are initialized, and the objective function value and the degree of constraint satisfaction are calculated in each iteration. The optimization variables are adjusted based on the calculation results until the algorithm converges or the maximum number of iterations is reached.

[0125] Based on the calculated optimal operating parameters, the power allocation plan for the electrolyzer group is determined. Based on the optimal operating parameters, the upper and lower power constraints for each electrolyzer are determined. Technical parameters such as the electrolyzer's rated power and minimum starting power need to be considered. Based on efficiency prediction data, the operating efficiency of each electrolyzer under different operating conditions is calculated. Electrolyzers are sorted from high to low according to operating efficiency, with power allocated preferentially to those with higher efficiency to maximize overall system efficiency. During the allocation process, constraints such as the electrolyzer's start-stop switching cost and power change rate also need to be considered to ensure the feasibility of the allocation plan.

[0126] For example, consider a hydrogen production plant that includes a photovoltaic power generation system, an energy storage system, and multiple electrolyzers. The system receives power forecast data for the next 24 hours, including time-series data such as photovoltaic power generation forecasts and grid load forecasts. It also obtains efficiency forecast data, including the efficiency forecasts for each electrolyzer at different load rates and the charge and discharge efficiency of the energy storage system.

[0127] The dynamic balance constraint setting specifies the fluctuation range of the system's surplus power. When PV power generation fluctuates, the system's dynamic balance is maintained by adjusting the energy storage system's charge and discharge power and the electrolyzer's operating power. A nonlinear programming algorithm is used for optimization, and through multiple iterative calculations, the optimal operating parameters that satisfy all constraints are obtained.

[0128] Based on the optimal operating parameters, the operating power range for each electrolyzer is determined. By analyzing efficiency prediction data, the operating efficiency of each electrolyzer under current operating conditions is calculated. After sorting the electrolyzers by efficiency, higher-efficiency electrolyzers are prioritized for greater operating power allocation, while lower-efficiency electrolyzers are allocated less power or placed in standby mode. This allocation strategy ensures dynamic system balance while improving overall operating efficiency.

[0129] In this embodiment, by using power prediction data and efficiency prediction data as calculation parameters and combining them with surplus power constraints, a complete system dynamic balance model is established, fully considering the mutual influence between the various components of the system, so that the optimization results can accurately reflect the actual operating status of the system. The iterative solution mechanism of the nonlinear programming algorithm ensures the convergence and reliability of the optimization results. Through the synergistic effect of multi-objective optimization and multiple constraints, the system can maximize operating efficiency while ensuring dynamic balance.

[0130] In an optional embodiment,

[0131] A Gaussian process-based deep reinforcement learning algorithm is used to model environmental uncertainty and predict system risks through a Bayesian optimization framework. Combined with an online dynamic programming method, the target power parameters and control parameters of each electrolyzer unit are calculated, including:

[0132] 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 obey the 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.

[0133] Constructing a Bayesian optimization framework based on the Gaussian process regression model, calculating the predicted distribution of new input points, and obtaining a system risk function based on the mean and variance of the predicted distribution, wherein 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 and environmental characteristics of each electrolyzer unit 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. The training process adopts a reward function that includes system efficiency terms, power deviation terms, and risk loss terms.

[0135] 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.

[0136] A Gaussian process regression approach is used to model environmental uncertainty. The input spatial dimensions are determined, including environmental parameters such as ambient temperature, humidity, and light intensity, which are used as input variables. Data at each input point is preprocessed, including data standardization and outlier processing. A radial basis function is selected as the kernel function, which contains two parameters to be optimized: the variance parameter and the length scale parameter. A marginal likelihood function is constructed and used as the optimization objective function. The marginal likelihood function is optimized using the gradient ascent method, and the optimal kernel function parameter values are obtained through iterative calculation. A complete Gaussian process regression model is constructed based on the optimized kernel function parameters.

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

[0138] Define the state space, which includes the electrolyzer operating parameters (power, efficiency, temperature, etc.) and environmental characteristic parameters. Define the action space as a set of power adjustment instructions, including power adjustment amounts of different amplitudes. Construct an experience replay pool to store training samples. Establish an evaluation network and a target network. The two networks have the same network structure but different parameter update frequencies. Design a reward function that includes system efficiency terms, power deviation terms, and risk loss terms. During the training process, the intelligent agent interacts with the environment and stores the gained experience in the experience replay pool. Regularly sample randomly from the experience replay pool for training, use the target network to calculate the target Q value, and update the evaluation network parameters through backpropagation. Regularly copy the evaluation network parameters to the target network.

[0139] An online dynamic programming approach is used to calculate control parameters. The target power parameters for each electrolyzer are calculated based on the current system state and the output of the deep reinforcement learning algorithm. Based on the target power parameters and the dynamic characteristic equations of the electrolyzer, the control variables for electrolyte concentration, electrolysis temperature, and current density are calculated. These control variables are then constrained to ensure they meet the equipment's operating limits. These processed control variables are then converted into specific control instructions and sent to the actuators to execute the control operations.

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

[0141] Based on a trained Gaussian process regression model, the system predicts environmental conditions for future periods. For example, it can predict changes in light intensity over the next four hours. The system calculates the probability of risk occurrence based on the predicted distribution and, combined with a predefined loss function, calculates the expected loss, resulting in a complete risk assessment.

[0142] In the deep reinforcement learning algorithm, the electrolyzer's operating parameters, such as power, efficiency, and temperature, along with environmental parameters, are combined into a state vector. The action space is defined as five power adjustment settings: -10%, -5%, 0%, +5%, and +10%. A deep neural network with multiple hidden layers is constructed as the evaluation network and target network. A reward function is designed that combines system efficiency, power deviation, and risk loss according to preset weights.

[0143] Based on the output of the deep reinforcement learning algorithm, the target power value for each electrolyzer is calculated. Based on this target power value and the dynamic characteristics of the electrolyzer, adjustment instructions for electrolyte concentration, electrolysis temperature, and current density are calculated. For example, when the electrolyzer power needs to be increased, the system will simultaneously increase the electrolyte concentration and current density, and adjust the electrolysis temperature accordingly to ensure stable operation of the electrolyzer under the new operating conditions.

[0144] In this embodiment, the radial basis function is used as the kernel function, combined with the optimization of the marginal likelihood function, so that 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 the 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 the efficient and stable operation of the electrolytic cell under various working conditions.

[0145] In an optional embodiment,

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

[0147] Obtaining target power parameters and control parameters of the electrolyzer unit as independent variables and an operating state as a dependent variable, performing intervention operations on the independent and dependent variables, and calculating a conditional probability distribution;

[0148] Performing counterfactual calculations based on the conditional probability distribution, performing difference operations on system outputs under different intervention conditions, and obtaining the degree of influence of the target power parameters and control parameters on the operating state;

[0149] Constructing a mapping function to map the target power parameter and the control parameter to a latent space, determining whether the output of the mapping function satisfies the constraints of the manifold space; if not, adjusting the target power parameter and the control parameter and re-performing the mapping operation until the constraints are satisfied, calculating the Jacobian determinant using the initial probability distribution in the latent space, performing probability distribution calculation based on the Jacobian determinant, and obtaining a probabilistic representation of the system state;

[0150] Select key target power parameters and control parameters according to the degree of influence, substitute the probability representation into the conditional probability distribution for iterative calculation, and output the target power parameters and control parameters of each electrolytic cell unit;

[0151] Among them, the manifold space constraints include the upper limit constraint of the power of a single electrolytic cell, the lower limit constraint of the power of a single electrolytic cell, the total power balance constraint, the power change rate constraint, the power difference constraint at adjacent moments, the current density range constraint, the temperature range constraint, the pressure range constraint, the voltage fluctuation constraint and the current fluctuation constraint.

[0152] The target power parameters (including power setpoint and power adjustment range) and control parameters (including electrolyte concentration, temperature setpoint, and pressure setpoint) of the electrolyzer unit are collected as independent variables, while operating status data (including actual power, efficiency, and temperature) are collected as dependent variables. Intervention operations are performed on the dependent variables, including setting different intervention conditions, changing the values of the independent variables, observing the changes in the dependent variables, and calculating the conditional probability distribution based on these changes. The intervention operation needs to consider the causal relationship between the variables to ensure the rationality of the intervention.

[0153] Based on the obtained conditional probability distribution, a comparative analysis of system output under different intervention conditions is performed. A baseline scenario is set, and comparison scenarios are generated by changing single or multiple parameters. The difference in system output under different scenarios is calculated. This difference calculation can quantify the impact of target power parameters and control parameters on operating conditions. This process needs to consider the interaction between parameters to ensure the accuracy of the calculation results.

[0154] Construct a mapping function to map the target power parameters and control parameters to the latent space. This mapping function needs to maintain the essential characteristics of the data while reducing the dimensionality of the data. Then determine whether the mapping result meets the constraints of the manifold space. The constraints include power limits, rate of change limits, process parameter range limits, and other aspects. If the constraints are not met, the original parameters need to be adjusted and the mapping operation needs to be re-executed. When the constraints are met, the Jacobian determinant is calculated using the initial probability distribution in the latent space, and the probability distribution calculation is performed based on the determinant to obtain a probabilistic representation of the system state.

[0155] Based on the degree of influence obtained in advance, the key parameters that have a significant impact on the system operating status are screened out. These key parameters are combined with the probability representation obtained in the third step and substituted into the conditional probability distribution for iterative calculation. After multiple iterations, the optimal target power parameters and control parameters of each electrolytic cell unit are finally output.

[0156] For example, we collected operational data for each of the five electrolyzers in a hydrogen production plant. Independent variables included power setpoint (0-1 MW), electrolyte concentration (25-35%), operating temperature (65-85°C), and operating pressure (0.1-3.0 MPa). Dependent variables included actual output power, system efficiency, and hydrogen production.

[0157] During the intervention operation, parameters such as the power setting (±10%), temperature setting (±5°C), and pressure setting (±0.2MPa) were varied, and the system responses were recorded. For example, when the power setting was increased from 0.8MW to 0.88MW, the changes in efficiency and hydrogen production were observed, and the conditional probability distribution was calculated based on these changes.

[0158] In the counterfactual calculation, a baseline operating condition (power 0.8 MW, temperature 75°C, pressure 2.0 MPa) was set, and multiple sets of comparative operating conditions were generated. By calculating the performance differences under different operating conditions, the influence of each parameter was determined. For example, it was found that the influence of temperature changes on efficiency was 0.8, and on hydrogen production was 0.6.

[0159] In the mapping function construction, a multi-layer neural network is used to map the original parameters into a low-dimensional space. The mapping result is checked to see if it meets the constraints, such as a power limit of 1MW and a temperature range of 65-85°C. If any constraints are not met, the original parameters are adjusted and remapped. Once the constraints are met, the Jacobian determinant is calculated and the probability distribution is obtained.

[0160] Key parameters (such as power and temperature settings) were selected based on their impact and then iteratively calculated using a probabilistic representation. After multiple iterations, the optimized parameters for each electrolyzer were obtained: power, temperature, and pressure settings. These parameters met all constraints while ensuring efficient system operation.

[0161] In this embodiment, by introducing intervention operations and conditional probability distribution calculations, a causal relationship model between parameters was established. This model can accurately identify the true impact of parameter changes on system performance. By comparing the system responses under different intervention conditions, the influence of each parameter is quantitatively evaluated, providing an objective measure of parameter importance and a scientific basis for subsequent parameter screening and optimization. By constructing a mapping function to map the parameter space to the latent space and introducing multidimensional constraints, it is ensured that the optimization results meet the physical limitations and process requirements of the system, significantly improving the feasibility of the optimization results.

[0162] In existing technologies, parameter optimization of electrolyzer systems mainly relies on simple statistical analysis and empirical models. These methods often focus only on superficial correlations between parameters, fail to deeply explore the causal relationships between parameters, and are unable to accurately assess the actual impact of parameter changes on system performance. When dealing with multi-parameter coupled optimization problems, they are prone to falling into local optimality and it is difficult to ensure that the optimization results meet multiple constraints.

[0163] This embodiment can effectively handle the complex interactions between parameters and has strong adaptability to system fluctuations. It not only improves the accuracy and reliability of optimization, but also enhances the interpretability and practicality of the optimization results. It provides new research ideas for technological development in related fields, realizes the efficient and stable operation of the electrolyzer system, and provides reliable technical support for the intelligent control of the hydrogen energy production system.

[0164] Figure 3 This is a radar chart of multi-dimensional performance indicators of the power dynamic balance control method of renewable energy-based hydrogen energy according to the embodiment of the present invention, which 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). It can be seen intuitively from the figure that the present technical solution (represented by circular data points) is superior to the other three methods in all indicators, forming the outermost pentagonal area. Specifically comparing the data in Table 1, it can be seen that in terms of system efficiency, the present technical solution reaches 76.5%, which is significantly higher than 69.2% of PID control, 72.1% of model predictive control and 67.8% of fixed parameter method; in terms of stabilization time index, the present technical solution only takes 3.2 minutes, which is much lower than 7.5 minutes of PID control, 5.1 minutes of model predictive control and 9.8 minutes of fixed parameter method, reflecting a faster system response speed; in terms of hydrogen production, the present technical solution can achieve 1085.3Nm 3 / h of output, 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 index, this technical solution is 4.32kWh / 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, this technical solution reaches 92.8%, which is much higher than other methods.

[0165] From the shape analysis of the radar chart, the performance curve of this technical solution expands outward evenly, indicating that it has balanced advantages in various indicators; while PID control (represented by square data points) performs poorly in terms of stabilization time and energy consumption, forming an inward concave area; although the overall performance of model predictive control (represented by diamond data points) is inferior to this technical solution, it is relatively balanced in various indicators; the fixed parameter method (represented by cross-shaped data points) lags significantly behind in the three indicators of stabilization time, hydrogen production and energy consumption, forming a significantly irregular shape. This comprehensive performance comparison fully demonstrates the superiority of the online dynamic programming method adopted by this technical solution, especially the innovative idea of mapping parameters to latent space by constructing a mapping function and optimizing under the constraints of manifold space, so that the system can achieve the highest comprehensive performance while meeting various process constraints. Compared with traditional PID control that relies on fixed gain parameters and model predictive control that relies on precise mathematical models, this technical solution can more accurately capture the complex relationship between system parameters through conditional probability distribution and counterfactual calculation, thereby achieving more accurate and efficient control effects.

[0166] Figure 4 This is a comparison chart of electrolyzer parameters and performance indicators in the power dynamic balance control method for producing hydrogen-based energy from renewable energy in an embodiment of the present invention, showing the multi-dimensional relationship comparison between electrolyzer parameters and performance indicators under different load conditions. Figure 4 Contains 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 this 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 to form three complete performance curves. Comparing the four optimization methods under medium load conditions, it can be seen that this technical solution (circular data point connection) shows the best system efficiency (74.5%) and hydrogen production (1012.7Nm 3 / h); while the system efficiency of PID control (square data points and short dashed line) is 69.2%, and the hydrogen production is 972.6Nm 3 / h; the model predictive control (diamond data points and long dashed line) efficiency is 72.1%, and the hydrogen production is 1023.4Nm3 / h; the fixed parameter method (cross-shaped data points and dot-dash line) has the lowest efficiency, only 67.8%, and the hydrogen production is only 945.2Nm 3 / h.

[0167] This technical solution performs well under high load conditions (90%). By optimizing parameters such as electrolyte concentration (32.5%), temperature (80°C) and pressure (2.0MPa), a system efficiency of 76.5% and a load of 1085.3Nm 3 / h of hydrogen production. In contrast, this technical solution can also maintain a high system efficiency (73.5%) by adjusting parameters under low load conditions (40%), although the hydrogen production is reduced (895.6Nm 3 / h). Figure 4 It clearly demonstrates that this technical solution can intelligently adjust process parameters according to different load conditions to achieve optimal performance. In particular, under high-load conditions, by increasing the electrolyte concentration and adjusting the operating temperature, a high system efficiency is maintained, while the efficiency of traditional methods is significantly reduced under the same conditions. Although model predictive control performs well in hydrogen production under medium load, its system efficiency is lower than that of this technical solution; and the fixed parameter method lags significantly behind in various performance indicators because it cannot dynamically adjust parameters according to load changes. This dynamic optimization capability of multi-dimensional parameters is the core advantage of this technical solution, ensuring the efficient and stable operation of the electrolyzer system under various load conditions.

[0168] In an optional embodiment,

[0169] The optimal power allocation plan for the electrolyzer is updated based on the power prediction data and the efficiency prediction data, including:

[0170] When the actual energy utilization efficiency value is lower than a preset threshold, obtaining historical operating data of the electrolyzer unit, including historical power data, historical efficiency data, and historical operating condition data, training a prediction model through a deep learning algorithm and generating power prediction data and efficiency prediction data for the next period;

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

[0172] The energy efficiency value during system operation is monitored in real time and compared with a preset threshold. If the actual efficiency value is higher than the preset efficiency threshold, the current electrolyzer power allocation plan is output as the optimal power allocation plan. Once the actual efficiency value is detected to be lower than the threshold, the data acquisition mechanism is triggered to extract the historical operation data of the electrolyzer unit from the system database, including power data, energy conversion efficiency data, and operating condition data recorded at various time points. The operating condition data covers historical records of key parameters such as temperature, pressure, and current density.

[0173] During the deep learning algorithm training phase, historical data is reorganized and normalized according to time series, building a deep neural network structure with multiple hidden layers. This network uses historical data as input, extracts data features through multiple layers of nonlinear transformations, and optimizes network parameters using a backpropagation algorithm. During training, network weights are continuously adjusted 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 dataset to ensure good generalization.

[0174] The trained deep learning model is used to predict the power demand and operating efficiency of the next period. Based on the time series characteristics and change patterns contained in the historical data, a data set containing predicted power values and predicted efficiency values is generated.

[0175] An optimization function targeting overall system efficiency was constructed, with the power allocation of each electrolyzer unit as the optimization variable. Two types of constraints needed to be considered simultaneously during the optimization process: a power balance constraint requiring the sum of the power allocations of all electrolyzers to meet the total system power demand; and a safe operation constraint ensuring that the operating parameters of each electrolyzer remained within a safe range.

[0176] An iterative optimization algorithm is used to solve this optimization problem. In each iteration, the power allocation of each electrolyzer is dynamically adjusted based on the objective function value and the satisfaction of the constraints. Through multiple rounds of iterative calculations, a power allocation solution is ultimately obtained that satisfies all constraints and achieves optimal system efficiency.

[0177] In this embodiment, the data-driven prediction method breaks through the limitations of traditional empirical models and can better capture the dynamic characteristics and changing laws of the system. By integrating power balance constraints and safe operation constraints into the optimization process, it not only ensures the feasibility of the optimization results, but also ensures the safe and stable operation of the system. Through the application of iterative optimization algorithms, efficient solutions to complex optimization problems are achieved, and the global optimality of the optimization results is guaranteed. In summary, this embodiment significantly improves the operating efficiency and control accuracy of the hydrogen production system, enhances the reliability and safety of operation, and provides a new technical path for the intelligent control of the hydrogen production system.

[0178] A second aspect of an embodiment of the present invention provides a power dynamic balancing control system for hydrogen-based energy produced from renewable energy, comprising:

[0179] The first unit is used to obtain real-time power generation data of the renewable energy power generation system, real-time energy storage status data of the energy storage device, and grid load demand data, establish a dynamic balance equation and calculate the surplus power value available for hydrogen production, collect equipment operating parameters of the hydrogen production device, and calculate the current operating efficiency value of the hydrogen production device;

[0180] The second unit is used to build a multi-scale spatiotemporal attention network. It extracts multi-dimensional temporal features through adaptive convolution kernels, uses a position-weighted attention mechanism to model the dynamic correlation between system components, and calculates power and efficiency prediction data for the next 24 hours.

[0181] The third unit is used to calculate the electrolytic cell power allocation plan based on the power prediction data, efficiency prediction data, dynamic balance equation and surplus power value;

[0182] A fourth unit is configured to calculate target power parameters and control parameters for each electrolytic cell unit based on the electrolytic cell power allocation scheme, employ a Gaussian process-based deep reinforcement learning algorithm, model environmental uncertainty and predict system risk through a Bayesian optimization framework, and combine this with an online dynamic programming method to dynamically adjust the operating parameters of each electrolytic cell unit according to the control parameters.

[0183] The fifth unit is used to monitor the actual energy utilization efficiency value of each electrolytic cell unit under the control parameters. When the actual energy utilization efficiency value is lower than the preset threshold, the electrolytic cell power allocation plan is updated based on the power prediction data and efficiency prediction data to obtain the optimal power allocation plan.

[0184] According to a third aspect of the embodiments of the present invention,

[0185] An electronic device is provided, comprising:

[0186] processor;

[0187] a memory for storing processor-executable instructions;

[0188] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0189] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0190] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for dynamically balancing power of renewable energy hydrogen-based energy, characterized in that: include: Acquire real-time power generation data from renewable energy power generation systems, real-time energy storage status data from energy storage devices, and grid load demand data, establish a dynamic balance equation, and calculate the surplus power available for hydrogen production. Collect equipment operating parameters from the hydrogen production device to calculate the current operating efficiency of the hydrogen production device. A multi-scale spatiotemporal attention network is constructed to extract multi-dimensional temporal features through adaptive convolution kernels. A position-weighted attention mechanism is used to model the dynamic correlations between system components, and power and efficiency prediction data for the next 24 hours is calculated. The power distribution plan of the electrolyzer is calculated based on the power prediction data, efficiency prediction data, dynamic balance equation and surplus power value; Based on the electrolyzer power allocation scheme, a deep reinforcement learning algorithm based on Gaussian processes is used to model environmental uncertainty and predict system risks through a Bayesian optimization framework. Combined with an online dynamic programming method, the target power parameters and control parameters of each electrolyzer unit are calculated, and the operating parameters of each electrolyzer unit are dynamically adjusted according to the control parameters. The actual energy utilization efficiency value of each electrolytic cell unit under the control parameters is monitored. When the actual energy utilization efficiency value is lower than the preset threshold, the electrolytic cell power allocation plan is updated based on the power prediction data and the efficiency prediction data to obtain the optimal power allocation plan.

2. The method according to claim 1, characterized in that Establish a dynamic balance equation and calculate the surplus power value available for hydrogen production. Collect the equipment operating parameters of the hydrogen production device and calculate the current operating efficiency value of the hydrogen production device, including: Substituting the real-time power generation data, the real-time energy storage status data, and the grid load demand data 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 grid load demand data, the charge and discharge power change of the energy storage device, and the surplus power value, and calculating the surplus power value that can be used for hydrogen production according to the dynamic balance equation; The temperature parameters, pressure parameters, current parameters and hydrogen production parameters of the hydrogen production device are collected, the product of the hydrogen production parameter and the higher calorific value of hydrogen is used as the numerator, the product of the current parameter and the electrolyzer voltage of the hydrogen production device is used 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.

3. The method according to claim 1, characterized in that A multi-scale spatiotemporal attention network is constructed. Adaptive convolution kernels are used to extract multi-dimensional temporal features. A position-weighted attention mechanism is used to model the dynamic correlation between system components. Power and efficiency forecast data for the next 24 hours are calculated, including: Construct a multi-scale spatiotemporal attention network, perform convolution operation on the input time series data through the adaptive convolution kernel to obtain feature data, and splice the feature data corresponding to the adaptive convolution kernel to obtain multi-dimensional time series features; Constructing a query matrix and a key matrix based on the multi-dimensional time series features, adding the product of the query matrix and the key matrix to a position weight matrix and then normalizing the result to obtain an attention weight matrix, and multiplying the attention weight matrix by the value matrix of the multi-dimensional time series features to obtain attention output data; Using a multi-layer perceptron to perform nonlinear transformation on the attention output data, modeling the dynamic correlation between system components to obtain a dynamic correlation graph; The dynamic association graph is input into the corresponding decoder network together with the power history data and the efficiency history data, and the power prediction data and efficiency prediction data for the next twenty-four hours are calculated. 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 optimized based on the joint loss function.

4. The method according to claim 3, characterized in that The attention output data is nonlinearly transformed using a multi-layer perceptron to model the dynamic correlation between system components to obtain a dynamic correlation graph including: Acquire attention output data, and construct 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, wherein features of the attention output data are used as the node set, associations between the attention output data are used as the edge set, and dynamic associations between the attention output data are used as the causal probability matrix; Performing counterfactual intervention through a multi-layer perceptron, applying different intervention conditions to the attention output data, calculating the expected output under the different intervention conditions, and calculating the difference between the expected outputs under the different intervention conditions to obtain the causal effect strength; Performing a nonlinear transformation on the attention output data based on the causal effect strength, and multiplying the result of the nonlinear transformation by the output value of the sigmoid function to obtain an adjusted dynamic correlation; 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 motion data, and combining them to obtain a physical constraint equation group, inputting the adjusted dynamic correlation into the physical constraint equation group, calculating the norm value of the physical constraint equation group and the norm value of the gradient of the physical constraint equation group, generating a weighted sum based on the norm value of the physical constraint equation group and the norm value of the gradient of the physical constraint equation group, and using the weighted sum as a physical consistency loss function; The adjusted dynamic correlation is subjected to nonlinear transformation by a multi-layer perceptron, the dynamic correlation between system components is modeled, and the parameters of the multi-layer perceptron are optimized according to the physical consistency loss function to obtain a dynamic correlation graph.

5. The method according to claim 1, wherein The electrolytic cell power distribution plan calculated based on power prediction data, efficiency prediction data, dynamic balance equation and surplus power value includes: Inputting the power prediction data and the efficiency prediction data as calculation parameters into the dynamic balance equation, and setting the surplus power value as a dynamic balance constraint condition; Solving the dynamic balance equation based on a nonlinear programming algorithm, and obtaining the optimal operating parameters that meet the system dynamic balance constraints and surplus power constraints through iterative calculation; The upper and lower power constraints of each electrolytic cell are determined according to the optimal operating parameters, the operating efficiency of each electrolytic cell is calculated based on the efficiency prediction data, and the power of the electrolytic cells is allocated in descending order of operating efficiency to obtain an electrolytic cell power allocation plan.

6. The method according to claim 1, characterized in that A Gaussian process-based deep reinforcement learning algorithm is used to model environmental uncertainty and predict system risks through a Bayesian optimization framework. Combined with an online dynamic programming method, the target power parameters and control parameters of 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 obey the 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. Constructing a Bayesian optimization framework based on the Gaussian process regression model, calculating the predicted distribution of new input points, and obtaining a system risk function based on the mean and variance of the predicted distribution, wherein 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 electrolyzer unit 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. The training process adopts a reward function that includes system efficiency terms, power deviation terms, and risk loss terms. 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.

7. The method according to claim 6, characterized in that The target power parameters and control parameters of each electrolytic cell unit are calculated using the online dynamic programming method, including: Obtaining target power parameters and control parameters of the electrolytic cell unit as independent variables and an operating state as a dependent variable, performing intervention operations on the independent variables and the dependent variables, and calculating a conditional probability distribution; Performing counterfactual calculations based on the conditional probability distribution, performing difference operations on system outputs under different intervention conditions, and obtaining the degree of influence of the target power parameters and control parameters on the operating state; Constructing a mapping function to map the target power parameter and the control parameter to a latent space, determining whether the output of the mapping function satisfies the constraints of the manifold space; if not, adjusting the target power parameter and the control parameter and re-performing the mapping operation until the constraints are satisfied, calculating the Jacobian determinant using the initial probability distribution in the latent space, performing probability distribution calculation based on the Jacobian determinant, and obtaining a probabilistic representation of the system state; Select key target power parameters and control parameters according to the degree of influence, substitute the probability representation into the conditional probability distribution for iterative calculation, and output the target power parameters and control parameters of each electrolytic cell unit; Among them, the manifold space constraints include the upper limit constraint of the power of a single electrolytic cell, the lower limit constraint of the power of a single electrolytic cell, the total power balance constraint, the power change rate constraint, the power difference constraint at adjacent moments, the current density range constraint, the temperature range constraint, the pressure range constraint, the voltage fluctuation constraint and the current fluctuation constraint.

8. The method according to claim 1, characterized in that The optimal power allocation plan for the electrolyzer is updated based on the power prediction data and the efficiency prediction data, including: When the actual energy utilization efficiency value is lower than a preset threshold, obtaining historical operating data of the electrolyzer unit, including historical power data, historical efficiency data, and historical operating condition data, training a prediction model through a deep learning algorithm and generating 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 efficiency of the system is set as the optimization target, the power allocation amount of each electrolytic cell unit is used as the optimization variable, and the power balance constraint and safe operation constraint are combined as constraint conditions. The target optimization function is solved by an iterative optimization algorithm to obtain the optimal power allocation scheme.

9. A power dynamic balance control system for hydrogen-based energy produced from renewable energy, used to implement the method according to any one of claims 1 to 8, characterized in that: include: The first unit is used to obtain real-time power generation data of the renewable energy power generation system, real-time energy storage status data of the energy storage device, and grid load demand data, establish a dynamic balance equation and calculate the surplus power value available for hydrogen production, collect equipment operating parameters of the hydrogen production device, and calculate the current operating efficiency value of the hydrogen production device; The second unit is used to build a multi-scale spatiotemporal attention network. It extracts multi-dimensional temporal features through adaptive convolution kernels, uses a position-weighted attention mechanism to model the dynamic correlation between system components, and calculates power and efficiency prediction data for the next 24 hours. The third unit is used to calculate the electrolytic cell power allocation plan based on the power prediction data, efficiency prediction data, dynamic balance equation and surplus power value; A fourth unit is configured to calculate target power parameters and control parameters for each electrolytic cell unit based on the electrolytic cell power allocation scheme, employ a Gaussian process-based deep reinforcement learning algorithm, model environmental uncertainty and predict system risk through a Bayesian optimization framework, and combine this with an online dynamic programming method 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 value of each electrolytic cell unit under the control parameters. When the actual energy utilization efficiency value is lower than the preset threshold, the electrolytic cell power allocation plan is updated based on the power prediction data and efficiency prediction data to obtain the optimal power allocation plan.

10. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 8.

Images