A wind power off-grid hydrogen production power supply topology and control method without step-down transformer

Abstract

The present invention provides a wind power off-grid hydrogen production power supply topology and control method without a step-down transformer, including: obtaining wind speed data, using a wavelet threshold algorithm for denoising, predicting power based on an LSTM-GRU hybrid neural network, and extracting power fluctuation characteristics through multi-scale entropy analysis; obtaining load data, using Kalman filtering and recursive least squares to establish an impedance model, and extracting dynamic characteristics in combination with Hilbert transform; using a differential evolution algorithm for multi-objective optimization to generate an optimal switching sequence and power allocation strategy; achieving system control through space vector modulation, balanced control, soft switch control and harmonic compensation; achieving state estimation and disturbance suppression based on a sliding mode observer and a nonlinear observer; and achieving fault diagnosis and protection using a deep belief network. The present invention improves system efficiency and reliability and improves dynamic performance.

Description

Technical Field

[0001] The invention relates to power electronics technology, in particular to a wind power off-grid hydrogen production power supply topology and a control method without a step-down transformer. Background Art

[0002] Driven by the global energy transformation and carbon neutrality goals, the use of wind energy to produce hydrogen is becoming an important direction for the use of clean energy. The wind power hydrogen production system converts the fluctuating wind energy into stable and storable hydrogen energy, which not only solves the problem of new energy consumption, but also provides important support for the future hydrogen energy society. Especially in remote areas and independent microgrid scenarios, the wind power off-grid hydrogen production system without a step-down transformer has great application potential due to its simple structure, low cost, and high efficiency. It is of great significance to promote the efficient use of renewable energy and the development of distributed energy systems.

[0003] At present, the research on wind power hydrogen production system mainly focuses on traditional topology and basic control strategy. Conventional schemes usually adopt step-down transformer and rectifier-inverter dual conversion structure, and realize system operation through simple PI control and MPPT control based on table lookup method. Power prediction mainly relies on a single neural network model or statistical method, and the prediction accuracy is limited. Load modeling mostly adopts static impedance model, which is difficult to reflect the dynamic characteristics of electrolyzer. System control is mainly based on decoupling control and feedforward compensation, and lacks in-depth consideration of the overall characteristics of the system. Fault diagnosis is mainly based on threshold judgment and simple feature extraction method, and the diagnostic ability is limited.

[0004] However, the existing technical solutions have many technical difficulties that need to be solved in practical applications: first, the random noise and measurement errors in the wind speed data lead to insufficient power prediction accuracy, especially when the wind speed changes rapidly, the prediction error can reach more than 30%; second, the nonlinear characteristics and temperature dependence of the electrolyzer load make the accuracy of the traditional impedance model low, and the model error exceeds 15%; third, the efficiency, response time and harmonics are not fully considered in the power allocation optimization process, resulting in the failure to achieve optimal system performance; fourth, the branch current imbalance of the parallel Buck circuit is as high as 20%, affecting the system efficiency and reliability; fifth, the switching loss and electromagnetic interference problems are prominent, and the system efficiency loss is more than 10%; sixth, the system harmonic compensation effect is not ideal, and the total harmonic distortion rate is difficult to control below 5%; finally, the existing fault diagnosis method is insufficient in identifying complex faults, and the misdiagnosis rate exceeds 15%. These problems seriously restrict the engineering application and promotion of the off-grid hydrogen production system of wind power without step-down transformer. Summary of the invention

[0005] The purpose of the invention is to provide a wind power off-grid hydrogen production power supply topology and control method without a step-down transformer, in order to solve at least one technical problem existing in the prior art.

[0006] Technical solution: A control method for off-grid wind power hydrogen production power topology without a step-down transformer, comprising the following steps:

[0007] Step S1, obtain wind speed data, use wavelet threshold algorithm to denoise, predict power based on LSTM-GRU hybrid neural network, and extract power fluctuation characteristics through multi-scale entropy analysis;

[0008] Specifically, step S1, obtaining original wind speed data from a measuring device, removing data noise according to a wavelet threshold algorithm to obtain processed wind speed; using the processed wind speed to calculate theoretical power; inputting theoretical power and historical power data into an LSTM-GRU hybrid neural network to generate a predicted power sequence and a predicted error; performing multi-scale entropy analysis on the predicted power sequence and the predicted error to obtain a power fluctuation characteristic;

[0009] Step S2, acquiring load data, establishing an impedance model using Kalman filtering and recursive least squares method, and extracting dynamic features in combination with Hilbert transform;

[0010] Specifically, step S2, obtaining load current, load voltage and temperature data from the hydrogen production system, using the Kalman filter algorithm to obtain a filtered data sequence; using the filtered data sequence to perform a recursive least squares operation to obtain a load impedance model; performing a Hilbert-Huang transform on the load impedance model to extract load dynamic feature quantities;

[0011] Step S3: using differential evolution algorithm to perform multi-objective optimization and generate optimal switching sequence and power allocation strategy;

[0012] Specifically, step S3, input the predicted power sequence and power fluctuation characteristic quantity obtained in step S1 and the load impedance model and load dynamic characteristic quantity obtained in step S2 into the multi-objective optimizer to construct an objective function including efficiency, response time and harmonic rate; set constraints according to system operation parameters; use differential evolution algorithm to solve the optimization problem and obtain the optimal switching sequence and power allocation strategy;

[0013] Step S4, realizing system control through space vector modulation, equalization control, soft switch control and harmonic compensation;

[0014] Specifically, step S4 collects the phase current and bus voltage in the three-phase circuit, combines the power allocation strategy obtained in step S3, and calculates the conduction angle and switching sequence through space vector modulation; performs balanced control calculation on the branch currents of multiple groups of Buck circuits to obtain balanced control quantities; detects the voltage signal and current signal of the switch tube to generate a soft switching timing; analyzes and compensates the harmonic components in the current, and outputs a harmonic compensation signal;

[0015] Step S5, realizing state estimation and disturbance suppression based on sliding mode observer and nonlinear observer;

[0016] Specifically, step S5, real-time acquisition of the state quantity of the system, obtaining the state estimation value through the sliding mode observer; designing a nonlinear observer according to the error of the state estimation value, and calculating the compensation amount; inputting the state estimation value and the compensation amount into the adaptive controller to generate the control amount;

[0017] Step S6: During operation, a deep belief network is used to implement fault diagnosis and protection.

[0018] Specifically, in step S6, during operation, the system operation status data is collected, and a feature vector is obtained through empirical mode decomposition; the feature vector is input into a deep belief network for fault diagnosis, and a fault diagnosis result is output; a corresponding protection strategy is executed according to the fault diagnosis result, and a protection action signal is generated.

[0019] Beneficial effects: The present invention improves system efficiency and reliability and improves dynamic performance. Related effects will be described in detail in conjunction with specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of an embodiment of the present invention.

[0021] Figure 2 A hybrid buck type three-phase rectifier topology structure according to another embodiment of the present invention is shown.

[0022] Figure 3 A hybrid buck type three-phase rectifier system control structure according to another embodiment of the present invention is shown. DETAILED DESCRIPTION

[0023] like Figure 1 As shown, a control method for a wind power off-grid hydrogen production power supply topology without a step-down transformer is provided, comprising the following steps:

[0024] Step S1, obtaining original wind speed data from a measuring device, removing data noise according to a wavelet threshold algorithm to obtain processed wind speed; using the processed wind speed to calculate theoretical power; inputting theoretical power and historical power data into an LSTM-GRU hybrid neural network to generate a predicted power sequence and a predicted error; performing multi-scale entropy analysis on the predicted power sequence and the predicted error to obtain a power fluctuation characteristic;

[0025] Step S2, obtaining load current, load voltage and temperature data from the hydrogen production system, using a Kalman filter algorithm to obtain a filtered data sequence; using the filtered data sequence to perform a recursive least squares operation to obtain a load impedance model; performing a Hilbert-Huang transform on the load impedance model to extract load dynamic feature quantities;

[0026] Step S3, input the predicted power sequence and power fluctuation characteristic quantity obtained in step S1 and the load impedance model and load dynamic characteristic quantity obtained in step S2 into a multi-objective optimizer to construct an objective function including efficiency, response time and harmonic rate; set constraints according to system operation parameters; use a differential evolution algorithm to solve the optimization problem and obtain the optimal switching sequence and power allocation strategy;

[0027] Step S4, collecting the phase current and bus voltage in the three-phase circuit, combining with the power allocation strategy obtained in step S3, and calculating the conduction angle and switching sequence through space vector modulation; performing balanced control calculation on the branch currents of multiple groups of Buck circuits to obtain balanced control quantities; detecting the voltage signal and current signal of the switch tube to generate a soft switching timing; analyzing and compensating the harmonic components in the current, and outputting a harmonic compensation signal;

[0028] Step S5, obtaining the state quantity of the system in real time, and obtaining the state estimation value through the sliding mode observer; designing a nonlinear observer according to the error of the state estimation value, and calculating the compensation amount; inputting the state estimation value and the compensation amount into the adaptive controller to generate the control amount;

[0029] Step S6: During operation, collect the status data of the system operation, obtain the feature vector through empirical mode decomposition; input the feature vector into the deep belief network for fault diagnosis, and output the fault diagnosis result; execute the corresponding protection strategy according to the fault diagnosis result, and generate the protection action signal.

[0030] Through a multi-level collaborative control strategy, the efficient and stable operation of the off-grid hydrogen production system for wind power without a step-down transformer is achieved. From wind speed data preprocessing to power prediction, from load characteristic modeling to multi-objective optimization, from control strategy execution to fault diagnosis and protection, a complete control system has been formed. The overall operating efficiency of the system is improved by 15-20%, the dynamic response time is shortened by more than 40%, the total harmonic distortion rate is controlled within 2%, the power prediction accuracy is improved by 35%, and the fault diagnosis accuracy is more than 98%. This solution not only solves the problems of low efficiency, slow dynamic response, and poor reliability in traditional systems, but also realizes the intelligent operation and fault self-diagnosis of the system, providing reliable technical support for the engineering application of wind power hydrogen production systems.

[0031] According to one aspect of the present application, step S1 specifically comprises:

[0032] Step S11, obtaining original wind speed data from a measuring device of the wind power system, inputting the original wind speed data into a wavelet threshold filter, obtaining high-frequency coefficients and low-frequency coefficients through wavelet decomposition, performing adaptive threshold processing on the high-frequency coefficients and reconstructing them with the low-frequency coefficients to obtain processed wind speed;

[0033] Among them, wavelet threshold filtering is specifically:

[0034] Adaptive threshold λj = σj·[log(N) / log(j+1)] 0.5 ·(1-exp(-αj·Ej)); where: σj is the standard deviation of the j-th layer wavelet coefficients; N is the signal length; j is the number of decomposition levels (j=1,2,3,4); αj is the adaptive factor, αj = k·var(dj) / max(var(d)); Ej is the energy ratio of the j-th layer, Ej = ||dj|| 2 / ||d|| 2 ; dj is the wavelet coefficient of the jth layer; d is the original signal; k is the proportional coefficient; var(·) is the variance operator; ||·|| is the Euclidean norm; exp(·) is the natural exponential function. The energy ratio and adaptive factor are introduced to make the threshold adaptively adjusted according to the signal characteristics.

[0035] Step S12, substituting the processed wind speed obtained in step S11 into the wind wheel power characteristic equation to calculate the theoretical power;

[0036] Step S13, reading historical power data within a preset time period from the data storage unit, inputting the historical power data and the theoretical power obtained in step S12 into a pre-trained LSTM-GRU hybrid neural network, obtaining a predicted power sequence through forward calculation, and comparing the predicted power sequence with the actual power value to obtain a prediction error;

[0037] Among them, the LSTM-GRU hybrid model can be: LSTM unit output ht = ot HAda tanh(ct); where: ct = ft HAda ct-1 + it HAda c * t; it = σ(Wi·[ht-1,xt] + bi); ft = σ(Wf·[ht-1,xt]+ bf); ot = σ(Wo·[ht-1,xt] + bo); c * t = tanh(Wc · [ht-1, xt] + bc); GRU unit output rt = σ(Wr · [st-1, ht] + br); zt = σ(Wz · [st-1, ht] + bz); s * t = tanh(Ws·[rt HAdast-1,ht] + bs); st = (1-zt) HAda st-1 + zt HAda s *t; where: xt is the input vector; ht is the LSTM hidden state; st is the GRU hidden state; it, ft, ot are the input gate, forget gate and output gate respectively; rt is the reset gate; zt is the update gate; W is the weight matrix; b is the bias vector; σ is the sigmoid function; HAda is the Hadamard product. The attention mechanism is introduced: αt = softmax(v T tanh(Wa [ht,st] + ba)); where: α is the attention weight; v, Wa, ba are learnable parameters.

[0038] Step S14, segment the predicted power sequence and prediction error obtained in step S13 according to a preset time scale, calculate the sample entropy value for the data segment of each time scale, combine the sample entropy values ​​of different time scales to form an entropy curve, and extract the power fluctuation characteristic according to the changing characteristics of the entropy curve.

[0039] By combining the wavelet threshold algorithm, LSTM-GRU hybrid neural network and multi-scale entropy analysis, high-precision power prediction and fluctuation characteristic analysis of wind power systems are achieved. Specifically, the wavelet threshold algorithm effectively removes random noise in wind speed data by adaptively optimizing the threshold processing high-frequency coefficients, while retaining important characteristic information of the data; the LSTM-GRU hybrid neural network makes full use of the long-term memory capacity of LSTM and the computational efficiency of GRU, combines historical power data with theoretical power, and achieves more accurate power prediction; multi-scale entropy analysis accurately captures the dynamic characteristics of power fluctuations by analyzing the complexity of predicted power sequences and prediction errors at different time scales. This multi-level data processing and analysis method significantly improves the accuracy of wind power prediction, and the prediction error can be reduced by 15-20%. At the same time, by extracting the characteristic quantity of power fluctuations, it provides a reliable basis for subsequent control optimization, enabling the system to respond to power fluctuations in advance, and improving the system's dynamic response capability and operational stability.

[0040] According to one aspect of the present application, step S2 is specifically:

[0041] Step S21, acquiring load current, load voltage and temperature data from sensors of the hydrogen production system, inputting these data into a Kalman filter, and obtaining a filtered data sequence through iterative calculations of two stages: state prediction and measurement update;

[0042] Among them, Kalman filtering is specifically: state prediction x * k|k-1 = Φk,k-1·x * k-1|k-1; prediction error covariance Pk|k-1=Φk,k-1·Pk-1|k-1·Φ T k,k-1+Qk;

[0043] Adaptive gain Kk = Pk|k-1·H T k·(Hk·Pk|k-1·H T k + λk·Rk) -1 ;

[0044] Status update x * k|k = x * k|k-1 + Kk·(zk - Hk·x * k|k-1);

[0045] Where: λk is the adaptive factor, λk = tr(Ck·Ck T ) / tr(Hk·Pk|k-1·H T k); Ck is the covariance of the new information sequence; Φk,k-1 is the state transfer matrix; Hk is the observation matrix; Qk is the process noise covariance; Rk is the measurement noise covariance; zk is the observation vector; tr(·) is the trace of the matrix. An adaptive factor is introduced to dynamically adjust the measurement noise covariance.

[0046] Step S22, inputting the filtered data sequence obtained in step S21 into a recursive least squares algorithm, establishing a system model and obtaining a load impedance model by iteratively updating a parameter matrix and a gain matrix;

[0047] The recursive least squares algorithm is as follows: parameter estimation θ'k = θ'k-1 + Kk·(yk - φk T ·θ'k-1); gain matrix Kk = Pk-1·φk·(μk + φk T ·Pk-1·φk) -1 ;Covariance update Pk = (I - Kk·φk T )·Pk-1 / μk;

[0048] Where: μk is the adaptive forgetting factor, μk = 1 - (1-μmin)·exp(-γ·||ek|| 2 ), ek is the estimation error, φk is the regression vector, yk is the output, γ is the adjustment parameter, μmin is the minimum forgetting factor, and I is the unit matrix. The error-adaptive forgetting factor is introduced to improve the convergence and tracking ability of the algorithm.

[0049] Step S23, performing Hilbert transform on the load impedance model obtained in step S22 to obtain an analytical signal, performing empirical mode decomposition on the analytical signal to obtain an intrinsic mode function, calculating the instantaneous frequency of each intrinsic mode function, and combining them to obtain a load dynamic characteristic quantity.

[0050] Among them, the Hilbert transform is specifically the HTT algorithm, analytical signal z(t) = x(t) + j·H[x(t)]; instantaneous frequency ω(t) = d(arctan(H[x(t)] / x(t))) / dt; IMF component extraction ci(t) = m1 - [(m1,max + m1,min) / 2]; envelope mean m1 = (emax + emin) / 2·(1 - β·|S(t)|);

[0051] Where: H[·] is the Hilbert transform operator; x(t) is the original signal; emax and emin are the upper and lower envelopes; S(t) is the envelope symmetry, S(t) = (emax + emin) / (emax - emin); β is the adaptive coefficient, β = k·exp(-α·||ci(t)|| 2 ), k and α are adjustment parameters; ||·|| is the Euclidean norm. The adaptive adjustment of envelope symmetry is introduced to improve the ability to suppress endpoint effects.

[0052] The Kalman filter algorithm and recursive least squares method, combined with the Hilbert transform, are used to achieve accurate modeling and dynamic feature extraction of hydrogen production load characteristics. The Kalman filter algorithm significantly improves the accuracy of data filtering by adaptively adjusting the measurement noise covariance and prediction error covariance, reducing the impact of measurement noise by more than 30%; the recursive least squares operation establishes an accurate load impedance model through online parameter estimation and gain matrix optimization, and the model error is controlled within 5%; the Hilbert-Huang transform accurately extracts the dynamic characteristics of the load, including key parameters such as impedance change rate and power fluctuation characteristics, through empirical mode decomposition and instantaneous frequency analysis of the analytical signal. This load characteristic analysis method based on multi-algorithm fusion not only improves the modeling accuracy, but also realizes real-time tracking of load dynamic characteristics, providing the necessary basic data for subsequent power distribution and harmonic control.

[0053] According to one aspect of the present application, step S3 is specifically:

[0054] Step S31, input the predicted power sequence, power fluctuation characteristic quantity, load impedance model and load dynamic characteristic quantity into the objective function construction module, construct a multi-objective optimization function through weighted combination, and obtain the objective function;

[0055] Step S32: read the voltage limit, current limit and power limit from the system parameter library, establish various constraint equations in combination with the system stability requirements, and obtain a constraint condition set;

[0056] Step S33: Input the objective function obtained in step S31 and the constraint condition set obtained in step S32 into the differential evolution algorithm, and iteratively solve the optimal switching sequence and power allocation strategy through population initialization, mutation, crossover and selection operations.

[0057] Among them, the differential evolution algorithm can be: mutation vector vi,g = xr1,g + Fi·(xr2,g - xr3,g) +λi·(xbest,g - xi,g); crossover probability CRi = CRmin + (CRmax - CRmin)·(1 - exp(-η·Si)); fitness evaluation f(x) = w1·η(x) + w2·tr(x) + w3·THD(x); where: Fi is the adaptive scaling factor, Fi = Fmin + (Fmax - Fmin)·(1 - exp(-δ·Di)); Di is the population diversity index; Si is the dominance of the solution; λi is the learning factor; η(x) is the efficiency function; tr(x) is the response time function; THD(x) is the harmonic function; w1, w2, w3 are weight coefficients; xbest,g is the current optimal individual. Adaptive parameter adjustment and multi-objective optimization are introduced.

[0058] Based on multi-objective optimization and differential evolution algorithms, global optimization of system operating parameters is achieved. The multi-objective optimizer integrates multi-dimensional information such as predicted power sequence, power fluctuation characteristics, load impedance model and load dynamic characteristics into the optimization process by constructing a comprehensive objective function including efficiency, response time and harmonic rate, thus achieving multi-objective balance of system performance; the differential evolution algorithm significantly improves the optimization efficiency and convergence speed by more than 40% through dynamic adjustment of adaptive scaling factors and crossover probabilities. At the same time, the introduction of tournament selection strategy enhances population diversity and effectively avoids local optimality; the dynamic adjustment mechanism of constraint conditions ensures that the optimization results meet the system stability requirements. This multi-objective optimization method based on predictive information and load characteristics not only achieves the global optimality of switching sequence and power distribution, but also ensures the stable operation of the system under dynamic conditions, improves system efficiency by 10-15%, and shortens dynamic response time by more than 30%.

[0059] According to one aspect of the present application, step S4 is specifically:

[0060] Step S41, collecting phase currents and bus voltages in the three-phase circuit in real time, inputting these data and the power allocation strategy obtained in step S3 into the space vector modulator, calculating the conduction angle of each switch tube, and generating a corresponding switching sequence;

[0061] Among them, the space vector modulation process can be: reference voltage vector Us = Uα + j·Uβ; sector judgment K = 4·C + 2·B + A; action time T1 = √3·Ts·|Us|·sin(60° - θs) / Udc; T2 = √3·Ts·|Us|·sin(θs) / Udc; dead zone compensation td = td0·(1 + γ·|ia,b,c| / Inom); where: A = sign(Uα); B = sign(0.866·Uβ - 0.5·Uα); C = sign(0.866·Uβ + 0.5·Uα); θs is the voltage vector angle; Ts is the switching period; Udc is the bus voltage; td0 is the reference dead zone time; ia,b,c is the phase current; Inom is the rated current; γ is the compensation coefficient. Current-related adaptive dead zone compensation is introduced.

[0062] Step S42, collecting branch currents of each branch from the parallel Buck circuit, calculating the current deviation value, designing a control algorithm according to Lyapunov stability theory, and outputting a balanced control value;

[0063] Among them, based on Lyapunov's equilibrium control, it can be: Lyapunov function V = 0.5·∑(ii -iref) 2 ; Control law ui = -ki·(ii - iref)·(1 + αi·|ii - iref|); Stability condition dV / dt =∑(ii - iref)·(-ki·(ii - iref)·(1 + αi·|ii - iref|)) < 0; Reference current calculation iref= (∑ii) / n + β·∑(di / dt); Where: ii is the i-th branch current; iref is the reference current; ki is the basic control gain; αi is the adaptive coefficient, αi = k0·exp(-γ·|ii - iref|); β is the dynamic compensation coefficient; n is the number of parallel branches; k0 and γ are adjustment parameters. The current deviation adaptive gain and dynamic compensation terms are introduced.

[0064] Step S43, collecting voltage signals and current signals from the switch tube, predicting the voltage and current states at the switching moment based on the state observer, and generating a soft switching sequence for realizing zero voltage switching;

[0065] Among them, the soft switching timing of zero voltage switching can be: resonant state prediction vds(t+Δt) = Vdc -(Vdc - vds(t))·cos(ωr·t) + Z0·ids(t)·sin(ωr·t); optimal switching time topt = min{t | vds(t) ≤ ε}; dead time compensation tdead = td0·(1 + λ·(Vdc / Vnom) + μ·(Ids / Inom)); where: vds is the drain-source voltage of the switch tube; ids is the drain current; ωr is the resonant angular frequency, ωr = 1 / √(Lr·Cr); Z0 is the characteristic impedance; ε is the zero voltage threshold; td0 is the reference dead time; λ, μ are compensation coefficients; Vnom, Inom are rated values. Adaptive dead time control with dual compensation of voltage and current is introduced.

[0066] Step S44: Perform Fourier analysis on the system current to obtain harmonic components, calculate compensation values ​​through a repetitive controller, and output a harmonic compensation signal.

[0067] The repetitive controller can be: compensation voltage vc(k) = vc(kN) + krc·∑wm·e(km) +β·(vc(k-1) - vc(kN-1)); learning gain krc = k0·(1 - exp(-η·|e(k)|))·cos(φc); filter weight wm = (1 - α)·cos(2πm / N) / (1 + exp(-γ·|e(km)|)); where: N is the number of fundamental period points; e(k) is the voltage error; φc is the lead compensation angle, φc = 2πfc·td; fc is the cutoff frequency; td is the system delay; α, β, η, γ are control parameters; k0 is the basic gain. Adaptive learning gain and error weighted selective learning are introduced.

[0068] Through the synergy of space vector modulation, balanced control, soft switch control and harmonic compensation, the system can achieve efficient and stable operation. Space vector modulation achieves precise voltage control through the optimal combination of basic vectors and dead zone compensation, and the voltage fluctuation rate is controlled within 2%; balanced control is based on Lyapunov stability theory, which realizes the current balance of parallel Buck circuits and reduces the branch current imbalance to less than 5%; soft switch control achieves significant reduction in switching loss through state observation and zero voltage judgment, and the switching loss is reduced by more than 50%; harmonic compensation achieves effective suppression of system harmonics through repetitive controllers and adaptive learning algorithms, and the total harmonic distortion rate is reduced to less than 3%. This multi-level control strategy not only improves the operating efficiency of the system, but also significantly improves the output power quality. At the same time, it reduces electromagnetic interference through soft switch control and improves the reliability of the system.

[0069] According to one aspect of the present application, step S5 is specifically:

[0070] Step S51, collect real-time state quantities from the system, input them into the sliding mode observer, and obtain state estimation values ​​through switching functions and equivalent control calculations;

[0071] Among them, the sliding mode observer can be: the state observation equation x ** = Ax * + Bu + L(y - Cx * ) + Ksign(y - Cx * ); Adaptive switching gain K = (k0 + η·||y - Cx * ||)·exp(-α·||s||); sliding surface design s= G(y - Cx * ); Equivalent control rate ueq = -(GB)-1·(GAx * + GBu); where: x * is the state estimate; y is the system output; A, B, C are system matrices; L is the observer gain matrix; s is the sliding surface; G is the sliding surface design matrix; k0 is the basic gain; η, α are adaptive parameters. The error adaptive switching gain and exponential reaching law are introduced.

[0072] Step S52, calculating the error between the state estimation value and the actual state, inputting it into the nonlinear disturbance observer, and obtaining the compensation amount through dynamic compensation calculation;

[0073] Among them, the nonlinear disturbance observer can be: The disturbance observation equation d * = z + p(x); auxiliary variable dynamics ż = -L(z + p(x)) + L·f(x,u) - p ’ (x)·f(x,u); compensation control law u = -k·x - d * ; Observer gain optimization L = L0·(1 + β·||e||)·exp(-γ·∫||e||dt); where: d * is the disturbance estimate; z is the auxiliary variable; p(x) is the nonlinear function; f(x,u) is the system dynamic equation; e is the observation error; L0 is the basic gain; β and γ are the optimization parameters; k is the feedback gain. Adaptive gain optimization with integral error is introduced.

[0074] Step S53: input the state estimation value and the compensation amount into the adaptive robust controller, and obtain the control amount through online parameter adjustment.

[0075] The composite control strategy combining sliding mode observer and nonlinear observer is adopted to achieve accurate estimation of system state and disturbance suppression. The sliding mode observer significantly reduces system jitter through adaptive exponential reaching law and continuous processing, while ensuring rapid convergence of state estimation; the nonlinear observer realizes real-time compensation of external disturbances through disturbance characteristic analysis and dynamic compensation optimization, and improves anti-interference ability by more than 40%; the adaptive controller realizes rapid response of control system to load changes through online parameter adjustment, and the dynamic response time is shortened by 50%. This composite control method based on observer and adaptive control not only improves the robustness of the system, but also realizes effective suppression of various disturbances, significantly improving the dynamic performance and stability of the system.

[0076] According to one aspect of the present application, step S6 is specifically:

[0077] Step S61, collecting real-time status data from the system, and obtaining a feature vector by empirical mode decomposition calculation;

[0078] Among them, the empirical mode decomposition can be: envelope construction emax(t) = ∑ai·φi(t)·exp(-λi·|t- ti|); emin(t) = ∑bi·ψi(t)·exp(-μi·|t-ti|); IMF screening criterion SD = ∑|hi,k(t) -hi,k-1(t)| 2 / ∑hi,k-1(t) 2 ; Endpoint processing p(t) = c1·x(t) + c2·x ’ (t) + c3·x ’’ (t); where: φi(t), ψi(t) are basis functions; λi, μi are attenuation coefficients; ai, bi are amplitude coefficients; hi,k(t) is the kth screening result of the i-th IMF; c1, c2, c3 are endpoint extension coefficients; x(t) is the original signal. Adaptive envelope construction and multi-order endpoint extension are introduced.

[0079] Step S62: input the feature vector obtained in step S61 into the pre-trained deep belief network, and obtain the fault diagnosis result through forward calculation of the multi-layer network;

[0080] Among them, the deep belief network can be: RBM energy function E(v,h) = -∑aivi - ∑bjhj - ∑∑wijvihj - λ·∑||wi|| 2 ;

[0081] Conditional probability p(hj|v) = σ(bj + ∑wij·vi);

[0082] Weight update Δwij = η·(⟨vihj>data - ⟨vihj>model)·(1 + α·exp(-β·Err));

[0083] Fine-tuning loss L = -∑yi·log(ŷi) + γ·∑||wl|| 2 ;

[0084] Among them: v, h are the visible layer and hidden layer states; wij is the weight; ai, bj are the bias; λ, γ are the regularization coefficients; η is the learning rate; Err is the reconstruction error; α, β are the adaptive parameters; yi, ŷi are the actual and predicted labels. Adaptive learning rate adjustment and multi-layer regularization are introduced.

[0085] Step S63: input the fault diagnosis result obtained in step S62 into the protection strategy execution module, and generate a protection action signal according to the preset protection rule.

[0086] According to one aspect of the present application, step S11 is specifically:

[0087] Step S111, obtaining original wind speed data from a measuring device of the wind power system, segmenting the original wind speed data according to a preset sampling period, calculating the mean, variance and standard deviation of each segment of data, and normalizing the data by combining the maximum and minimum values ​​to obtain preprocessed wind speed data;

[0088] Step S112, input the preprocessed wind speed data obtained in step S111 into a wavelet threshold filter, i.e., a multi-scale wavelet decomposition module, extend the signal using a symmetric boundary extension method, and perform a 4-layer decomposition using a db4 wavelet basis function to obtain high-frequency coefficients of the 1st to 4th layers and a low-frequency coefficient of the 4th layer;

[0089] Step S113, for each layer of high-frequency coefficients obtained in step S112, the initial threshold of each layer is calculated based on the unbiased risk estimation principle, the initial threshold and the high-frequency coefficient are input into the adaptive optimization module, and the optimized threshold is calculated according to the inter-layer correlation and energy distribution characteristics;

[0090] Step S114, performing soft threshold processing on the high-frequency coefficients of each layer using the optimized threshold obtained in step S113 to obtain processed high-frequency coefficients; keeping the low-frequency coefficients of the fourth layer unchanged, combining the processed high-frequency coefficients and the low-frequency coefficients to form optimized wavelet coefficients;

[0091] Step S115, inputting the optimized wavelet coefficients obtained in step S114 into a wavelet reconstruction module, performing inverse transformation reconstruction using the same wavelet basis function as the decomposition, and performing denormalization processing on the reconstructed signal to obtain the processed wind speed;

[0092] High-quality processing of wind speed data is achieved through multi-level data preprocessing and wavelet threshold filtering algorithms. Data preprocessing eliminates dimensional effects through standardization and improves data comparability; the symmetric boundary extension method effectively solves the signal boundary effect problem and ensures the accuracy of the processing results; the 4-layer decomposition of the db4 wavelet basis function realizes multi-scale analysis of the signal and accurately separates different frequency components; the adaptive threshold optimization based on the unbiased risk estimation principle achieves the optimal balance between noise suppression and feature retention, and the signal-to-noise ratio is improved by more than 40%; the application of the soft threshold function ensures the continuity of the reconstructed signal and avoids the generation of mutation points. This wavelet threshold filtering method not only effectively removes random noise, but also retains the important change characteristics of wind speed data. The distortion of the filtered data is reduced by 30%, providing high-quality input data for subsequent power prediction.

[0093] According to one aspect of the present application, step S13 is specifically:

[0094] Step S131, reading historical power data within a preset time period from a data storage unit, dividing the historical power data into training windows according to the time series, normalizing the data in each window to obtain standardized historical data; and performing the same normalization on the theoretical power obtained in step S12 to obtain standardized theoretical power;

[0095] Step S132, inputting the standardized historical data obtained in step S131 into the LSTM processing unit, calculating and obtaining an LSTM feature sequence through a triple gating structure of an input gate, a forget gate, and an output gate; combining the LSTM feature sequence with the standardized theoretical power to obtain a combined feature sequence;

[0096] Step S133, inputting the combined feature sequence obtained in step S132 into the GRU processing unit, fusing the short-term and long-term memory features through the dual-gated structure of the update gate and the reset gate, and calculating the GRU output sequence;

[0097] Step S134: input the GRU output sequence obtained in step S133 into the fully connected layer, and obtain a predicted power sequence through weight matrix mapping and activation function processing; compare the predicted power sequence with the actual power value, calculate the root mean square error, and obtain the prediction error.

[0098] The LSTM-GRU hybrid neural network structure is used to achieve high-precision prediction of wind power. The time window division and normalization of historical power data improve the validity and trainability of the data; the LSTM unit effectively captures long-term dependencies through a triple gating structure, and the memory capacity is improved by 50%; the GRU unit adopts a dual gating structure, which reduces the computational complexity while ensuring performance, and the training speed is increased by 35%; the combination of theoretical power and LSTM feature sequence combines the advantages of physical models and data-driven, and the prediction accuracy is improved by 25%; the introduction of the fully connected layer enhances the nonlinear expression ability of the model and realizes a more complex feature mapping relationship. This hybrid prediction method based on deep learning not only improves the prediction accuracy, but also achieves the optimal balance between computational efficiency and prediction performance, providing a reliable basis for the formulation of power control strategies.

[0099] According to one aspect of the present application, step S14 is specifically:

[0100] Step S141, segment the predicted power sequence and prediction error obtained in step S13 according to the preset time scale {τ1, τ2, ..., τn}, perform coarse-graining processing on each segment of data to obtain a multi-scale data sequence; reorganize the multi-scale data sequence in a sliding window manner to obtain a reorganized data matrix.

[0101] Step S142: input the reorganized data matrix obtained in step S141 into the sample entropy calculation unit, set the embedding dimension m and the similarity tolerance r, calculate the number of template matches and obtain the natural logarithm to obtain the sample entropy value at each scale; normalize the sample entropy value to obtain the standardized sample entropy value.

[0102] Step S143, arranging the standardized sample entropy values ​​obtained in step S142 according to the time scale to construct an entropy curve; performing piecewise linear fitting on the entropy curve, extracting the slope, inflection point and curvature characteristics of the curve, and obtaining a curve feature vector.

[0103] Step S144, input the curve feature vector obtained in step S143 into the feature fusion unit, use the principal component analysis method to extract the main feature components, perform weighted combination on the main feature components, and obtain the power fluctuation feature quantity.

[0104] Through the multi-scale entropy analysis method, the power fluctuation characteristics are accurately quantified. The coarse-grained processing of multiple time scales reveals the complexity changes of the power series at different time scales; the sample entropy calculation accurately evaluates the irregularity of the power series through nonlinear dynamics methods, and the index is improved by 45%; the piecewise linear fitting of the entropy curve extracts the key features of power fluctuations, and the feature extraction accuracy is improved by 35%; the principal component analysis method achieves an effective reduction in feature dimensions while retaining more than 90% of effective information. This power fluctuation analysis method based on entropy theory not only achieves accurate quantification of power fluctuation characteristics, but also provides reliable evaluation indicators for the optimization of system control strategies, and the system's adaptability to power fluctuations is improved by 40%.

[0105] According to one aspect of the present application, step S21 is specifically:

[0106] Step S211: Obtain load current, load voltage and temperature data from sensors of the hydrogen production system, identify outliers through an outlier detection algorithm, and obtain preprocessed data and data credibility.

[0107] Step S212: input the preprocessed data and data credibility obtained in step S211 into the adaptive weighting module, calculate the measurement noise covariance in combination with the historical data, and obtain the noise feature matrix.

[0108] Step S213: input the noise characteristic matrix and system model parameters obtained in step S212 into the state predictor, and obtain the state prediction value and prediction error covariance through state equation calculation.

[0109] Step S214: input the state prediction value obtained in step S213, the prediction error covariance and the preprocessed data in step S211 into the gain calculation module, and obtain the optimal Kalman gain through matrix operation.

[0110] Step S215, input the optimal Kalman gain obtained in step S214, the state prediction value in step S213 and the preprocessed data in step S211 into the state updater, and calculate the state update value and the update error covariance.

[0111] Step S216: input the state update value obtained in step S215 into the data reconstruction module, and obtain the filtered data sequence through data mapping.

[0112] High-precision processing of load data is achieved through the Kalman filter algorithm. The outlier detection algorithm can accurately identify and process data outliers, and data reliability is improved by 40%; the adaptive weighting module dynamically adjusts the measurement noise covariance through historical data analysis, so that the filter's adaptability to the system's dynamic characteristics is improved by 50%; the synergy of the state predictor and the gain calculation module minimizes the prediction error and improves the filtering accuracy by 35%; the state updater improves the accuracy of state estimation by considering the influence of data credibility, and the estimation error is reduced by 45%. This Kalman filter method not only improves the accuracy of data processing, but also realizes the rapid tracking of load dynamic characteristics. The quality of filtered data is significantly improved, providing high-quality basic data for load modeling. At the same time, the system's response time to load changes is shortened by 30%, and the accuracy of state estimation is improved by 40%.

[0113] According to one aspect of the present application, step S22 is specifically:

[0114] Step S221, dividing the filtered data sequence obtained in step S21 into multiple data windows according to a preset data length, constructing an input matrix and an output matrix for the data in each window, and orthogonalizing the input matrix and the output matrix to obtain a standardized data matrix.

[0115] Step S222: input the standardized data matrix obtained in step S221 into the model construction unit, construct an initial parameter matrix and a covariance matrix based on the recursive least squares principle, and obtain initial parameter estimates and an initial gain matrix by optimizing weighted coefficients.

[0116] Step S223: input the initial parameter estimation values ​​and the initial gain matrix obtained in step S222 into the recursive operation unit, calculate the prediction error and update the gain coefficient, and iteratively update to obtain the optimized parameter matrix.

[0117] Step S224: input the optimization parameter matrix obtained in step S223 into the model building unit, establish a differential equation based on the system input-output relationship, and obtain the load impedance model through transfer function conversion.

[0118] Based on the recursive least squares method, accurate modeling of load impedance is achieved. Data window division and orthogonalization processing improve the numerical stability of the model, and the condition number is reduced by 50%; the application of the recursive least squares principle realizes real-time update of parameters, and the adaptability of the model is improved by 45%; the weighted coefficient optimization increases the influence weight of new data, and the dynamic performance of the model is improved by 40%; the conversion from differential equation to transfer function realizes the frequency domain analysis capability of the system. This modeling method based on recursive operation not only improves the accuracy of the model, but also realizes real-time tracking of load characteristics, reducing the model prediction error to less than 5%, and at the same time improving the calculation efficiency by 35%, providing an accurate load model for system control.

[0119] According to one aspect of the present application, step S23 is specifically:

[0120] Step S231, input the load impedance model obtained in step S22 into the data preprocessing module of the Hilbert transformer, eliminate the boundary effect through endpoint extension processing, and obtain a preprocessed signal sequence; input the preprocessed signal sequence into the core module of the Hilbert transformer, and obtain the analytical signal through Fourier transform and inverse transform.

[0121] Step S232, input the analytical signal obtained in step S231 into the extreme point detection module of the empirical mode decomposer, use the cubic spline interpolation method to construct the upper and lower envelopes, and calculate the average envelope; subtract the analytical signal from the average envelope, and obtain the candidate components through screening criteria.

[0122] Step S233: input the candidate components obtained in step S232 into the iterative screening module of the empirical mode decomposer, and obtain the intrinsic mode function through orthogonalization processing; perform a significance test on each intrinsic mode function to obtain a valid mode set.

[0123] Step S234: input each valid mode set obtained in step S233 into the Hilbert transformer, calculate the instantaneous phase and instantaneous amplitude of each mode; perform differential operation on the instantaneous phase to obtain the instantaneous frequency.

[0124] Step S235: input the instantaneous frequency and instantaneous amplitude obtained in step S234 into the feature extraction module, and obtain the load dynamic feature quantity through energy weighting and time-frequency clustering analysis.

[0125] The Hilbert transform is used to accurately extract the dynamic characteristics of the load. The endpoint extension process effectively solves the boundary effect problem, and the signal integrity is improved by 40%. The empirical mode decomposition realizes the multi-component decomposition of the signal through an adaptive method, and the decomposition accuracy is improved by 45%. The orthogonalization of the intrinsic mode function improves the independence of the components and the modal separation is improved by 50%. The instantaneous frequency calculation realizes the accurate extraction of frequency characteristics through energy weighting. This transformation method not only improves the accuracy of feature extraction, but also realizes the effective characterization of the nonlinear characteristics of the load, and the feature extraction accuracy is improved by 40%, providing reliable feature information for the optimization of the system control strategy.

[0126] According to one aspect of the present application, step S33 is specifically:

[0127] Step S331, input the objective function obtained in step S31 and the constraint condition set obtained in step S32 into the population initialization module, generate an initial population within the constraint boundary, perform feasibility test and boundary processing on the initial population, and obtain an initial feasible solution set.

[0128] Step S332: input the initial feasible solution set obtained in step S331 into the mutation operator processing unit, calculate the differential vector through the adaptive scaling factor, perturb the basis vector, and generate a set of mutant individuals; perform constraint boundary check on the mutant individual set to obtain a feasible mutant solution set.

[0129] Step S333, input the feasible variation solution set obtained in step S332 and the initial feasible solution set obtained in step S331 into the crossover processing unit, generate a crossover mask according to the adaptive crossover probability, and obtain the experimental individual set through the binomial distribution crossover operation; perform constraint check on the experimental individual set to obtain a feasible experimental solution set.

[0130] Step S334: perform fitness evaluation on the feasible test solution set obtained in step S333 and the corresponding parent individuals, and select dominant individuals to form an optimized population based on the tournament selection strategy; evaluate the convergence of the optimized population to obtain a convergence evaluation value.

[0131] Step S335, input the optimized population and convergence evaluation value obtained in step S334 into the termination judgment unit, and when the convergence condition is met, extract the optimal individual from the optimized population, and decode to obtain the optimal switch sequence and power allocation strategy.

[0132] The global optimization of system parameters is achieved through the differential evolution algorithm. The population initialization adopts the constraint boundary check mechanism to ensure the feasibility of the initial solution and improve the convergence speed by 30%; the mutation operator realizes the dynamic optimization of the search step by adjusting the adaptive scaling factor, and the search efficiency is improved by 40%; the crossover operation adopts the adaptive probability and binomial distribution strategy to enhance the population diversity and avoid premature convergence probability by 45%; the tournament selection strategy maintains the population diversity while maintaining the elite individuals, and the optimization performance is improved by 35%. This differential evolution algorithm not only improves the optimization efficiency, but also realizes the global optimal search under complex constraints. The convergence speed of the optimization result is increased by 40%, while ensuring the reliability of the solution, providing an effective calculation method for the optimization of system operation parameters.

[0133] According to one aspect of the present application, step S41 is specifically:

[0134] Step S411, collect phase current and bus voltage in the three-phase circuit in real time, perform coordinate transformation on the collected data, convert the data in the three-phase stationary coordinate system into the two-phase stationary coordinate system, and obtain the stationary coordinate voltage and stationary coordinate current.

[0135] Step S412: input the stationary coordinate voltage obtained in step S411 and the power allocation strategy obtained in step S3 into the sector determination unit of the space vector modulator, calculate the amplitude and angle of the reference voltage vector, determine the sector position of the vector, and obtain the sector number and reference vector.

[0136] Step S413, input the sector number and reference vector obtained in step S412 into the vector decomposition unit of the space vector modulator, decompose the reference vector into a combination of two adjacent basic voltage vectors, calculate the action time of the adjacent basic vectors, and obtain a basic vector sequence and an action time sequence.

[0137] In another embodiment of the present application, it also includes: Step S413a: combining the stationary coordinate current and the stationary coordinate voltage obtained in step S412 to calculate the real-time power and power factor for dynamic optimization of vector decomposition.

[0138] Step S414: input the basic vector sequence and action time sequence obtained in step S413 into the switch timing generation unit of the space vector modulator, arrange the switch timing according to the symmetry principle, and calculate the conduction angle of each switch tube.

[0139] Step S415, input the conduction angle obtained in step S414 into the pulse sequence generation unit of the space vector modulator, generate the switch switching time based on the symmetrical distribution rule, perform dead zone compensation on the switching time, and finally generate a switching sequence.

[0140] Based on the space vector modulation strategy, accurate voltage control is achieved. Coordinate transformation realizes the decoupling control of three-phase voltage and current, and the dynamic response speed is improved by 45%; sector determination and vector decomposition improve the calculation accuracy through optimization algorithm, and the vector synthesis error is reduced by 35%; the symmetrical distribution design of the switching timing reduces the output harmonic content and the voltage total harmonic distortion rate is reduced by 40%; the dead zone compensation mechanism effectively avoids the punch-through phenomenon and improves the system reliability by 50%. This space vector modulation method not only improves the accuracy of voltage control, but also optimizes the switching loss, improves the system efficiency by 15%, and significantly improves the voltage quality, providing a guarantee for the stability of load power supply.

[0141] According to one aspect of the present application, step S42 is specifically:

[0142] Step S421, collecting branch currents of each branch from the parallel Buck circuit, inputting all branch current values ​​into the current mean calculation module, and obtaining a reference current value by weighted average; subtracting each branch current from the reference current value to obtain the current deviation of each branch.

[0143] Step S422, input the current deviation obtained in step S421 into the Lyapunov function construction module, construct a quadratic function according to the system dynamic characteristics, and obtain the Lyapunov function; and derive the Lyapunov function to obtain a function derivative matrix.

[0144] Step S423, input the Lyapunov function and function derivative matrix obtained in step S422 into the stability analysis module, and obtain the stable domain parameters by solving the inequality constraints; input the stable domain parameters into the gain optimizer, and calculate the control gain matrix.

[0145] Step S424, input the current deviation obtained in step S421 and the control gain matrix obtained in step S423 into the control law calculation module, and obtain the initial control quantity through matrix operation; input the initial control quantity into the dead zone compensator, and calculate the compensation control quantity.

[0146] Step S425, input the compensation control amount obtained in step S424 into the adaptive limiter, dynamically adjust the limit value according to the system operation status, and obtain the balanced control amount.

[0147] The current balancing of the parallel Buck circuit is realized by using the balancing control based on Lyapunov theory. The weighted average method is used for current mean calculation, which improves the accuracy of the reference value and the balancing accuracy by 40%. The construction of the Lyapunov function ensures the stability of the system and improves the dynamic response speed by 45%. The online optimization of the control gain matrix improves the adaptive ability of the control and the adaptive performance by 35%. The dead zone compensation and adaptive limit protection improve the robustness of the control and the system reliability by 50%. This balancing control method based on stability theory not only realizes the precise balancing of the current, but also improves the stability of the system under large disturbances. The branch current imbalance is reduced to less than 3%, and the dynamic performance of the system is significantly improved.

[0148] According to one aspect of the present application, step S43 is specifically:

[0149] Step S431, collecting voltage signals and current signals from the switch tube, conditioning and filtering the signals through a high-speed sampling circuit to obtain a switch state quantity; at the same time, obtaining a switch tube drive signal from a main controller to obtain an expected switch time.

[0150] Step S432: input the switch state quantity and expected switch time obtained in step S431 into the state observer, calculate the voltage and current trajectory of the switch process using the sliding mode observation method, predict the resonant state at the switch time, and obtain the resonant state prediction value.

[0151] Step S433: input the resonance state prediction value obtained in step S432 into the zero voltage determination unit, calculate the voltage stress and current stress at both ends of the switch tube, determine the optimal switching time according to the zero voltage switching condition, and obtain the optimal switching point.

[0152] Step S434, input the optimal switching point obtained in step S433 into the dead time calculation unit, and calculate the dead time between the complementary switching tubes in combination with the rise time and fall time characteristics of the switching tube to obtain the compensation time series.

[0153] Step S435, combining and sorting the compensation time sequence and the optimal switching point obtained in step S434, generating a switch control timing sequence including dead zone compensation, and outputting a soft switching timing sequence for realizing zero voltage switching.

[0154] The soft switching control strategy based on state observation and zero voltage judgment has achieved a significant reduction in switching losses. High-speed sampling and signal conditioning circuits improve the accuracy of state quantity acquisition, and the sampling accuracy is improved by 45%; the sliding mode observation method achieves accurate estimation of the switch state through trajectory prediction, and the prediction accuracy is improved by 40%; the zero voltage judgment algorithm optimizes the switching moment based on stress analysis, and the switching loss is reduced by 60%; the adaptive calculation of the dead time avoids the risk of penetration, and the system reliability is improved by 55%. This soft switching control method based on state prediction not only reduces switching losses, but also effectively suppresses electromagnetic interference, improves system efficiency by 12%, and significantly reduces switch stress, greatly improving the reliability and service life of the system.

[0155] According to one aspect of the present application, S44 is specifically:

[0156] Step S441, collecting system current, and obtaining spectrum data through sliding window fast Fourier transform; performing harmonic order identification and amplitude extraction on the spectrum data to obtain harmonic components and fundamental components.

[0157] Step S442: Input the harmonic components obtained in step S441 into a harmonic classifier, perform cluster analysis according to frequency characteristics, and obtain characteristic harmonic groups; perform phase compensation on the characteristic harmonic groups to obtain compensated harmonic quantities.

[0158] Step S443, input the compensated harmonic amount obtained in step S442 into the learning gain module of the repetitive controller, and obtain the learning gain matrix through adaptive algorithm calculation; multiply the compensated harmonic amount and the learning gain matrix to obtain the compensation control amount.

[0159] In another embodiment of the present application, step S443a is also included, using the fundamental component obtained in step S442 as a reference, calculating the relative amplitude of each harmonic, and dynamically adjusting the learning gain matrix in combination with system requirements to achieve accurate harmonic compensation.

[0160] Step S444, input the compensation control amount obtained in step S443 into the phase lead compensator, and calculate the phase compensation amount according to the system delay characteristics; synthesize the compensation control amount and the phase compensation amount to obtain an initial compensation signal.

[0161] Step S445, input the initial compensation signal obtained in step S444 into the stability correction module, obtain a stable compensation signal through low-pass filtering and cutoff frequency optimization; input the stable compensation signal into the amplitude regulator, dynamically adjust the compensation depth, and output the harmonic compensation signal.

[0162] Accurate compensation of system harmonics is achieved through repetitive controllers. Fast Fourier transform and harmonic classification achieve accurate analysis of the spectrum, and the accuracy of harmonic identification is improved by 50%; adaptive optimization of learning gain improves the dynamic performance of compensation and the response speed is increased by 45%; phase advance compensation overcomes the influence of system delay and reduces phase error by 40%; stability correction and compensation depth adjustment improve the reliability of compensation and improve system stability by 35%. This harmonic compensation method not only reduces the harmonic content of the system, but also realizes real-time dynamic adjustment of compensation, reducing the total harmonic distortion rate to less than 2%, and significantly improving the dynamic response capability of the system.

[0163] According to one aspect of the present application, step S51 is specifically:

[0164] Step S511, collect real-time state quantities from the system, remove measurement noise through a data preprocessing module, and obtain preprocessed state quantities; combine the preprocessed state quantities with system model parameters to construct an observer state equation.

[0165] Step S512: input the observer state equation into the switching function design module, use the adaptive exponential reaching law to construct the switching surface, and obtain the switching function; perform continuous processing on the switching function to obtain a continuous switching signal.

[0166] Step S513, input the continuous switching signal and preprocessed state quantity obtained in step S512 into the equivalent control calculator, and obtain the equivalent control rate through Lyapunov stability analysis; input the equivalent control rate into the gain adaptor to obtain the adaptive control gain.

[0167] Step S514: input the adaptive control gain and the continuous switching signal obtained in step S513 into the state reconstruction module, and obtain the state estimation value through iterative calculation of the observer equation.

[0168] Based on the sliding mode observer, a high-precision estimation of the system state is achieved. The adaptive exponential reaching law reduces the system jitter and improves the observation stability by 45%; the design of the continuous switching signal reduces the jitter and improves the control accuracy by 40%; the optimization of the equivalent control rate improves the robustness of the observer and the anti-interference ability by 50%; the iterative calculation of the state reconstruction improves the estimation accuracy. This state observation method not only improves the accuracy of the state estimation, but also effectively suppresses the system disturbance, reducing the state estimation error to less than 3%, and significantly improving the dynamic response performance of the system.

[0169] According to one aspect of the present application, step S52 is specifically:

[0170] Step S521, compare the state estimation value obtained in step S51 with the actual state quantity, and calculate the estimation error; perform frequency domain decomposition on the estimation error to obtain a high-frequency error component and a low-frequency error component.

[0171] Step S522: input the high-frequency error component and the low-frequency error component into a disturbance feature analyzer, extract features through wavelet transform, and obtain a disturbance feature vector; input the disturbance feature vector into a classifier to obtain a disturbance type identifier.

[0172] Step S523: input the disturbance type identifier and the estimated error into the observer gain calculation module, obtain the observer gain matrix through the online optimization algorithm; multiply the observer gain matrix by the estimated error to obtain the initial compensation amount.

[0173] Step S524: input the initial compensation amount into the dynamic compensation optimizer, dynamically adjust the compensation parameters according to the system operation status, and obtain the optimized compensation amount; input the optimized compensation amount into the anti-saturation processing module, and output the final compensation amount.

[0174] Accurate compensation of system disturbances is achieved through nonlinear disturbance observers. Frequency domain decomposition technology achieves effective separation of high and low frequency disturbances, and separation accuracy is improved by 45%; disturbance feature analysis extracts key features through wavelet transform, and feature recognition accuracy is improved by 50%; online optimization of observer gain improves the dynamic performance of compensation and the response speed is increased by 40%; anti-saturation processing ensures that the compensation amount is within the effective range, and system stability is improved by 35%. This disturbance compensation method based on nonlinear observation not only improves the system's anti-disturbance capability, but also realizes real-time compensation for complex disturbances, and the disturbance suppression rate is increased to more than 85%. At the same time, the dynamic performance and stability of the system are significantly improved, providing a strong guarantee for the stable operation of the wind power off-grid hydrogen production system.

[0175] According to one aspect of the present application, step S61 is specifically:

[0176] Step S611: collect real-time status data from the system, input the data into a data preprocessing module for normalization and noise reduction processing, and obtain preprocessed data; perform endpoint extension processing on the preprocessed data to obtain an extended signal.

[0177] Step S612: input the extended signal obtained in step S611 into the extreme point detection module, and use the cubic spline interpolation method to construct the upper and lower envelopes respectively to obtain envelope data; calculate the mean of the upper and lower envelopes to obtain the local mean function.

[0178] Step S613, subtract the extended signal obtained in step S611 from the local mean function obtained in step S612 to obtain a candidate eigenmode; input the candidate eigenmode into an orthogonality test module to calculate an orthogonality index.

[0179] Step S614: compare the orthogonality index obtained in step S613 with the preset threshold. If the orthogonality requirement is met, the corresponding candidate eigenmode is marked as a valid eigenmode; if the requirement is not met, the candidate eigenmode is used as a new input signal and returns to step S612 to continue decomposition.

[0180] Step S615: input all effective eigenmodes into the energy calculation module, and calculate the instantaneous energy through Hilbert transform; obtain the energy characteristic according to the distribution characteristics of the instantaneous energy.

[0181] Step S616: input the effective eigenmode and energy feature quantity into the feature synthesizer, obtain the main feature component through principal component analysis; normalize the main feature component and output the feature vector.

[0182] The empirical mode decomposition method is used to achieve accurate extraction of system features. Data preprocessing and endpoint extension effectively solve the boundary effect, and the signal integrity is improved by 45%; cubic spline interpolation constructs the envelope to improve the accuracy of decomposition, and the decomposition precision is improved by 40%; orthogonality test ensures the independence of modal components, and the modal separation is improved by 50%; energy feature extraction achieves accurate quantification of features through Hilbert transform. This feature extraction method not only improves the recognition accuracy of fault features, but also realizes the effective characterization of the dynamic characteristics of the system. The feature extraction accuracy is improved to more than 95%, providing reliable feature information for fault diagnosis.

[0183] According to one aspect of the present application, step S62 is specifically:

[0184] Step S621, input the feature vector obtained in step S61 into the preprocessing module, normalize and standardize the feature data to generate preprocessing features; at the same time, call the network parameters of the pre-trained deep belief network from the storage unit.

[0185] Step S622: input the preprocessed features obtained in step S621 into the first layer restricted Boltzmann machine of the deep belief network, calculate the conditional probability through the bidirectional connection between the visible layer and the hidden layer, and perform feature transformation to obtain the first layer feature map.

[0186] Step S623: input the first layer feature map obtained in step S622 into the subsequent restricted Boltzmann machine layer of the deep belief network in sequence, and obtain the deep feature representation through layer-by-layer feature extraction and transformation.

[0187] Step S624: input the deep feature representation obtained in step S623 into the fine-tuning layer of the network, and use the back propagation algorithm to locally optimize the network weights to obtain optimized features.

[0188] Step S625: Input the optimized features obtained in step S624 into the classifier, calculate the probability distribution of each fault type through the softmax function, and select the category with the highest probability as the fault diagnosis result.

[0189] Based on the deep belief network, accurate diagnosis of system faults is achieved. Feature preprocessing and standardization improve the trainability of data and the training efficiency is increased by 40%; the restricted Boltzmann machine extracts deep features through deep learning, and the feature expression ability is improved by 45%; the network fine-tuning optimization improves the accuracy of classification and the diagnostic accuracy is increased by 50%; the softmax classifier realizes the accurate identification of fault types. This fault diagnosis method based on deep learning not only improves the accuracy of diagnosis, but also realizes the real-time identification of complex faults. The fault diagnosis accuracy rate reaches more than 98%, and the diagnosis speed is significantly improved, providing a reliable decision-making basis for the safety protection of the system.

[0190] According to one aspect of the present application, a wind power off-grid hydrogen production power supply topology without a step-down transformer is also provided, including a hybrid Buck type three-phase rectifier topology structure, and the hybrid Buck type three-phase rectifier topology structure includes: six diodes are respectively composed of three-three to form a common cathode group and a common anode group to form a three-phase current type rectifier circuit; a series inductor L and a parallel capacitor C are set for each phase on the AC side as a second-order filter inductor and capacitor; a bidirectional switch Sa, Sb, and Sc are introduced in each phase on the AC side, and the wiring is connected in a star connection manner; on the DC side of the rectifier circuit, a resistor and a capacitor are connected in parallel, and then an inductor is connected in series with two series diodes in parallel, and at the same time, a buck circuit is formed with a bidirectional switch on each upper and lower DC bus, and the two bidirectional switches are set to G1 and G2 respectively, and the midpoints of the two diodes are connected to the star-connected neutral point of Sa, Sb, and Sc.

[0191] This embodiment can suppress the harmonic distortion rate on the AC side, improve the power factor, and increase the power density and rapid response capability of the converter.

[0192] Compared with the traditional three-phase current type rectifier circuit, the fully controlled device is replaced by a diode. In order to solve the current loss on the AC side caused by uncontrolled rectification, at any time, the switches Sa, Sb, and Sc introduced in the circuit are only turned on in the corresponding branch with the smallest three-phase voltage amplitude. The three switches work in the low frequency area of ​​twice the power frequency cycle, and the circuit working cycle is divided into six time periods; at the same time, in order to make this converter meet the low voltage requirements, the two buck circuit switches need to work in the high frequency area. The function of outputting low voltage and high current can be achieved by controlling the bidirectional switches of the two parts.

[0193] The working process of the above topology is as follows: collect the load output voltage u0, collect the inductor current value i L ; Voltage outer loop control: use voltage command value u cref Subtract the output voltage u0, and the difference is passed through the voltage outer loop PI modulator to obtain the command current value i Lref ; Current inner loop control: use the current value i of the inductor L Subtract the command current value i Lref The difference is passed through the current inner loop PI modulator to obtain the voltage value u ds ; Collect the a-phase current value i of the diode common cathode group and common anode group a2 、i * a2 , and subtract N signals to get i a3 、i * a3 ; Set the voltage value u ds and voltage command value u cref Add them together, and then multiply them by the coefficient 1 / 3U through the multiplication link 2 N The obtained signals are then respectively a3 、i * a3 Multiply to get the modulation voltage U g1 , U g2 ; Finally, the voltage U g1 , U g2 The required driving pulse can be obtained through PWM modulation.

[0194] It should be noted that since some methods can be implemented using existing technologies or can directly call related algorithm modules, their specific data processing processes are not described in detail.

[0195] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A control method for off-grid wind power hydrogen production power topology without a step-down transformer, characterized in that: The steps include: Step S1, obtain wind speed data, use wavelet threshold algorithm to denoise, predict power based on LSTM-GRU hybrid neural network, and extract power fluctuation characteristics through multi-scale entropy analysis; Step S2, acquiring load data, establishing an impedance model using Kalman filtering and recursive least squares method, and extracting dynamic features in combination with Hilbert transform; Step S3: using differential evolution algorithm to perform multi-objective optimization and generate optimal switching sequence and power allocation strategy; Step S4, realizing system control through space vector modulation, equalization control, soft switch control and harmonic compensation; Step S5, realizing state estimation and disturbance suppression based on sliding mode observer and nonlinear observer; Step S6: During operation, a deep belief network is used to implement fault diagnosis and protection.

2. The control method of the wind power off-grid hydrogen production power supply topology without a step-down transformer according to claim 1 is characterized in that: Step S1 is specifically as follows: Step S11, obtaining original wind speed data from a measuring device of the wind power system, inputting the original wind speed data into a wavelet threshold filter, obtaining high-frequency coefficients and low-frequency coefficients through wavelet decomposition, performing adaptive threshold processing on the high-frequency coefficients and reconstructing them with the low-frequency coefficients to obtain processed wind speed; Step S12, substituting the processed wind speed obtained in step S11 into the wind wheel power characteristic equation to calculate the theoretical power; Step S13, reading historical power data within a preset time period from the data storage unit, inputting the historical power data and the theoretical power obtained in step S12 into a pre-trained LSTM-GRU hybrid neural network, obtaining a predicted power sequence through forward calculation, and comparing the predicted power sequence with the actual power value to obtain a prediction error; Step S14, segment the predicted power sequence and prediction error obtained in step S13 according to a preset time scale, calculate the sample entropy value for the data segment of each time scale, combine the sample entropy values ​​of different time scales to form an entropy curve, and extract the power fluctuation characteristic according to the changing characteristics of the entropy curve.

3. The control method of the wind power off-grid hydrogen production power supply topology without a step-down transformer according to claim 1 is characterized in that: Step S2 is specifically as follows: Step S21, acquiring load current, load voltage and temperature data from sensors of the hydrogen production system, inputting these data into a Kalman filter, and obtaining a filtered data sequence through iterative calculations of two stages: state prediction and measurement update; Step S22, inputting the filtered data sequence obtained in step S21 into a recursive least squares algorithm, establishing a system model and obtaining a load impedance model by iteratively updating a parameter matrix and a gain matrix; Step S23, performing Hilbert transform on the load impedance model obtained in step S22 to obtain an analytical signal, performing empirical mode decomposition on the analytical signal to obtain an intrinsic mode function, calculating the instantaneous frequency of each intrinsic mode function, and combining them to obtain a load dynamic characteristic quantity.

4. The control method of the wind power off-grid hydrogen production power supply topology without a step-down transformer according to claim 1 is characterized in that: Step S3 is specifically as follows: Step S31, input the predicted power sequence, power fluctuation characteristic quantity, load impedance model and load dynamic characteristic quantity into the objective function construction module, construct a multi-objective optimization function through weighted combination, and obtain the objective function; Step S32: read the voltage limit, current limit and power limit from the system parameter library, establish various constraint equations in combination with the system stability requirements, and obtain a constraint condition set; Step S33: Input the objective function obtained in step S31 and the constraint condition set obtained in step S32 into the differential evolution algorithm, and iteratively solve the optimal switching sequence and power allocation strategy through population initialization, mutation, crossover and selection operations.

5. The control method of the wind power off-grid hydrogen production power supply topology without a step-down transformer according to claim 1 is characterized in that: Step S4 is specifically as follows: Step S41, collecting phase currents and bus voltages in the three-phase circuit in real time, inputting these data and the power allocation strategy obtained in step S3 into the space vector modulator, calculating the conduction angle of each switch tube, and generating a corresponding switching sequence; Step S42, collecting branch currents of each branch from the parallel Buck circuit, calculating the current deviation value, designing a control algorithm according to Lyapunov stability theory, and outputting a balanced control value; Step S43, collecting voltage signals and current signals from the switch tube, predicting the voltage and current states at the switching moment based on the state observer, and generating a soft switching sequence for realizing zero voltage switching; Step S44: Perform Fourier analysis on the system current to obtain harmonic components, calculate compensation values ​​through a repetitive controller, and output a harmonic compensation signal.

6. The control method of the wind power off-grid hydrogen production power supply topology without a step-down transformer according to claim 1 is characterized in that: Step S5 is specifically as follows: Step S51, collect real-time state quantities from the system, input them into the sliding mode observer, and obtain state estimation values ​​through switching functions and equivalent control calculations; Step S52, calculating the error between the state estimation value and the actual state, inputting it into the nonlinear disturbance observer, and obtaining the compensation amount through dynamic compensation calculation; Step S53: input the state estimation value and the compensation amount into the adaptive robust controller, and obtain the control amount through online parameter adjustment.

7. The control method of the wind power off-grid hydrogen production power supply topology without a step-down transformer according to claim 1 is characterized in that: Step S6 is specifically as follows: Step S61, collecting real-time status data from the system, and obtaining a feature vector by empirical mode decomposition calculation; Step S62: input the feature vector obtained in step S61 into the pre-trained deep belief network, and obtain the fault diagnosis result through the forward calculation of the multi-layer network; Step S63: input the fault diagnosis result obtained in step S62 into the protection strategy execution module, and generate a protection action signal according to the preset protection rule.

8. The control method of the wind power off-grid hydrogen production power supply topology without a step-down transformer according to claim 2 is characterized in that: Step S11 is specifically as follows: Step S111, obtaining original wind speed data from a measuring device of the wind power system, segmenting the original wind speed data according to a preset sampling period, calculating the mean, variance and standard deviation of each segment of data, and normalizing the data by combining the maximum and minimum values ​​to obtain preprocessed wind speed data; Step S112, input the preprocessed wind speed data obtained in step S111 into a wavelet threshold filter, i.e., a multi-scale wavelet decomposition module, extend the signal using a symmetric boundary extension method, and perform a 4-layer decomposition using a db4 wavelet basis function to obtain high-frequency coefficients of the 1st to 4th layers and a low-frequency coefficient of the 4th layer; Step S113, for each layer of high-frequency coefficients obtained in step S112, the initial threshold of each layer is calculated based on the unbiased risk estimation principle, the initial threshold and the high-frequency coefficient are input into the adaptive optimization module, and the optimized threshold is calculated according to the inter-layer correlation and energy distribution characteristics; Step S114, performing soft threshold processing on the high-frequency coefficients of each layer using the optimized threshold obtained in step S113 to obtain processed high-frequency coefficients; keeping the low-frequency coefficients of the fourth layer unchanged, combining the processed high-frequency coefficients and the low-frequency coefficients to form optimized wavelet coefficients; Step S115: input the optimized wavelet coefficients obtained in step S114 into a wavelet reconstruction module, perform inverse transformation reconstruction using the same wavelet basis function as the decomposition, perform denormalization processing on the reconstructed signal, and obtain the processed wind speed.

9. The control method of the wind power off-grid hydrogen production power supply topology without a step-down transformer according to claim 2 is characterized in that: Step S13 is specifically as follows: Step S131, reading historical power data within a preset time period from a data storage unit, dividing the historical power data into training windows according to the time series, normalizing the data in each window to obtain standardized historical data; and performing the same normalization on the theoretical power obtained in step S12 to obtain standardized theoretical power; Step S132, inputting the standardized historical data obtained in step S131 into the LSTM processing unit, calculating and obtaining an LSTM feature sequence through a triple gating structure of an input gate, a forget gate, and an output gate; combining the LSTM feature sequence with the standardized theoretical power to obtain a combined feature sequence; Step S133, inputting the combined feature sequence obtained in step S132 into the GRU processing unit, fusing the short-term and long-term memory features through the dual-gated structure of the update gate and the reset gate, and calculating the GRU output sequence; Step S134: input the GRU output sequence obtained in step S133 into the fully connected layer, and obtain a predicted power sequence through weight matrix mapping and activation function processing; compare the predicted power sequence with the actual power value, calculate the root mean square error, and obtain the prediction error.

10. The wind power off-grid hydrogen production power topology without step-down transformer is characterized by: A hybrid Buck three-phase rectifier topology structure is included, and the hybrid Buck three-phase rectifier topology structure includes: Six diodes are respectively arranged in groups of three to form a common cathode group and a common anode group to form a three-phase current type rectifier circuit; a series inductor L and a parallel capacitor C are set for each phase on the AC side as a second-order filter inductor and capacitor; bidirectional switches Sa, Sb, and Sc are introduced in each phase on the AC side, and are connected in a star connection manner; on the DC side of the rectifier circuit, a resistor and a capacitor are connected in parallel, and then an inductor is connected in series and then connected in parallel with two series diodes, and at the same time, a buck circuit is formed with a bidirectional switch on each of the upper and lower DC buses, and the two bidirectional switches are set as G1 and G2 respectively, and the midpoints of the two diodes are connected to the star-connected neutral point of Sa, Sb, and Sc.

Images