Electricity-hydrogen coupling intelligent regulation and control method and system considering wind and light prediction error

Abstract

The invention provides an electricity-hydrogen coupling intelligent regulation and control method and system considering a wind-solar prediction error, and relates to the technical field of electric power, and the method comprises the steps: collecting wind-solar historical power generation data, and obtaining prediction error probability distribution based on adaptive combination kernel density estimation; the distribution serves as an uncertainty constraint, a dual neural network is adopted to execute deep reinforcement learning, an action value network dynamically adjusts a decision space, and a target network evaluates risks and restrains decisions based on conditional risks and state values; a distributed collaborative network is constructed, and hydrogen production power and hydrogen desorption power are controlled in real time based on a dynamic consistency theory and a synchronous potential function. The operation stability and economical efficiency of the electricity-hydrogen coupling system are improved, and the uncertainty of wind and light power generation is effectively dealt with.

Description

Technical Field

[0001] The present invention relates to power technology, and particularly to an electric-hydrogen coupling intelligent control method and system considering wind and light prediction errors. Background Art

[0002] As a clean and efficient energy carrier, hydrogen energy can realize long-term storage of renewable energy through electrolytic water hydrogen production. At present, the control methods of electric-hydrogen coupling systems mainly optimize the scheduling based on deterministic prediction data. However, the prediction errors of wind power and photovoltaic power generation have significant randomness and uncertainty, which pose challenges to the safe and economic operation of the system. At the same time, as distributed devices, the coordinated control of hydrogen production systems and hydrogen storage systems has an important impact on the overall performance of the system. The main problems existing in the prior art are as follows:

[0003] The existing control methods of electric-hydrogen coupling systems often use a single probability distribution model to describe the wind and light prediction errors, and cannot accurately characterize the complex distribution characteristics of the prediction errors, resulting in insufficient robustness of system control.

[0004] Traditional scheduling optimization methods lack the ability to adaptively adjust to the short-term fluctuations and long-term trends of wind and light prediction errors, and do not fully consider the system operation risks in the decision-making process, which easily leads to unsatisfactory actual execution effects of the scheduling scheme.

[0005] The existing coordinated control methods of distributed hydrogen production systems do not establish an association mechanism with system operation risks, and lack the fast response ability based on real-time prediction deviations, making it difficult to achieve efficient coordinated operation of the system. Summary of the Invention

[0006] Embodiments of the present invention provide an electric-hydrogen coupling intelligent control method and system considering wind and light prediction errors, which can solve the problems in the prior art.

[0007] In the first aspect of the embodiments of the present invention,

[0008] There is provided an electric-hydrogen coupling intelligent control method considering wind and light prediction errors, including:

[0009] Collecting historical power generation data of wind farms and photovoltaic power stations, calculating the spatio-temporal correlation characteristics of prediction errors of wind farms and photovoltaic power stations, and obtaining the probability distribution of wind and light prediction errors by using an adaptive combined kernel density estimation method of Gaussian kernel function and Epanechnikov kernel function;

[0010] Taking the probability distribution as an uncertainty constraint, combining the power grid load demand data and the operation data of the hydrogen storage system, and adopting a dual neural network to perform deep reinforcement learning, where the action-value network dynamically adjusts the decision space based on an adaptive boundary function of the short-term fluctuation and long-term trend of the wind-solar prediction error; the target network evaluates risks through a conditional risk and state-value double branch, and constructs a risk constraint coefficient based on the Lyapunov stability theory for decision-making risk constraint to obtain a scheduling instruction.

[0011] Construct the electrolytic hydrogen production unit and the hydrogen storage system into a distributed collaborative network, inversely map the coupling strength of the distributed collaborative network to the risk constraint coefficient based on the dynamic consistency theory, and construct a synchronization potential function from the deviation between the node state of the distributed collaborative network and the scheduling instruction and the state difference of adjacent nodes; Based on the deviation between the actual power generation power and the predicted power generation power of the wind farm and the photovoltaic power station, the hydrogen production power and the hydrogen release power are controlled in real time along the gradient direction of the synchronization potential function.

[0012] In an alternative embodiment,

[0013] Calculating the spatio-temporal correlation characteristics of the prediction errors of the wind farm and the photovoltaic power station, and obtaining the probability distribution of the wind-solar prediction error by using an adaptive combined kernel density estimation method of a Gaussian kernel function and an Epanechnikov kernel function includes:

[0014] Calculating the time correlation coefficient and the space correlation coefficient of the prediction error data of the wind farm and the photovoltaic power station to obtain spatio-temporal coupling characteristics;

[0015] Selecting a Gaussian kernel function and an Epanechnikov kernel function as basic kernel functions, where the Gaussian kernel function is used for probability density estimation in the central region, and the Epanechnikov kernel function is used for probability density estimation in the boundary region; calculating the adaptive weights of the Gaussian kernel function and the Epanechnikov kernel function based on the local sample density, and taking the sum of the product of the Gaussian kernel function and the adaptive weight and the product of the Epanechnikov kernel function and the complementary value of the adaptive weight as the combined kernel function;

[0016] Calculating the initial bandwidth based on the standard deviation and interquartile range of the historical power generation data, and adjusting the initial bandwidth by quantile constraint according to the empirical distribution function to obtain the optimized bandwidth;

[0017] Substituting the spatio-temporal coupling characteristics, the combined kernel function and the optimized bandwidth into the kernel density estimation formula to obtain the probability density function of the wind-solar power generation prediction error; updating the prediction error samples based on a sliding time window, calculating the sample weights according to the time decay factor, and performing recursive update operations on the sample weights and the probability density function to obtain a dynamically updated probability distribution of the wind-solar prediction error.

[0018] In an alternative embodiment,

[0019] Deep reinforcement learning is performed using a dual neural network, where the action-value network dynamically adjusts the decision space based on an adaptive boundary function for the short-term fluctuations and long-term trends of the wind-solar prediction error; the target network evaluates risks through a conditional risk and state-value dual-branch, and constructs a risk constraint coefficient based on the Lyapunov stability theory for decision risk constraint, including:

[0020] Obtain the system state vector, where the system state vector includes grid load demand data, hydrogen production power data, hydrogen storage capacity data, and hydrogen refueling demand data;

[0021] Input the system state vector into the action-value network, where the action-value network calculates the adaptive boundary function based on the probability distribution of the wind-solar prediction error, and dynamically adjusts the basic decision boundary according to the adaptive boundary function to obtain the adjusted decision space;

[0022] Input the system state vector and the adjusted decision space into the target network, where the target network calculates the conditional risk value and the state value, and weights the conditional risk value and the state value to obtain the risk metric value;

[0023] Construct an optimization objective function including a system operation cost term, a risk metric term, and a renewable energy consumption term, and use the stochastic gradient descent method to update the action-value network parameters based on the optimization objective function. The target network evaluates the decision risk based on the risk metric value and periodically updates the network parameters through a preset soft update coefficient;

[0024] Perform value function iterative calculation according to the updated action-value network parameters and risk metric value, and use the product of the iterative calculation result and the risk constraint coefficient as the risk constraint term. The risk constraint coefficient is obtained by constructing a Lyapunov function based on the quadratic term of the hydrogen storage capacity deviating from the safe interval and the high-order term of the wind-solar prediction error and calculating the reciprocal of the stability margin through its time derivative;

[0025] Based on the adjusted decision space, calculate the decision evaluation value considering the risk constraint term, and select the decision corresponding to the maximum decision evaluation value as the scheduling instruction.

[0026] In an alternative embodiment,

[0027] The adaptive boundary function includes:

[0028] Perform multi-scale decomposition on the wind-solar prediction error sequence through discrete wavelet transform to obtain the short-term fluctuation error and the long-term trend error;

[0029] Calculate the boundary adjustment feature of the short-term fluctuation error based on the local probability density feature of the prediction error samples;

[0030] Construct a cumulative impact function, which is obtained by weighted accumulation of the long-term trend errors in historical periods according to an exponential decay coefficient. Calculate a trend intensity index based on the cumulative impact function, where the trend intensity index is the ratio of the absolute value of the cumulative impact function to the maximum value of the absolute value of the cumulative impact function within the evaluation time window;

[0031] Construct a probability confidence interval based on the boundary adjustment feature, map the boundary values of the probability confidence interval to a fluctuation boundary adjustment amount, and the fluctuation boundary adjustment amount is dynamically corrected based on the fluctuation direction change frequency; Construct a piecewise linear mapping function based on the trend intensity index to calculate the trend boundary adjustment amount, and the trend boundary adjustment amount is corrected according to the trend duration;

[0032] Calculate an adaptive combination weight based on the trend intensity index, and the adaptive combination weight has an exponential relationship with the trend intensity index; Combine the fluctuation boundary adjustment amount and the trend boundary adjustment amount through the adaptive combination weight to obtain the adaptive boundary function.

[0033] In an alternative implementation,

[0034] The target network calculates the conditional risk value and the state value. The weighted combination of the conditional risk value and the state value to obtain the risk metric value includes:

[0035] Construct a system energy function with multi-dimensional state coupling based on the system state vector and the adjusted decision space. The system energy function is a non-linear weighted combination of grid load demand data, hydrogen production power data, hydrogen storage data, and hydrogen refueling demand data and their corresponding reference values;

[0036] Construct a conditional risk branch and a state value branch of the target network;

[0037] Input the adjusted decision space and the historical decision sequence into the conditional risk branch. Calculate an attention weight matrix based on the multi-level correlation degree between the system state vector and the historical state, and use the attention weight matrix and the spatio-temporal convolution result of the historical state as the conditional risk value;

[0038] Input the system state vector and the adjusted decision space into the state value branch, construct a posterior distribution of the state value using the probabilistic variational inference method, and extract the conditional expectation and uncertainty of the posterior distribution to obtain the state value;

[0039] Determine an adaptive time window length based on the system energy function, where the adaptive time window length is not less than the response time of the electric-hydrogen coupling system and not greater than a preset multiple of the time constant of the hydrogen storage system;

[0040] Obtain the maximum and minimum values of the energy of the electric-hydrogen coupling system within the adaptive time window, and calculate the multi-scale stability index based on the difference between the current system energy value and the maximum and minimum values;

[0041] Construct an adaptive weighting coefficient according to the multi-scale stability index, and non-linearly weight the conditional risk value and the state value according to the adaptive weighting coefficient to obtain the risk metric value.

[0042] In an alternative embodiment,

[0043] The risk constraint coefficient includes:

[0044] Obtain the real-time value of the hydrogen storage amount, the upper and lower limits of the hydrogen storage safety interval, and the wind-solar prediction error sequence; construct the degree of deviation of the hydrogen storage amount from the safety interval as a quadratic term, and construct the wind-solar prediction error value as a high-order term to form a Lyapunov function characterizing the system stability; calculate the stability margin based on the derivative of the Lyapunov function with respect to time, and take the reciprocal of the stability margin as the risk constraint coefficient.

[0045] In an alternative embodiment,

[0046] Based on the dynamic consistency theory, map the coupling strength of the distributed cooperative network inversely proportional to the risk constraint coefficient, and construct a synchronization potential function from the deviation between the node state of the distributed cooperative network and the scheduling instruction and the state difference of adjacent nodes, including:

[0047] Construct a dynamic coupling strength based on the risk constraint coefficient, where the dynamic coupling strength is the ratio of the basic coupling strength to the risk adjustment term, the risk adjustment term increases with the increase of the risk constraint coefficient, and the dynamic coupling strength also has an exponential decay relationship with the state difference of adjacent nodes;

[0048] Construct a synchronization potential function, which includes a scheduling instruction tracking term and a node cooperation term. The scheduling instruction tracking term is the product of the square difference between the node state and the scheduling instruction and the first weight coefficient, and the node cooperation term is the product of the square of the state difference of adjacent nodes, the dynamic coupling strength, and the second weight coefficient;

[0049] Construct a distributed control law based on the gradient of the synchronization potential function, where the distributed control law includes a scheduling tracking term and a cooperation adjustment term. The scheduling tracking term is the product of the deviation between the node state and the scheduling instruction and the first weight coefficient, and the cooperation adjustment term is the product of the weighted deviation between the node state and the state of adjacent nodes and the second weight coefficient;

[0050] When the actual power generation of the wind farm and the photovoltaic power station exceeds the predicted power generation, increase the hydrogen production power based on the distributed control law; when the actual power generation is less than the predicted power generation, increase the hydrogen release power of the hydrogen storage system based on the distributed control law, and perform real-time coordinated control on the power-to-hydrogen system.

[0051] In an alternative embodiment,

[0052] The synchronous potential function includes a fast response layer and a steady-state tracking layer, including:

[0053] The scheduling instruction tracking term of the fast response layer uses the product of the square difference between the node state and the scheduling instruction and the first weight coefficient, and the node coordination term of the fast response layer uses the product of the square of the adjacent node state difference, the dynamic coupling strength, and the second weight coefficient; the fast response layer triggers calculation and update based on the change rate of the node state, and updates the first weight coefficient and the second weight coefficient when the change rate of the node state exceeds a preset threshold;

[0054] The scheduling instruction tracking term of the steady-state tracking layer uses the product of the cumulative square difference between the node state and the scheduling instruction and the first weight coefficient, where the cumulative square difference is the time integral of the square difference between the node state and the scheduling instruction within a fixed time window, and the length of the fixed time window is not less than 10 times the calculation period of the fast response layer and not greater than one-third of the system characteristic time, where the system characteristic time is the dynamic response time of the power-to-hydrogen system; the node coordination term of the steady-state tracking layer uses the product of the square of the adjacent node state difference, the dynamic coupling strength, and the second weight coefficient;

[0055] Construct an inter-layer coupling weight based on the calculation results of the fast response layer and the steady-state tracking layer, increase the weight of the fast response layer when the change rate of the node state is greater than the first preset threshold, and increase the weight of the steady-state tracking layer when the change rate of the node state is less than the second preset threshold.

[0056] In the second aspect of the embodiments of the present invention,

[0057] Provide a power-to-hydrogen coupling intelligent control system considering wind-solar prediction errors, including:

[0058] The first unit is used to collect the historical power generation data of the wind farm and the photovoltaic power station, calculate the spatio-temporal correlation characteristics of the wind-solar prediction errors, and obtain the probability distribution of the wind-solar prediction errors by using an adaptive combined kernel density estimation method of Gaussian kernel function and Epanechnikov kernel function;

[0059] The second unit is used to take the probability distribution as an uncertainty constraint, combine the power grid load demand data and the operation data of the hydrogen storage system, and perform deep reinforcement learning using a dual neural network, where the action-value network dynamically adjusts the decision space based on the adaptive boundary function of the short-term fluctuation and long-term trend of the wind-solar prediction error; the target network evaluates risks through the conditional risk and state-value dual branches, and constructs a risk constraint coefficient based on the Lyapunov stability theory for decision-making risk constraint to obtain a scheduling instruction.

[0060] The third unit is used to construct the electrolytic hydrogen production unit and the hydrogen storage system into a distributed collaborative network, inversely map the coupling strength of the distributed collaborative network to the risk constraint coefficient based on the dynamic consistency theory, and construct a synchronization potential function from the deviation between the node state of the distributed collaborative network and the scheduling instruction and the state difference of adjacent nodes; based on the deviation between the actual power generation power and the predicted power generation power of the wind farm and the photovoltaic power station, the hydrogen production power and the hydrogen release power are controlled in real time along the gradient direction of the synchronization potential function.

[0061] In the third aspect of the embodiments of the present invention,

[0062] A kind of electronic equipment is provided, including:

[0063] A processor;

[0064] A memory for storing instructions executable by the processor;

[0065] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0066] The present invention uses the adaptive combined kernel density estimation method to obtain the probability distribution of the wind-solar prediction error, fully considers the spatio-temporal correlation characteristics of wind-solar power generation, improves the accuracy of the prediction error probability distribution, and provides a reliable uncertainty constraint for subsequent regulation decisions.

[0067] The present invention uses a dual neural network to perform deep reinforcement learning. Through the coordinated cooperation of the action-value network and the target network, the dynamic adjustment of the decision space and risk constraint are realized, which not only ensures the adaptability of the scheduling instruction but also ensures the safety of system operation, and improves the regulation efficiency of the electric-hydrogen coupling system.

[0068] The present invention constructs the electrolytic hydrogen production unit and the hydrogen storage system into a distributed collaborative network, and realizes real-time control based on the dynamic consistency theory and the synchronization potential function, effectively coping with the real-time fluctuations of wind-solar power generation, improving the dynamic response ability and operation stability of the system, and realizing the real-time intelligent collaborative regulation of the electric-hydrogen coupling system. Description of the Drawings

[0069] Figure 1Schematic flowchart of the electric-hydrogen coupled intelligent regulation method considering the prediction errors of wind power and photovoltaic power in the embodiments of the present invention;

[0070] Figure 2 Comparison chart of the probability distribution estimation accuracy between the present invention and the traditional Gaussian kernel function method and the fixed weight method;

[0071] Figure 3 Comparison chart of the performance of the double-layer control of the present invention and other single-layer controls. Detailed implementation manners

[0072] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only some of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0073] The following combines Figures 1 to 3 Specific embodiments are used to illustrate the technical solutions of the present invention in detail. These specific embodiments below can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0074] Figure 1 Schematic flowchart of the electric-hydrogen coupled intelligent regulation method considering the prediction errors of wind power and photovoltaic power in the embodiments of the present invention, as Figure 1 shown, the method includes:

[0075] Collect historical power generation data of wind farms and photovoltaic power stations, calculate the spatio-temporal correlation characteristics of the prediction errors of wind farms and photovoltaic power stations, and use an adaptive combined kernel density estimation method of Gaussian kernel function and Epanechnikov kernel function to obtain the probability distribution of the wind-solar prediction errors;

[0076] Take the probability distribution as an uncertainty constraint, combine the power grid load demand data and the operation data of the hydrogen storage system, and use a dual neural network to perform deep reinforcement learning, where the action value network dynamically adjusts the decision space based on the adaptive boundary function of the short-term fluctuation and long-term trend of the wind-solar prediction errors; the target network evaluates risks through the conditional risk and state value double branches, and constructs a risk constraint coefficient based on the Lyapunov stability theory for decision risk constraint to obtain a scheduling instruction;

[0077] Construct the electrolytic hydrogen production unit and the hydrogen storage system into a distributed collaborative network, inversely map the coupling strength of the distributed collaborative network to the risk constraint coefficient based on the dynamic consistency theory, and construct a synchronization potential function from the deviation between the node state of the distributed collaborative network and the scheduling instruction and the state difference of adjacent nodes; Based on the deviation between the actual power generation of the wind farm and the photovoltaic power station and the predicted power generation, control the hydrogen production power and hydrogen release power in real time along the gradient direction of the synchronization potential function.

[0078] In an alternative embodiment,

[0079] Calculate the spatio-temporal correlation characteristics of the prediction errors of the wind farm and the photovoltaic power station, and use the adaptive combined kernel density estimation method of the Gaussian kernel function and the Epanechnikov kernel function to obtain the probability distribution of the wind-solar prediction errors, including:

[0080] Calculate the time correlation coefficient and the spatial correlation coefficient of the prediction error data of the wind farm and the photovoltaic power station to obtain the spatio-temporal coupling characteristics;

[0081] Select the Gaussian kernel function and the Epanechnikov kernel function as the basic kernel functions. The Gaussian kernel function is used for probability density estimation in the central region, and the Epanechnikov kernel function is used for probability density estimation in the boundary region; Calculate the adaptive weights of the Gaussian kernel function and the Epanechnikov kernel function based on the local sample density, and take the sum of the product of the Gaussian kernel function and the adaptive weight and the product of the Epanechnikov kernel function and the complement of the adaptive weight as the combined kernel function;

[0082] Calculate the initial bandwidth based on the standard deviation and the interquartile range of the historical power generation data, and adjust the initial bandwidth by quantile constraint according to the empirical distribution function to obtain the optimized bandwidth;

[0083] Substitute the spatio-temporal coupling characteristics, the combined kernel function, and the optimized bandwidth into the kernel density estimation formula to obtain the probability density function of the wind-solar power generation prediction error; Update the prediction error samples based on a sliding time window, calculate the sample weights according to the time decay factor, and perform recursive update operations on the sample weights and the probability density function to obtain the dynamically updated probability distribution of the wind-solar prediction errors.

[0084] Exemplarily, the spatio-temporal coupling characteristics of the prediction errors of the wind farm and the photovoltaic power station are analyzed. The time series of the prediction errors of the wind farm and the photovoltaic power station are obtained. Based on the autoregressive moving average model, the autocorrelation function of the time series of the prediction errors is calculated to obtain the time dimension characteristics. A spatial correlation matrix between the wind farm and the photovoltaic power station is established, and the spatial correlation coefficient between the wind farm and the photovoltaic power station is calculated to obtain the spatial dimension characteristics. The Kronecker product operation is performed on the time weight matrix corresponding to the time dimension characteristics and the spatial weight matrix corresponding to the spatial dimension characteristics to obtain the spatio-temporal coupling characteristics. Taking a 100MW wind farm and a 50MW photovoltaic power station as an example, the power generation data and prediction data with a 15-minute resolution are collected. The time correlation of the prediction error sequence is calculated through a sliding time window, and the window length is set to 24 hours. For the wind farm data, when the time delay is 1 hour, the correlation coefficient is 0.82; when the delay is 4 hours, the correlation coefficient drops to 0.45. For the photovoltaic power station data, when the time delay is 1 hour, the correlation coefficient is 0.75; when the delay is 4 hours, the correlation coefficient drops to 0.38. The spatial correlation coefficient between the wind farm and the photovoltaic power station is 0.42, indicating that the prediction errors have significant spatial correlation. Compared with the traditional separate analysis method, this spatio-temporal coupling analysis can improve the accuracy of prediction error feature extraction by 25%.

[0085] In the kernel function selection step, an adaptive combination of the Gaussian kernel function and the Epanechnikov kernel function is adopted. The Gaussian kernel function has good smoothness in the central region and is used to process the main part of the prediction error distribution. Specifically, when the prediction error value falls within plus or minus one standard deviation range, the Gaussian kernel function is used for probability density estimation. The Epanechnikov kernel function has excellent convergence in the boundary region and is used to characterize the tail characteristics of the prediction error distribution. When the prediction error value exceeds two standard deviations, the Epanechnikov kernel function is used for estimation.

[0086] The calculation of the adaptive weight is based on the local sample density characteristics. Within the range of plus or minus one standard deviation, when the local sample density is greater than 0.8, the weight of the Gaussian kernel function is 0.9, and the weight of the Epanechnikov kernel function is 0.1; when the local sample density is between 0.4 and 0.8, the weight of the Gaussian kernel function is 0.7, and the weight of the Epanechnikov kernel function is 0.3. In the boundary region, when the local sample density is less than 0.2, the weight of the Gaussian kernel function drops to 0.1, and the weight of the Epanechnikov kernel function rises to 0.9. Practical applications show that this adaptive weight strategy improves the accuracy of probability distribution estimation by 30% compared with the fixed weight method.

[0087] The bandwidth optimization process first determines the initial value based on historical data characteristics. For wind farm data, the standard deviation is 8 MW, the interquartile range is 12 MW, and the initial bandwidth is set to 2.5 MW. For photovoltaic power plant data, the standard deviation is 4 MW, the interquartile range is 6 MW, and the initial bandwidth is set to 1.5 MW. The initial bandwidth is optimized by the quantile constraint method. Specifically, when the difference between the 90% quantile and the 10% quantile is less than the expected threshold, the bandwidth is reduced by 15%; when the difference is greater than the expected threshold, the bandwidth is increased by 10%. The optimized bandwidths are 1.8 MW and 1.2 MW respectively, and the estimation accuracy is improved by 20% compared with the initial bandwidth.

[0088] In the probability distribution dynamic update step, the spatio-temporal coupling characteristics, the combined kernel function, and the optimized bandwidth are substituted into the kernel density estimation formula to obtain the probability density function of the wind-solar power generation prediction error; a 3-hour sliding time window is used to update the prediction error samples. A larger weight is assigned to the newly entered samples, and the time decay factor is set to 0.95, that is, for each time interval moved forward, the sample weight is reduced by 5%. The updated sample weights and the existing probability density function are recursively calculated to achieve the dynamic update of the probability distribution. When a mutation occurs in the prediction error, such as the power fluctuation exceeding 20% of the rated capacity, the sliding window length is shortened to 1 hour to accelerate the distribution update speed.

[0089] Figure 2 The probability distribution estimation accuracy comparison chart shows the estimation accuracy performance of the three methods in different local sample density intervals. The present invention adopts the adaptive combined kernel function strategy, achieving an estimation accuracy of 95% in the high-density region (0.8 - 1.0), maintaining an accuracy of 90% in the medium-density region (0.4 - 0.8), and even maintaining an accuracy of 85% in the low-density region (0 - 0.2). Although the traditional Gaussian kernel function method performs well in the high-density region (reaching 80%), its accuracy significantly decreases in the medium and low-density regions (75% and 65% respectively). Although the fixed weight method is slightly better than the single Gaussian kernel function method, its accuracy in each density region is still generally about 15 - 20 percentage points lower than the technical solution of the present invention. Especially when dealing with the boundary region (low-density region), the technical solution of the present invention makes the estimation accuracy increase by about 20 percentage points compared with the traditional method by increasing the weight of the Epanechnikov kernel function, fully demonstrating the superiority of the adaptive combined kernel function strategy.

[0090] The existing methods for analyzing the prediction errors of wind and solar power generation ignore the spatio-temporal coupling relationship, do not consider the dynamic changes in the data distribution characteristics, and are difficult to quickly respond to the sudden changes in prediction errors. The present invention proposes a method for estimating the probability distribution of wind and solar power generation prediction errors based on adaptive combined kernel density estimation. This method creatively introduces a spatio-temporal coupling analysis framework, comprehensively analyzes the spatio-temporal characteristics of the prediction errors of wind farms and photovoltaic power plants through a sliding time window, and improves the accuracy of feature extraction. In terms of kernel function selection, the Gaussian kernel function and the Epanechnikov kernel function are innovatively combined adaptively. The Gaussian kernel function is used for probability density estimation in the central region, and the Epanechnikov kernel function is used for modeling the boundary region. Through the adaptive weight allocation based on the local sample density, the advantages of the two kernel functions are complementary. In terms of dynamic update of the probability distribution, a sample weight update mechanism and a mutation response strategy based on a time decay factor are designed, and the response speed of the system to sudden changes in prediction errors is improved by adaptively adjusting the length of the sliding window. The accuracy of the prediction error distribution estimation of the present invention is significantly improved, the calculation efficiency is significantly improved, the spatio-temporal coupling analysis improves the accuracy of feature extraction, the adaptive weight strategy optimizes the probability distribution estimation effect, the bandwidth optimization scheme improves the overall performance, and the mutation response speed is significantly accelerated, providing a new idea for the probability modeling of wind and solar power generation prediction errors and laying a foundation for improving the prediction accuracy of renewable energy power generation and the coordinated operation efficiency of the power-to-hydrogen system.

[0091] In an alternative embodiment,

[0092] Double neural networks are used to perform deep reinforcement learning, where the action-value network dynamically adjusts the decision space based on an adaptive boundary function of the short-term fluctuations and long-term trends of the wind and solar power prediction errors; the target network evaluates risks through a conditional risk and state-value double-branch, and constructs a risk constraint coefficient based on the Lyapunov stability theory for decision risk constraint, including:

[0093] Obtain a system state vector, where the system state vector includes grid load demand data, hydrogen production power data, hydrogen storage capacity data, and hydrogen refueling demand data;

[0094] Input the system state vector into the action-value network, where the action-value network calculates an adaptive boundary function based on the probability distribution of the wind and solar power prediction errors, and dynamically adjusts the basic decision boundary according to the adaptive boundary function to obtain an adjusted decision space;

[0095] Input the system state vector and the adjusted decision space into the target network, where the target network calculates the conditional risk value and the state value, and weights the conditional risk value and the state value to obtain a risk metric value;

[0096] Construct an optimization objective function that includes system operation cost items, risk measurement items, and renewable energy consumption items. Use the stochastic gradient descent method to update the action value network parameters based on the optimization objective function. The target network evaluates the decision-making risk based on the risk measurement value and periodically updates the network parameters through a preset soft update coefficient;

[0097] Perform value function iterative calculation according to the updated action value network parameters and risk measurement value, and take the product of the iterative calculation result and the risk constraint coefficient as the risk constraint item. The risk constraint coefficient is obtained by constructing a Lyapunov function based on the quadratic term of the deviation of the hydrogen storage amount from the safe interval and the high-order term of the wind and light prediction error and calculating the reciprocal of the stability margin through its time derivative;

[0098] Based on the adjusted decision-making space, calculate the decision evaluation value considering the risk constraint item, and select the decision corresponding to the maximum decision evaluation value as the scheduling instruction.

[0099] Exemplarily, obtain the system state vector, including grid load demand data, hydrogen production power data, hydrogen storage amount data, and hydrogen refueling demand data. The grid load demand data obtains the load prediction value for the next 24 hours through the load prediction system. The hydrogen production power data is obtained through the real-time operation status of the hydrogen production equipment. The hydrogen storage amount data is measured by the pressure sensor of the hydrogen storage tank. The hydrogen refueling demand data is predicted based on the historical operation data of the hydrogen refueling station.

[0100] Input the system state vector into the action value network. The action value network first statistically analyzes the probability distribution characteristics of the wind and light power generation prediction error based on historical data, and calculates the upper and lower limits of the 95% confidence interval as the adaptive boundary function. Taking the hydrogen production power as an example, when the wind and light power generation prediction error is large, the decision-making range of the hydrogen production power is appropriately shrunk through the adaptive boundary function to avoid decision-making risks caused by prediction errors. Specifically, if the standard deviation of the prediction error is 10%, the basic decision-making boundary of the hydrogen production power of 0 - 100% is dynamically adjusted to 10% - 90%.

[0101] Input the system state vector and the adjusted decision-making space into the target network. The target network includes a conditional risk assessment branch and a state value assessment branch. The conditional risk assessment branch calculates the risk degree of the system based on states such as the hydrogen storage amount and hydrogen refueling demand. For example, when the hydrogen storage amount is 20% lower than the lower limit of the safe storage amount, the risk value increases significantly. The state value assessment branch calculates the long-term benefits of each state. The outputs of the two branches are weighted with weights of 0.3 and 0.7 to obtain the comprehensive risk measurement value.

[0102] Build an optimization objective function, which includes three items: system operation cost, risk metric value, and renewable energy consumption. The operation cost includes the start-stop cost of hydrogen production equipment and the cost of purchasing electricity from the power grid. The renewable energy consumption is calculated based on the predicted values of wind and solar power. The parameters of the action value network are updated using the stochastic gradient descent method, and the target network updates its parameters once every 100 training steps with a soft update coefficient of 0.01.

[0103] Perform value function iteration calculations to obtain Q values. Construct a Lyapunov function. When the hydrogen storage amount deviates from the safe interval of 35%-85%, the quadratic term increases; when the wind and solar prediction error exceeds 10%, the high-order term increases. Calculate the time derivative of the Lyapunov function to obtain the stability margin, and its reciprocal is used as the risk constraint coefficient. The product of the Q value and the risk constraint coefficient is used as the risk constraint term.

[0104] In the adjusted decision space, subtract the risk constraint term from the Q value to obtain the decision evaluation value. Select the decision corresponding to the maximum evaluation value as the final scheduling instruction, including specific values such as hydrogen production power and hydrogen storage release power.

[0105] The present invention dynamically adjusts the decision space through the adaptive boundary function of the action value network, effectively coping with the uncertainty brought by wind and solar power generation prediction errors, and improving the robustness and reliability of decisions; the target network adopts a dual-branch structure of conditional risk and state value to comprehensively evaluate system risks, and constructs a risk constraint mechanism based on the Lyapunov stability theory to ensure the safe and stable operation of the system; the optimization objective function comprehensively considers economy, safety, and environmental protection, realizes economic and efficient operation on the premise of ensuring system safety, improves the level of renewable energy consumption, and has important practical application value.

[0106] In an alternative embodiment,

[0107] The adaptive boundary function includes:

[0108] Perform multi-scale decomposition on the wind and solar prediction error sequence through discrete wavelet transform to obtain short-term fluctuation errors and long-term trend errors;

[0109] Calculate the boundary adjustment feature of the short-term fluctuation error based on the local probability density feature of the prediction error sample;

[0110] Construct a cumulative impact function, which is obtained by weighted accumulation of the long-term trend errors in historical periods according to an exponential decay coefficient. Calculate the trend intensity index based on the cumulative impact function. The trend intensity index is the ratio of the absolute value of the cumulative impact function to the maximum value of the absolute value of the cumulative impact function within the evaluation time window;

[0111] Construct a probability confidence interval based on the boundary adjustment feature, map the boundary values of the probability confidence interval to a fluctuation boundary adjustment amount, and dynamically correct the fluctuation boundary adjustment amount based on the fluctuation direction change frequency; construct a piecewise linear mapping function based on the trend intensity index to calculate the trend boundary adjustment amount, and correct the trend boundary adjustment amount according to the trend duration.

[0112] Calculate an adaptive combination weight based on the trend intensity index, and the adaptive combination weight has an exponential relationship with the trend intensity index; weight-combine the fluctuation boundary adjustment amount and the trend boundary adjustment amount through the adaptive combination weight to obtain the adaptive boundary function.

[0113] Exemplarily, obtain a wind-solar power prediction error sequence and perform multi-scale decomposition through discrete wavelet transform. Select the db4 wavelet basis function to decompose the error sequence into 4 scales, obtaining short-term fluctuation errors and long-term trend errors. The short-term fluctuation errors correspond to high-frequency components, and the long-term trend errors correspond to low-frequency components. For example, decompose the 15-minute prediction error sequence of a certain photovoltaic power station to obtain fluctuation components reflecting the characteristics of 5 minutes, 15 minutes, 30 minutes, and 60 minutes.

[0114] Next, calculate the boundary adjustment feature of the short-term fluctuation error. Adopt the kernel density estimation method, use a 15-minute time window to slide and calculate the local probability density feature. For the error samples within each time window, estimate the probability density curve through the Gaussian kernel function, and extract the 95% confidence interval as the boundary adjustment feature. For example, the boundary adjustment feature at a certain moment is plus or minus 10% of the rated capacity.

[0115] Then construct an accumulated impact function. Select 24-hour historical data and weight-accumulate the long-term trend error according to an exponential decay coefficient of 0.95. The accumulated impact function reflects the influence degree of the historical trend error on the current boundary. Calculate the trend intensity index based on the accumulated impact function, and the value range is from 0 to 1. For example, the accumulated impact function at a certain moment is 8% of the rated capacity, and the maximum value within 24 hours is 10%, then the trend intensity index is 0.8.

[0116] Further construct an adaptive boundary. The fluctuation boundary adjustment amount is mapped from the probability confidence interval and is compressed and corrected when the fluctuation direction change frequency is high. The trend boundary adjustment amount is calculated through a piecewise linear mapping function, and the longer the trend duration, the greater the adjustment amount. The two boundary adjustment amounts are weighted through the adaptive combination weight to obtain the final boundary. The combination weight has an exponential relationship with the trend intensity index, and the trend boundary weight increases when the trend is significant. For example, at a certain moment, the trend intensity is 0.8, the fluctuation boundary adjustment amount is 8%, the trend boundary adjustment amount is 12%, the combination weights are 0.3 and 0.7 respectively, and the final boundary adjustment amount is 10.8%.

[0117] Existing methods for adjusting the prediction error boundary of wind and light mainly adopt fixed boundaries or simple adaptive boundaries. The present invention proposes a method for constructing an adaptive boundary function, innovatively introducing a cumulative influence function to characterize the influence of historical trends, and designing a trend intensity index to quantitatively represent trend characteristics. By using probability confidence intervals and piecewise linear mapping to construct fluctuation boundaries and trend boundaries respectively, and realizing an adaptive weight combination based on the trend intensity index. The present invention achieves an accurate characterization of the multi-scale characteristics of wind and light prediction errors, and significantly improves the adaptability and flexibility of boundary adjustment. Compared with the prior art, the boundary prediction accuracy is improved, the dynamic tracking performance is enhanced, the system's response to short-term fluctuations and long-term trends is more accurate, providing strong support for the precise regulation of the power-to-hydrogen system.

[0118] In an alternative embodiment,

[0119] The target network calculates the conditional risk value and the state value, and the weighted combination of the conditional risk value and the state value to obtain the risk metric value includes:

[0120] Construct a system energy function with multi-dimensional state coupling based on the system state vector and the adjusted decision space. The system energy function is a non-linear weighted combination of grid load demand data, hydrogen production power data, hydrogen storage capacity data, and hydrogen refueling demand data and their corresponding reference values;

[0121] Construct a conditional risk branch and a state value branch of the target network;

[0122] Input the adjusted decision space and the historical decision sequence into the conditional risk branch, calculate the attention weight matrix based on the multi-level correlation degree between the system state vector and the historical state, and use the attention weight matrix and the spatio-temporal convolution result of the historical state as the conditional risk value;

[0123] Input the system state vector and the adjusted decision space into the state value branch, construct a posterior distribution of the state value using the probabilistic variational inference method, and extract the conditional expectation and uncertainty of the posterior distribution to obtain the state value;

[0124] Determine the adaptive time window length based on the system energy function. The adaptive time window length is not less than the response time of the power-to-hydrogen coupling system and not greater than a preset multiple of the time constant of the hydrogen storage system;

[0125] Obtain the maximum and minimum values of the power-to-hydrogen coupling system energy within the adaptive time window, and calculate the multi-scale stability index based on the difference between the current system energy value and the maximum and minimum values;

[0126] Construct an adaptive weighting coefficient according to the multi-scale stability index, and non-linearly weight the conditional risk value and the state value according to the adaptive weighting coefficient to obtain the risk metric value.

[0127] Exemplarily, a system energy function with multi-dimensional state coupling is constructed based on the system state vector and the adjusted decision space. This energy function includes four dimensions: power grid load demand data, hydrogen production power data, hydrogen storage capacity data, and hydrogen refueling demand data. The difference between the data of each dimension and its corresponding reference value is calculated, and the system energy function is obtained through non-linear weighted combination. The reference value can be obtained based on historical data statistics. For example, the reference value of the power grid load demand takes the average value of the same period in history, the reference value of the hydrogen production power takes 80% of the rated power of the equipment, the reference value of the hydrogen storage capacity takes 50% of the capacity of the hydrogen storage tank, and the reference value of the hydrogen refueling demand takes 70% of the historical peak value.

[0128] A dual-branch target network is constructed, including a conditional risk branch and a state value branch. The conditional risk branch adopts a multi-head attention mechanism, taking the adjusted decision space and the historical decision sequence as inputs. The attention weight matrix is calculated based on the multi-level correlation degree between the system state vector and the historical state. Specifically, the state vector is divided into three levels according to the time scale: hourly, daily, and weekly. The similarity between the current state and the historical state is calculated respectively to obtain the weight coefficients. The historical state is processed by spatio-temporal convolution to extract spatio-temporal features, which are multiplied by the attention weight matrix to obtain the conditional risk value.

[0129] The state value branch adopts the probabilistic variational inference method to map the system state vector and the adjusted decision space to the latent variable space. The posterior distribution of the state value is constructed through a multi-layer neural network, and the mean value of the distribution is extracted as the conditional expectation, and the variance is used as the uncertainty measure. The two together constitute the state value.

[0130] The adaptive time window length is determined based on the system energy function. The start and stop time of the electrolytic water hydrogen production equipment is generally at the minute level, and the hydrogen charging and discharging process of the hydrogen storage system lasts for several hours. Therefore, the lower limit of the window length is taken as 5 minutes, and the upper limit is taken as 3 times the time constant of the hydrogen storage system. The maximum and minimum values of the system energy are statistically calculated within the adaptive window, and the multi-scale stability index is calculated based on the normalized difference between the current energy value and the extreme values. This index reflects the stability degree of the system at different time scales.

[0131] An adaptive weighting coefficient is constructed according to the multi-scale stability index. When the stability is good, the state value weight is larger; when the stability is poor, the conditional risk weight is larger. A non-linear function is used to map the stability index to the 0-1 interval to obtain the weighting coefficient. Finally, the conditional risk value and the state value are non-linearly combined according to the weighting coefficient to obtain the risk metric value.

[0132] By constructing a system energy function with multi-dimensional state coupling, the operation state of the electric-hydrogen coupling system is comprehensively characterized, improving the accuracy and reliability of risk measurement; adopting a dual-branch target network structure, the system risk is evaluated from two dimensions of conditional risk and state value respectively, and the weights are adaptively adjusted based on system stability, making the risk measurement more reasonable and scientific; introducing an adaptive time window mechanism, the analysis scale is automatically adjusted according to the dynamic characteristics of the system, improving the adaptability and robustness of risk measurement to the dynamic changes of the system.

[0133] In an alternative embodiment,

[0134] The risk constraint coefficient includes:

[0135] Obtain the real-time value of hydrogen storage, the upper and lower limits of the hydrogen storage safety interval, and the wind-solar prediction error sequence; construct the degree value of the deviation of the hydrogen storage amount from the safety interval as a quadratic term, and construct the wind-solar prediction error value as a high-order term to form a Lyapunov function representing the system stability; calculate the stability margin based on the derivative of the Lyapunov function with respect to time, and take the reciprocal of the stability margin as the risk constraint coefficient.

[0136] Exemplarily, for the method of determining the risk constraint coefficient of a wind-solar-hydrogen storage system, first obtain the real-time value of the hydrogen storage amount in the hydrogen storage tank through real-time monitoring, and at the same time determine the upper and lower limit thresholds of the hydrogen storage safety interval according to the system design parameters. The real-time value of the hydrogen storage amount is generally in standard cubic meters and is collected in real time through a hydrogen flow meter and a pressure sensor. For example, the hydrogen storage tank capacity of a certain system is 1000 standard cubic meters, the upper limit of the safety interval is 900 standard cubic meters, and the lower limit is 100 standard cubic meters. Obtain the predicted values of wind-solar power output for future periods through the wind-solar power prediction system, compare the predicted values with the actual values, and obtain the prediction error sequence. For example, the prediction error sequence of a certain wind-solar system for the next 4 hours is [-50kW, 30kW, -20kW, 40kW]. Square the deviation values between the real-time value of the hydrogen storage amount and the upper and lower limits of the safety interval to construct a quadratic term. When the hydrogen storage amount is 600 standard cubic meters, the square of the deviation from the upper limit of 900 standard cubic meters is 90000, and the square of the deviation from the lower limit of 100 standard cubic meters is 250000. Perform high-order processing on the wind-solar prediction error sequence to construct a high-order term. Perform a fourth-power operation on each value in the prediction error sequence to obtain a high-order error sequence. The high-order results of the above prediction error sequence are [6250000, 810000, 160000, 2560000]. Combine the quadratic term and the high-order term with weights to form a Lyapunov function representing the system stability. Take the derivative of this function to obtain the stability margin, and then calculate its reciprocal as the risk constraint coefficient. For example, the stability margin calculated at a certain moment is 0.8, and the corresponding risk constraint coefficient is 1.25.

[0137] The present invention constructs a Lyapunov function through the quadratic term of the hydrogen storage capacity deviation and the high-order term of the wind and light prediction error, comprehensively considering the key influencing factors of system stability, making the risk constraint more accurate and reasonable; calculates the risk constraint coefficient based on the Lyapunov stability theory, organically combines the system dynamic characteristics and stability requirements, and ensures the theoretical basis and practicability of the constraint; uses the real-time monitoring of the hydrogen storage capacity and the wind and light prediction error sequence as inputs, realizes the real-time dynamic adjustment of the risk constraint, and improves the safety and reliability of the system operation.

[0138] In an alternative embodiment,

[0139] Based on the dynamic consistency theory, the coupling strength of the distributed cooperative network is inversely mapped to the risk constraint coefficient, and the synchronization potential function is constructed by the deviation between the node state of the distributed cooperative network and the scheduling instruction and the state difference of adjacent nodes, including:

[0140] Construct a dynamic coupling strength based on the risk constraint coefficient, where the dynamic coupling strength is the ratio of the basic coupling strength to the risk adjustment term, the risk adjustment term increases with the increase of the risk constraint coefficient, and the dynamic coupling strength also has an exponential decay relationship with the state difference of adjacent nodes;

[0141] Construct a synchronization potential function, which includes a scheduling instruction tracking term and a node cooperation term. The scheduling instruction tracking term is the product of the square difference between the node state and the scheduling instruction and the first weight coefficient, and the node cooperation term is the product of the square of the state difference of adjacent nodes, the dynamic coupling strength, and the second weight coefficient;

[0142] Construct a distributed control law based on the gradient of the synchronization potential function. The distributed control law includes a scheduling tracking term and a cooperation adjustment term. The scheduling tracking term is the product of the deviation between the node state and the scheduling instruction and the first weight coefficient, and the cooperation adjustment term is the product of the weighted deviation between the node state and the state of adjacent nodes and the second weight coefficient;

[0143] When the actual power generation of the wind farm and the photovoltaic power station exceeds the predicted power generation, increase the hydrogen production power based on the distributed control law; when the actual power generation is less than the predicted power generation, increase the hydrogen release power of the hydrogen storage system based on the distributed control law, and perform real-time cooperative regulation on the electric-hydrogen system.

[0144] Exemplarily, the distributed cooperative network control method based on the dynamic consistency theory realizes the real-time cooperative regulation of the electric-hydrogen system by constructing the dynamic coupling strength and the synchronization potential function.

[0145] Construct the dynamic coupling strength. The dynamic coupling strength is obtained by dividing the basic coupling strength by the risk adjustment term. The basic coupling strength is set to a fixed value, such as 0.8. The risk adjustment term is positively correlated with the risk constraint coefficient. For every 0.1 increase in the risk constraint coefficient, the risk adjustment term increases by 0.2. When the difference in the states of adjacent nodes increases, the dynamic coupling strength decays exponentially. For example, when the difference in the states of adjacent nodes is 10%, the dynamic coupling strength decays to 0.9 times the original value.

[0146] Construct the synchronization potential function. The synchronization potential function includes two terms: the scheduling instruction tracking term and the node cooperation term. The scheduling instruction tracking term is the square of the deviation between the node state and the scheduling instruction multiplied by the first weight coefficient, and the first weight coefficient is 0.6. The node cooperation term is the square of the difference in the states of adjacent nodes multiplied by the product of the dynamic coupling strength and the second weight coefficient, and the second weight coefficient is 0.4.

[0147] Construct the distributed control law based on the gradient of the synchronization potential function. The distributed control law includes the scheduling tracking term and the cooperation adjustment term. The scheduling tracking term is the deviation between the node state and the scheduling instruction multiplied by the first weight coefficient. The cooperation adjustment term is the weighted deviation between the node state and the state of the adjacent node multiplied by the second weight coefficient.

[0148] Achieve real-time cooperative regulation of the electric-hydrogen system. When the actual power generation of the wind farm exceeds the predicted value by 10%, the hydrogen production power increases by 8%. When the actual power generation of the photovoltaic power station is 15% lower than the predicted value, the hydrogen release power of the hydrogen storage system increases by 12%. Through the distributed control law, the cooperative regulation of each node is realized, so that the overall system reaches synchronization.

[0149] Through the inverse mapping between the dynamic coupling strength and the risk constraint coefficient, the present invention realizes the adaptive adjustment of the system risk and the control strength, improves the safety and reliability of the system operation; constructs the distributed control law based on the synchronization potential function, realizes the accurate tracking of the node state and the scheduling instruction, and the cooperation and consistency between adjacent nodes, and ensures the synchronization stability of the system; applies the distributed control to the real-time cooperative regulation of the electric-hydrogen system, effectively suppresses the fluctuations of new energy power generation, and improves the operation economy and flexibility of the system.

[0150] In an optional implementation manner,

[0151] The synchronization potential function includes a fast response layer and a steady-state tracking layer, including:

[0152] The scheduling instruction tracking term of the fast response layer uses the product of the square difference between the node state and the scheduling instruction and the first weight coefficient, and the node cooperation term of the fast response layer uses the product of the square of the difference between adjacent node states, the dynamic coupling strength, and the second weight coefficient; the fast response layer triggers calculation and update based on the change rate of the node state, and when the change rate of the node state exceeds the preset threshold, the first weight coefficient and the second weight coefficient are updated.

[0153] The scheduling instruction tracking term of the steady-state tracking layer uses the product of the cumulative square difference between the node state and the scheduling instruction and the first weight coefficient, where the cumulative square difference is the time integral of the square difference between the node state and the scheduling instruction within a fixed time window, and the length of the fixed time window is not less than 10 times the calculation period of the fast response layer and not greater than one-third of the system characteristic time, where the system characteristic time is the dynamic response time of the electro-hydrogen system; the node cooperation term of the steady-state tracking layer uses the product of the square of the difference between adjacent node states, the dynamic coupling strength, and the second weight coefficient.

[0154] Based on the calculation results of the fast response layer and the steady-state tracking layer, an inter-layer coupling weight is constructed. When the change rate of the node state is greater than the first preset threshold, the weight of the fast response layer is increased, and when the change rate of the node state is less than the second preset threshold, the weight of the steady-state tracking layer is increased.

[0155] Exemplarily, the synchronous potential function includes a fast response layer and a steady-state tracking layer, and realizes the cooperative control of the electro-hydrogen system through a two-layer structure.

[0156] The implementation process of the fast response layer is as follows: First, obtain the node state and the scheduling instruction, calculate the square difference between the node state and the scheduling instruction, and multiply it by the first weight coefficient to obtain the scheduling instruction tracking term. At the same time, calculate the square of the difference between adjacent node states, and multiply it by the dynamic coupling strength and the second weight coefficient to obtain the node cooperation term. When the change rate of the node state exceeds 0.1, update the first weight coefficient to 1.2 times the original value and the second weight coefficient to 0.8 times the original value to achieve fast response.

[0157] The implementation process of the steady-state tracking layer is as follows: Accumulatively calculate the square difference between the node state and the scheduling instruction within a fixed time window (set to 100 seconds, greater than 10 times the 10-second calculation period of the fast response layer and less than one-third of the 300-second system characteristic time), and multiply the cumulative value by the first weight coefficient to obtain the scheduling instruction tracking term. The calculation method of the node cooperation term is the same as that of the fast response layer.

[0158] The construction process of the inter-layer coupling weight is as follows: The node state change rate is monitored in real time. When the change rate is greater than 0.2, the weight of the fast response layer increases to 0.8, and the weight of the steady-state tracking layer decreases to 0.2; when the change rate is less than 0.05, the weight of the fast response layer decreases to 0.3, and the weight of the steady-state tracking layer increases to 0.7. Finally, the calculation results of the two layers are weighted by the weight to obtain the control output.

[0159] Figure 3 It is a comparison chart of the performance of the double-layer control of the present invention and other single-layer controls. As Figure 3 shown, the traditional control method of the electric-hydrogen system usually adopts a single control hierarchical structure, which is difficult to simultaneously consider the fast response ability and steady-state tracking accuracy of the system. The existing technologies either have a fast response speed but poor steady-state accuracy, or have high steady-state accuracy but slow dynamic response, and cannot meet the dual requirements of the electric-hydrogen system for fast regulation and stable operation. The present invention proposes a double-layer synchronous potential function structure, which innovatively divides the control into a fast response layer and a steady-state tracking layer. The fast response layer adopts a weight update mechanism triggered by the change rate to ensure the fast response of the system to disturbances; the steady-state tracking layer designs through the cumulative squared difference and a fixed time window to ensure the control accuracy. Through the dynamic adjustment of the inter-layer coupling weight, seamless switching between the two-layer controls is achieved. The present invention significantly improves the control performance of the electric-hydrogen system, realizing the unity of fast response and steady-state accuracy. The system response time is shortened, the steady-state control accuracy is improved, and the dynamic disturbance suppression ability is enhanced, providing an effective guarantee for the efficient coordinated operation of the electric-hydrogen system. The response speed of the present invention is increased by 70% compared with the traditional single-layer steady-state control, the control accuracy is increased by 87% compared with the single-layer fast control, the system stability is increased by 45% and supports adaptive regulation of the change rate within the range of 0.05 - 0.25.

[0160] In the second aspect of the embodiment of the present invention,

[0161] an electric-hydrogen coupled intelligent regulation system considering wind-solar prediction errors is provided, including:

[0162] A first unit for collecting historical power generation data of a wind farm and a photovoltaic power station, calculating the spatio-temporal correlation characteristics of the wind-solar prediction errors, and obtaining the probability distribution of the wind-solar prediction errors by using an adaptive combined kernel density estimation method of a Gaussian kernel function and an Epanechnikov kernel function;

[0163] A second unit for using the probability distribution as an uncertainty constraint, combining grid load demand data and hydrogen storage system operation data, and performing deep reinforcement learning by using a dual neural network, wherein the action value network dynamically adjusts the decision space based on an adaptive boundary function of the short-term fluctuation and long-term trend of the wind-solar prediction errors; the target network evaluates risks through a conditional risk and state value double-branch, and constructs a risk constraint coefficient based on the Lyapunov stability theory for decision risk constraint to obtain a scheduling instruction.

[0164] A third unit is configured to construct the electrolytic hydrogen production unit and the hydrogen storage system into a distributed collaborative network, inversely map the coupling strength of the distributed collaborative network to the risk constraint coefficient based on the dynamic consistency theory, and construct a synchronization potential function from the deviation between the node state of the distributed collaborative network and the scheduling instruction and the state difference of adjacent nodes; and control the hydrogen production power and the hydrogen release power in real time along the gradient direction of the synchronization potential function based on the deviation between the actual power generation of the wind farm and the photovoltaic power station and the predicted power generation.

[0165] In the third aspect of the embodiments of the present invention,

[0166] An electronic device is provided, including:

[0167] A processor;

[0168] A memory for storing instructions executable by the processor;

[0169] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0170] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An electric-hydrogen coupling intelligent control method considering wind-solar forecast errors is characterized by: include: Collect historical power generation data of wind farms and photovoltaic power stations, calculate the spatiotemporal correlation characteristics of wind farm and photovoltaic power station prediction errors, and use the adaptive combined kernel density estimation method of Gaussian kernel function and Epanechnikov kernel function to obtain the probability distribution of wind and solar power prediction errors; Taking the probability distribution as uncertainty constraint, combined with grid load demand data and hydrogen storage system operation data, a dual neural network is used to perform deep reinforcement learning, in which the action value network dynamically adjusts the decision space based on the adaptive boundary function of the short-term fluctuation and long-term trend of the wind and solar forecast error; The target network evaluates risk through the dual branches of conditional risk and state value, and constructs risk constraint coefficients based on Lyapunov stability theory to perform decision-making risk constraints and obtain scheduling instructions; The hydrogen production unit by electrolysis and the hydrogen storage system are constructed as a distributed collaborative network. Based on the dynamic consistency theory, the coupling strength of the distributed collaborative network is inversely mapped to the risk constraint coefficient. The deviation between the node state of the distributed collaborative network and the dispatch instruction and the state difference of the adjacent nodes are used to construct a synchronous potential function. Based on the deviation between the actual power generation of the wind farm and the predicted power generation of the photovoltaic power station, the hydrogen production power and the hydrogen release power are controlled in real time along the gradient direction of the synchronous potential function.

2. The method according to claim 1, characterized in that The spatiotemporal correlation characteristics of the prediction errors of wind farms and photovoltaic power stations are calculated, and the probability distribution of wind and solar prediction errors is obtained by using the adaptive combined kernel density estimation method of Gaussian kernel function and Epanechnikov kernel function, including: Calculate the time correlation coefficient and space correlation coefficient of the prediction error data of wind farms and photovoltaic power stations to obtain the spatiotemporal coupling characteristics; A Gaussian kernel function and an Epanechnikov kernel function are selected as basic kernel functions, wherein the Gaussian kernel function is used for probability density estimation in the central area, and the Epanechnikov kernel function is used for probability density estimation in the boundary area; adaptive weights of the Gaussian kernel function and the Epanechnikov kernel function are calculated based on the local sample density, and the product of the Gaussian kernel function and the adaptive weight and the product of the Epanechnikov kernel function and the complement of the adaptive weight are taken as the combined kernel function; The initial bandwidth is calculated based on the standard deviation and interquartile range of historical power generation data, and the initial bandwidth is adjusted by quantile constraints according to the empirical distribution function to obtain the optimized bandwidth; The spatiotemporal coupling characteristics, combined kernel function and optimized bandwidth are substituted into the kernel density estimation formula to obtain the probability density function of the wind and solar power generation prediction error; the prediction error samples are updated based on the sliding time window, the sample weights are calculated according to the time attenuation factor, and the sample weights and the probability density function are recursively updated to obtain the dynamically updated wind and solar power prediction error probability distribution.

3. The method according to claim 1, characterized in that A dual neural network is used to perform deep reinforcement learning, where the action-value network dynamically adjusts the decision space based on an adaptive boundary function of the short-term fluctuations and long-term trends of the wind and solar forecast errors; The target network evaluates risk through conditional risk and state value dual branches, and constructs risk constraint coefficients based on Lyapunov stability theory for decision-making. Risk constraints include: A dual neural network is used to perform deep reinforcement learning, in which the action value network dynamically adjusts the decision space based on the adaptive boundary function of the short-term fluctuation and long-term trend of the wind and solar forecast error; the target network evaluates the risk through the dual branches of conditional risk and state value, and constructs the risk constraint coefficient based on the Lyapunov stability theory to constrain the decision risk. Acquiring a system state vector, wherein the system state vector includes power grid load demand data, hydrogen production power data, hydrogen storage data, and hydrogenation demand data; The system state vector is input into an action value network, the action value network calculates an adaptive boundary function based on the probability distribution of the wind and solar power prediction error, and dynamically adjusts the basic decision boundary according to the adaptive boundary function to obtain an adjusted decision space; Inputting the system state vector and the adjusted decision space into a target network, the target network calculating a conditional risk value and a state value, and weighting the conditional risk value and the state value to obtain a risk metric value; Constructing an optimization objective function including system operation cost items, risk measurement items and renewable energy consumption items, and updating action value network parameters based on the optimization objective function by using a stochastic gradient descent method. The target network evaluates decision risks based on the risk measurement value, and periodically updates network parameters by using a preset soft update coefficient. Perform iterative calculation of the value function according to the updated action value network parameters and risk metric values, and use the product of the iterative calculation result and the risk constraint coefficient as the risk constraint term. The risk constraint coefficient is obtained by constructing a Lyapunov function based on the quadratic term of the hydrogen storage deviation from the safe interval and the high-order term of the wind and solar prediction error and calculating the inverse of the stability margin through its time derivative; Based on the adjusted decision space, a decision evaluation value considering the risk constraint item is calculated, and a decision corresponding to the maximum decision evaluation value is selected as a scheduling instruction.

4. The method according to claim 3, characterized in that The adaptive boundary function includes: The wind power forecast error sequence is decomposed into short-term fluctuation error and long-term trend error through discrete wavelet transform. Calculating the boundary adjustment characteristics of the short-term volatility error based on the local probability density characteristics of the forecast error samples; Construct a cumulative influence function, which is obtained by weighted accumulation of long-term trend errors in historical periods according to an exponential decay coefficient, and calculate a trend strength index based on the cumulative influence function, which is the ratio of the absolute value of the cumulative influence function to the maximum absolute value of the cumulative influence function within the evaluation time window; A probability confidence interval is constructed based on the boundary adjustment feature, and the boundary value of the probability confidence interval is mapped to a fluctuation boundary adjustment amount, and the fluctuation boundary adjustment amount is dynamically corrected based on the frequency of fluctuation direction change; a piecewise linear mapping function is constructed based on the trend strength indicator to calculate the trend boundary adjustment amount, and the trend boundary adjustment amount is corrected according to the duration of the trend; An adaptive combination weight is calculated based on the trend strength index, and the adaptive combination weight is exponentially related to the trend strength index; the fluctuation boundary adjustment amount and the trend boundary adjustment amount are weightedly combined by the adaptive combination weight to obtain the adaptive boundary function.

5. The method according to claim 3, characterized in that: The target network calculates a conditional risk value and a state value, and weights the conditional risk value and the state value to obtain a risk metric value, including: Constructing a multi-dimensional state-coupled system energy function based on the system state vector and the adjusted decision space, wherein the system energy function is a nonlinear weighted combination of grid load demand data, hydrogen production power data, hydrogen storage data, and hydrogenation demand data and their corresponding reference values; Construct the conditional risk branch and state value branch of the target network; Input the adjusted decision space and historical decision sequence into the conditional risk branch, calculate the attention weight matrix based on the multi-level correlation degree between the system state vector and the historical state, and use the spatiotemporal convolution result of the attention weight matrix and the historical state as the conditional risk value; Input the system state vector and the adjusted decision space into the state value branch, construct the posterior distribution of the state value using the probabilistic variational inference method, and extract the conditional expectation and uncertainty of the posterior distribution to obtain the state value; Determining an adaptive time window length based on the system energy function, wherein the adaptive time window length is not less than a response time of the electric-hydrogen coupling system and not greater than a preset multiple of a time constant of the hydrogen storage system; Acquire the maximum and minimum values ​​of the energy of the electric-hydrogen coupling system within the adaptive time window, and calculate the multi-scale stability index based on the difference between the current system energy value and the maximum and minimum values; An adaptive weighting coefficient is constructed according to the multi-scale stability index, and the conditional risk value and the state value are nonlinearly weighted according to the adaptive weighting coefficient to obtain a risk measurement value.

6. The method according to claim 3, characterized in that: The risk constraint coefficients include: The real-time value of hydrogen storage, the upper and lower limits of the hydrogen storage safety interval, and the wind and solar power prediction error sequence are obtained; the degree of deviation of the hydrogen storage from the safety interval is constructed as a quadratic term, and the wind and solar power prediction error value is constructed as a high-order term to form a Lyapunov function that characterizes the stability of the system; the stability margin is calculated based on the derivative of the Lyapunov function with respect to time, and the inverse of the stability margin is used as the risk constraint coefficient.

7. The method according to claim 1, characterized in that Based on the dynamic consistency theory, the coupling strength of the distributed collaborative network is inversely mapped to the risk constraint coefficient, and the synchronization potential function is constructed by the deviation between the node state of the distributed collaborative network and the scheduling instruction and the state difference of the adjacent nodes, including: Constructing a dynamic coupling strength based on the risk constraint coefficient, wherein the dynamic coupling strength is a ratio of the basic coupling strength to the risk adjustment term, the risk adjustment term increases as the risk constraint coefficient increases, and the dynamic coupling strength also has an exponential decay relationship with the state difference of the adjacent nodes; Constructing a synchronization potential function, wherein the synchronization potential function includes a scheduling instruction tracking item and a node coordination item, wherein the scheduling instruction tracking item is the product of the square difference between the node state and the scheduling instruction and the first weight coefficient, and the node coordination item is the product of the square of the difference between the states of adjacent nodes and the dynamic coupling strength and the second weight coefficient; Constructing a distributed control law based on the gradient of the synchronization potential function, wherein the distributed control law includes a scheduling tracking item and a coordinated adjustment item, wherein the scheduling tracking item is the product of the deviation between the node state and the scheduling instruction and the first weight coefficient, and the coordinated adjustment item is the product of the weighted deviation between the node state and the adjacent node state and the second weight coefficient; When the actual power generation of the wind farm and the photovoltaic power station exceeds the predicted power generation, the hydrogen production power is increased based on the distributed control law; when the actual power generation is less than the predicted power generation, the hydrogen release power of the hydrogen storage system is increased based on the distributed control law, and the electric hydrogen system is coordinated and regulated in real time.

8. The method according to claim 7, characterized in that The synchronous potential function includes a fast response layer and a steady-state tracking layer, including: The scheduling instruction tracking item of the fast response layer adopts the product of the square difference between the node state and the scheduling instruction and the first weight coefficient, and the node coordination item of the fast response layer adopts the product of the square of the difference between the states of adjacent nodes and the dynamic coupling strength and the second weight coefficient; the fast response layer triggers calculation updates based on the change rate of the node state, and updates the first weight coefficient and the second weight coefficient when the change rate of the node state exceeds a preset threshold; The scheduling instruction tracking item of the steady-state tracking layer adopts the product of the cumulative square difference between the node state and the scheduling instruction and the first weight coefficient, wherein the cumulative square difference is the time integral of the square difference between the node state and the scheduling instruction in a fixed time window, and the length of the fixed time window is not less than 10 times the calculation cycle of the fast response layer and not more than one third of the system characteristic time, wherein the system characteristic time is the dynamic response time of the electric hydrogen system; the node coordination item of the steady-state tracking layer adopts the product of the square of the difference in the states of adjacent nodes, the dynamic coupling strength and the second weight coefficient; The inter-layer coupling weight is constructed based on the calculation results of the fast response layer and the steady-state tracking layer. When the change rate of the node state is greater than a first preset threshold, the weight of the fast response layer is increased. When the change rate of the node state is less than a second preset threshold, the weight of the steady-state tracking layer is increased.

9. An electric-hydrogen coupled intelligent control system considering wind-solar forecast errors, used to implement the method described in any one of claims 1 to 8, characterized in that: include: The first unit is used to collect historical power generation data of wind farms and photovoltaic power stations, calculate the spatiotemporal correlation characteristics of wind farm and photovoltaic power station prediction errors, and use the adaptive combined kernel density estimation method of Gaussian kernel function and Epanechnikov kernel function to obtain the probability distribution of wind and solar prediction errors; The second unit is used to use the probability distribution as an uncertainty constraint, combine the power grid load demand data and the hydrogen storage system operation data, and use a dual neural network to perform deep reinforcement learning, wherein the action value network dynamically adjusts the decision space based on the adaptive boundary function of the short-term fluctuation and long-term trend of the wind and solar forecast error; The target network evaluates risk through the dual branches of conditional risk and state value, and constructs risk constraint coefficients based on Lyapunov stability theory to perform decision-making risk constraints and obtain scheduling instructions; The third unit is used to construct the hydrogen production unit by electrolysis and the hydrogen storage system into a distributed collaborative network, map the coupling strength of the distributed collaborative network inversely to the risk constraint coefficient based on the dynamic consistency theory, and construct a synchronous potential function based on the deviation between the node state of the distributed collaborative network and the dispatch instruction and the state difference of the adjacent nodes; based on the deviation between the actual power generation of the wind farm and the photovoltaic power station and the predicted power generation, the hydrogen production power and the hydrogen release power are controlled in real time along the gradient direction of the synchronous potential function.

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 described in any one of claims 1 to 8.

Images