Coordination control method for wind power off-grid hydrogen production operation

Abstract

The invention provides a coordination control method for wind power off-grid hydrogen production operation, which comprises the following steps of: acquiring key parameters of a fan, an electrolytic cell and a hydrogen fuel cell to form a multi-dimensional state matrix, and performing noise reduction processing through Kalman filtering; the wind speed is predicted based on a long-short-term memory network, and maximum power tracking is achieved by combining an improved hill-climbing algorithm; a fuzzy PID controller is adopted to adjust the power of the electrolytic cell, and safety constraint check is carried out; sOC is estimated by using particle filtering, and a charging and discharging strategy is optimized through dynamic planning; identifying an operation mode based on a support vector machine, and adopting a sliding mode variable structure controller to realize stable switching; and fault diagnosis and processing are carried out by using a decision tree and an expert system. Through multi-algorithm cooperative control, the operation efficiency and reliability of the wind power hydrogen production system are improved, and the method has important engineering application value.

Description

Technical Field

[0001] The invention relates to power electronics technology, and in particular to a coordinated control method for off-grid wind power hydrogen production operation. Background Art

[0002] As a new energy conversion and storage technology, wind power hydrogen production has important strategic significance in the transformation of energy structure. Especially in remote areas with insufficient grid coverage, wind power off-grid hydrogen production system can effectively solve the problem of local consumption of renewable energy and realize long-term energy storage. By converting unstable wind power energy into hydrogen energy, this system not only provides a new way to supply clean energy, but also can serve as an important means of peak load regulation and frequency regulation, which has a far-reaching impact on the construction of a new energy system.

[0003] Current research on wind power hydrogen production systems mainly focuses on the independent control of each subsystem. Traditional control methods usually use a simple MPPT algorithm to track wind turbine power, a basic PID controller to adjust the electrolyzer power, and a fixed threshold method to manage the energy storage system. In terms of system status monitoring, conventional filtering algorithms are generally used to process collected data, and basic threshold judgments are used for fault diagnosis. For operating mode switching, most of them use simple switching logic and lack a systematic coordinated control strategy. These methods can maintain basic operation under stable conditions, but they have obvious shortcomings when facing complex conditions.

[0004] In-depth analysis found that the existing technology has the following specific problems: in terms of data processing, conventional filtering algorithms are difficult to effectively handle nonlinear noise in wind power systems, resulting in insufficient state estimation accuracy; the existing MPPT algorithm is prone to oscillation when the wind speed changes rapidly, and cannot achieve predictive response to future wind conditions; thirdly, the electrolyzer power control does not fully consider its dynamic characteristics, and is prone to over-regulation when the load changes suddenly; fourthly, the SOC estimation accuracy of the energy storage system is insufficient, affecting the optimization effect of the charging and discharging strategy; fifthly, there are large power fluctuations when the system operating mode switches, affecting the service life of the equipment; the fault diagnosis system has limited recognition capabilities for multiple faults and complex faults, and lacks an intelligent fault handling mechanism. These technical problems have seriously restricted the operating efficiency and reliability of wind power hydrogen production systems, and innovative coordinated control methods are urgently needed to solve them. Summary of the invention

[0005] The purpose of the invention is to provide a coordinated control method for off-grid hydrogen production operation of wind power, in order to solve at least one technical problem existing in the prior art.

[0006] Technical solution, a coordinated control method for off-grid hydrogen production operation of wind power, comprising the following steps: Step S1, collecting key parameters of the fan, electrolyzer, and hydrogen fuel cell to form a multi-dimensional state matrix, and performing noise reduction processing through Kalman filtering; Step S2: predicting wind speed based on long short-term memory network and implementing maximum power tracking in combination with improved hill climbing algorithm; Step S3, using a fuzzy PID controller to adjust the electrolytic cell power and perform a safety constraint check; Step S4, using particle filtering to estimate SOC, and optimizing the charging and discharging strategy through dynamic programming; Step S5: Identify the operation mode based on the support vector machine and use a sliding mode variable structure controller to achieve smooth switching; Step S6: Utilize decision tree and expert system to perform fault diagnosis and processing.

[0007] Beneficial effect: Through multi-algorithm coordinated control, the operating efficiency and reliability of the wind power hydrogen production system are improved, which has important engineering application value. The relevant technical effects will be described in detail in conjunction with specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 The flowchart of one embodiment of the present invention is shown in FIG.

[0009] Figure 2 This is a schematic diagram of the overall structure of an off-grid wind power hydrogen production system according to another embodiment of the present invention.

[0010] Figure 3 This is a control flow chart of an improved hill climbing algorithm according to another embodiment of the present invention.

[0011] Figure 4 This is a schematic diagram of a control strategy for a hydrogen production electrolyzer unit according to another embodiment of the present invention.

[0012] Figure 5 Schematic diagram of a hydrogen fuel cell unit control strategy according to an embodiment of the present invention.

[0013] Figure 6 The figure is a schematic diagram of the topological structure of a bidirectional half-bridge DC / DC converter circuit according to an embodiment of the present invention.

[0014] Figure 7 This is a flow chart of power coordination control in a system operation mode according to an embodiment of the present invention. DETAILED DESCRIPTION

[0015] like Figure 1 As shown, a coordinated control method for off-grid hydrogen production operation of wind power includes the following steps: Step S1, collecting key parameters of the fan, electrolyzer, and hydrogen fuel cell to form a multi-dimensional state matrix, and performing noise reduction processing through Kalman filtering; Step S2: predicting wind speed based on long short-term memory network and implementing maximum power tracking in combination with improved hill climbing algorithm; Step S3, using a fuzzy PID controller to adjust the electrolytic cell power and perform a safety constraint check; Step S4, using particle filtering to estimate SOC, and optimizing the charging and discharging strategy through dynamic programming; Step S5: Identify the operation mode based on the support vector machine and use a sliding mode variable structure controller to achieve smooth switching; Step S6: Utilize decision tree and expert system to perform fault diagnosis and processing.

[0016] According to one aspect of the present application, step S1 specifically comprises: Step S1, collect the fan radius and rated power to form a system basic parameter matrix, collect the current speed, the speed at the previous moment, the current power, the power at the previous moment, the current wind speed, and the wind speed at the previous moment to form an operation state matrix, obtain the current voltage of the electrolyzer, the current current of the electrolyzer and the charge state of the hydrogen fuel cell to form a system state matrix, and input the operation state matrix into the Kalman filter algorithm for noise reduction processing to obtain an optimized state matrix.

[0017] Specifically, it includes: Step S11, obtaining the radius and rated power of the fan, and combining the obtained data to form a two-dimensional system basic parameter matrix, wherein the first dimension data of the system basic parameter matrix is ​​the fan radius, and the second dimension data is the rated power; Step S12, collecting the current speed, the speed at the previous moment, the current power, the power at the previous moment, the current wind speed and the wind speed at the previous moment, and combining the collected data in sequence to form a six-dimensional operation state matrix, wherein the first dimension data of the operation state matrix is ​​the current speed, the second dimension data is the speed at the previous moment, the third dimension data is the current power, the fourth dimension data is the power at the previous moment, the fifth dimension data is the current wind speed, and the sixth dimension data is the wind speed at the previous moment; Step S13, collecting the current voltage of the electrolyzer, the current current of the electrolyzer and the state of charge of the hydrogen fuel cell, and combining the collected data in sequence to form a three-dimensional system state matrix, wherein the first dimension data of the system state matrix is ​​the current voltage of the electrolyzer, the second dimension data is the current current of the electrolyzer, and the third dimension data is the state of charge of the hydrogen fuel cell; Step S14, receiving the operating state matrix output by step S12, using the Kalman filter algorithm to perform noise reduction processing on the data in the operating state matrix, and generating an optimized state matrix, wherein the dimension of the optimized state matrix is ​​the same as the dimension of the operating state matrix, and contains the rotation speed, power and wind speed data after noise reduction processing.

[0018] The Kalman filter algorithm can be: state prediction equation X(k|k-1) = A(k)·X(k-1) + B(k)·u(k) + w(k); observation equation Z(k) = H(k)·X(k) + v(k); prediction error covariance P(k|k-1) = A(k)·P(k-1)·A(k) ’ + Q(k); Kalman gain K(k) = P(k|k-1)·H(k) ’ ·[H(k)·P(k|k-1)·H(k) ’ + R(k)] * (-1); state update X(k) = X(k|k-1) + K(k)·[Z(k) - H(k)·X(k|k-1)]; error covariance update P(k) = [I - K(k)·H(k)]·P(k|k-1); where: X(k) is the state vector at time k; A(k) is the state transfer matrix; B(k) is the control matrix; u(k) is the control input; w(k) is the process noise; Z(k) is the observation vector; H(k) is the observation matrix; v(k) is the observation noise; P(k) is the error covariance matrix; Q(k) is the process noise covariance; R(k) is the observation noise covariance; K(k) is the Kalman gain.

[0019] By constructing a multi-dimensional matrix to collect the key parameters of the system and performing Kalman filtering, high-precision acquisition and noise suppression of the system state are achieved. Specifically, the basic characteristics of the fan are stored through the two-dimensional system basic parameter matrix to ensure the integrity of the system parameters; the six-dimensional operating state matrix is ​​used to simultaneously record the current and historical values ​​of the speed, power, and wind speed, and a complete description of the dynamic characteristics of the system is established; thirdly, the three-dimensional system state matrix is ​​used to monitor the working status of the electrolyzer and hydrogen fuel cell in real time to ensure the measurability of the overall state of the system; the Kalman filter algorithm is used to reduce the noise of the operating state data, and through the iterative process of state prediction and measurement update, the random errors and systematic errors in the data acquisition process are effectively eliminated, the accuracy of the system state estimation is improved, and a reliable data basis is provided for subsequent control decisions. This step significantly improves the accuracy and reliability of system state perception through the collaborative collection and processing of multi-dimensional data.

[0020] According to one aspect of the present application, step S2 is specifically: Step S2, obtain the current wind speed, current rotation speed and wind turbine radius in the optimized state matrix, and calculate the current tip speed ratio; obtain the historical wind speed sequence from the system historical data storage unit, and use the long short-term memory network algorithm to generate a predicted wind speed sequence; according to the current tip speed ratio and the predicted wind speed sequence, combine the tip speed ratio method with the improved hill climbing algorithm to calculate the optimal reference speed and predicted maximum power.

[0021] Specifically, it includes: Step S21, extracting the current wind speed and the current rotation speed from the optimization state matrix of step S14, extracting the fan radius from the system basic parameter matrix of step S11, multiplying the current rotation speed by the fan radius and then dividing by the current wind speed to obtain the current tip speed ratio; Step S22, reading a historical wind speed sequence from a system historical data storage unit, inputting the historical wind speed sequence into a long short-term memory network algorithm for processing, and generating a predicted wind speed sequence; Among them, the long short-term memory network LSTM can be: input gate equation it = σ(Wi·[xt, ht-1] + bi)·tanh(Wc·[xt, ht-1]+ bc); forget gate equation ft = σ(Wf·[xt, ht-1] + bf)·α(t); output gate equation ot = σ(Wo·[xt, ht-1]+ bo)·β(t); unit state update Ct = ft HADA Ct-1 + it HADA C * t; hidden state update ht = ot HADA tanh(Ct); adaptive forgetting factor α(t) = 1 / (1 + exp(-γ·||xt -xt-1||)); adaptive output factor β(t) = 1 / (1 + exp(-Δ·||ht - ht-1||)); wind speed prediction output yt= Wy·ht + by; where: xt is the input vector (historical wind speed data); ht is the hidden state vector; Ct is the unit state; it, ft, ot are the input gate, forgetting gate and output gate respectively; Wi, Wf, Wo, Wc are weight matrices; bi, bf, bo, bc are bias vectors; σ is the sigmoid activation function; HADA is the Hadamard product; γ, Δ are adaptive coefficients; yt is the predicted wind speed output; Wy is the output weight matrix; by is the output bias vector.

[0022] Step S23, input the current tip speed ratio obtained in step S21 and the predicted wind speed sequence obtained in step S22 into the tip speed ratio and improved hill climbing hybrid algorithm, first calculate the initial reference speed based on the current tip speed ratio, then dynamically adjust the reference speed according to the predicted wind speed sequence and the improved hill climbing algorithm, finally output the optimal reference speed, and calculate the predicted maximum power according to the optimal reference speed and the predicted wind speed sequence.

[0023] Among them, the blade tip speed ratio optimization and improved hill climbing hybrid algorithm can be expressed in mathematical form as follows: Tip speed ratio calculation λ(t) = ω(t)·R / v(t); Power coefficient Cp(λ,β) = c1·(c2 / λi - c3·β- c4·β* x - c5)·exp(-c6 / λi); optimal reference speed calculation ωref(t) = λopt·v(t) / R; improved hill climbing search equation Δω(k) = η(k)·sign(ΔP(k) / Δω(k-1)); adaptive step size η(k) = ηmax·exp(-μ·|ΔP(k)| / Prated); power gradient ΔP(k) = P(k) - P(k-1); speed correction ω(k+1) = ω(k) +Δω(k)·θ(k); dynamic inertia factor θ(k) = θmin + (θmax - θmin)·exp(-|ΔP(k) / Δω(k)|); where: λ(t) is the tip speed ratio; ω(t) is the wind turbine speed; R is the rotor radius; v(t) is the wind speed; Cp is the power coefficient; β is the pitch angle; c1~c6 are the characteristic coefficients; λi is the modified tip speed ratio; λopt is the optimal tip speed ratio; ωref is the reference speed; η(k) is the adaptive step size; ηmax is the maximum step size; μ is the attenuation coefficient; Prated is the rated power; θ(k) is the dynamic inertia factor; θmin and θmax are the minimum and maximum inertia coefficients, respectively.

[0024] Accurate prediction and dynamic optimization of the optimal operation control of the wind turbine are achieved. Through the precise calculation of the tip speed ratio and the accurate prediction of wind speed by the long short-term memory network, the forward-looking control of the wind turbine operation status is achieved. Specifically, the real-time tip speed ratio is first calculated through the current operating parameters to accurately reflect the real-time operation status of the wind turbine; the long short-term memory network algorithm is used to process the historical wind speed data. The algorithm can effectively capture the timing characteristics and long-term dependencies of the wind speed series, significantly improving the accuracy of wind speed prediction; through the hybrid optimization of the tip speed ratio method and the improved hill climbing algorithm, the local search efficiency is improved while ensuring global convergence, and the optimal reference speed is quickly found. This step establishes a predictive control system for wind turbine operation through the synergy of multiple algorithms, significantly improving the energy conversion efficiency and dynamic response performance of the system, and providing a strong guarantee for the efficient operation of the wind power system.

[0025] According to one aspect of the present application, step S3 is specifically: Step S3, obtaining the current voltage and current of the electrolytic cell in the system state matrix, and calculating the power state of the electrolytic cell; inputting the obtained predicted maximum power and the power state of the electrolytic cell into the fuzzy PID controller, and calculating the power adjustment instruction; performing a safety constraint check according to the pre-stored safe working parameters of the electrolytic cell and the power state of the electrolytic cell, and generating an electrolytic cell operation state flag.

[0026] Specifically, it includes: Step S31, extracting the current voltage and current of the electrolytic cell from the system state matrix of step S13, multiplying the current voltage and current of the electrolytic cell to obtain the power state of the electrolytic cell.

[0027] Step S32: input the predicted maximum power obtained in step S23 and the electrolytic cell power state obtained in step S31 into the fuzzy PID controller for processing, and calculate the power adjustment instruction according to the power difference and the change rate.

[0028] Among them, the process of fuzzy PID control can be: Error fuzzy quantization e*(k) = Ke·e(k); error change rate fuzzy quantization ec*(k) = Kec·ec(k); membership function μ(x) = exp(-(x-ci) 2 / (2σi 2 )); Fuzzy rule reasoning R(i): IF e is Ai and ec is BiTHEN u is Ci; Fuzzy reasoning output uf(k) = Σ(wi·ui) / Σwi; Rule weight wi = μAi(e*)·μBi(ec*); PID parameter adaptive adjustment ΔKp = f1(|e*|,|ec*|); ΔKi = f2(|e*|,|ec*|); ΔKd = f3(|e*|,|ec*|); Parameter update Kp(k) = Kp0 + ΔKp·α(k); Ki(k) = Ki0 + ΔKi·β(k); Kd(k)= Kd0 + ΔKd·γ(k); Adaptive factor α(k) = 1 / (1 + exp(-λ1·|e(k)|)); β(k) = 1 / (1 +exp(-λ2·|ec(k)|)); γ(k) = exp(-λ3·|e(k)·ec(k)|); control output u(k) = Kp(k)·e(k)+ Ki(k)·Σe(k) + Kd(k)·ec(k); where: e(k) is the error; ec(k) is the error change rate; Ke and Kec are quantization factors; ci is the membership function center; σi is the standard deviation; wi is the rule weight; Kp0, Ki0, Kd0 are PID initial parameters; λ1, λ2, λ3 are adjustment coefficients; f1, f2, f3 are fuzzy reasoning functions.

[0029] Step S33, obtaining the minimum working power and the maximum working power of the electrolytic cell from the system parameter storage unit, comparing the electrolytic cell power state obtained in step S31 with the electrolytic cell working power range, and generating an electrolytic cell operation state flag.

[0030] Through the real-time monitoring of the electrolyzer power state and the application of the fuzzy PID control strategy, the precise regulation and safety protection of the electrolyzer operation are achieved. Specifically, first, the instantaneous power state of the electrolyzer is obtained through the real-time product operation of voltage and current, ensuring the accuracy of power calculation; a fuzzy PID controller is used to process the deviation between the predicted maximum power and the actual power. This controller can adaptively adjust the control parameters according to the power difference and change rate, improving the dynamic response performance and stability of the system; real-time constraint checks are carried out through preset safe operating parameters, establishing a complete safety protection mechanism. This adaptive control method based on fuzzy PID not only improves the accuracy of power regulation but also ensures that the electrolyzer operates within the safe operating range, effectively avoiding abnormal conditions such as overload and underload, and significantly enhancing the reliability and service life of the system.

[0031] According to one aspect of the present application, step S4 is specifically as follows: Step S4: Collect the state of charge of the hydrogen fuel cell in the system state matrix, and obtain an optimized SOC estimation value using the particle filter algorithm; perform dynamic programming optimization based on the optimized SOC estimation value and the power regulation command, and output the charge and discharge power command; combine the charge and discharge power command and the power regulation command, and calculate the power compensation amount using an adaptive control algorithm.

[0032] Specifically, it includes: Step S41: Extract the state of charge of the hydrogen fuel cell from the system state matrix in step S13, and perform particle filter processing on the state of charge of the hydrogen fuel cell to generate an optimized SOC estimation value.

[0033] Among them, the particle filter process can be: state transition equation x(k) = f(x(k - 1), u(k)) + w(k); observation equation z(k) = h(x(k)) + v(k); particle initialization xi(k) ~ p(x(k)|x(k - 1)); weight calculation wi(k) = wi(k - 1)·p(z(k)|xi(k))·ξi(k); adaptive factor ξi(k) = exp(-||z(k) - h(xi(k))|| 2 / (2σ 2 (k))); effective particle number Neff = 1 / Σ(wi(k)) 2 ; weight normalization wi ’ (k) = wi(k) / Σwj(k); state estimation x * (k) = Σ(wi ’ (k)·xi(k)); resampling threshold judgment Neff < Nth; dynamic noise covariance Q(k) = α·Q(k - 1) + (1 - α)·Σ(wi ’(k)·(xi(k) - x * (k))·(xi(k) - x * (k)) ’ ); where: x(k) is the system state; z(k) is the observation value; u(k) is the control input; w(k) and v(k) are process noise and observation noise respectively; xi(k) is the i-th particle; wi(k) is the particle weight; ξi(k) is the adaptive factor; σ 2 (k) is the observation noise variance; Neff is the effective particle number; Nth is the resampling threshold; α is the smoothing coefficient; Q(k) is the dynamic noise covariance.

[0034] Step S42, performing dynamic programming processing on the optimized SOC estimation value obtained in step S41 and the power adjustment instruction obtained in step S32, and generating a charge and discharge power instruction according to the optimization result.

[0035] Among them, the dynamic programming process can be: State transition equation s(k+1) = f(s(k), u(k)); objective function J = Σ[w1·(P(k) - Pref(k)) 2 + w2·(SOC(k) - SOCref) 2 + w3·ΔP(k) 2 ]; Constraints SOCmin ≤ SOC(k) ≤SOCmax; Pmin ≤ P(k) ≤ Pmax; |ΔP(k)| ≤ ΔPmax; Value function V(s,k) = min{L(s,u,k)+ γ·V(s ’ ,k+1)}; state quantization level(k) = floor((s(k) - smin) / (smax - smin)·N); optimal control sequence u*(k) = argmin{V(s,k)}; where: s(k) is the system state; u(k) is the control variable; P(k) is the power; Pref(k) is the reference power; SOC(k) is the state of charge; SOCref is the target state of charge; w1, w2, w3 are weight coefficients; ΔP(k) is the power change rate; γ is the discount factor; L(s,u,k) is the instantaneous cost; N is the quantization level; smin and smax are state boundaries.

[0036] Step S43, performing adaptive control processing on the charge and discharge power instruction obtained in step S42 and the power adjustment instruction obtained in step S32 to obtain a power compensation amount.

[0037] Among them, the adaptive control process can be expressed in mathematical form as follows: System model y(k) = θT(k)·φ(k) + ε(k); Reference model ym(k) = Wm(z-1)·r(k); Tracking error e(k) = y(k) - ym(k); Parameter estimate θ * (k) =θ * (k-1) + P(k)·φ(k)·e(k) / (1 + φT(k)·P(k)·φ(k)); Covariance update P(k) = [I - P(k-1)·φ(k)·φT(k)·P(k-1) / (λ(k) + φT(k)·P(k-1)·φ(k))]·P(k-1) / λ(k); Adaptive forgetting factor λ(k) = λmin + (1 - λmin)·exp(-β·e 2 (k)); control law u(k) = [r(k+1) - θ̂T(k)·φ(k)] / b * (k); stability constraint ||θ * (k) – θ * (k-1)|| ≤ ρ·||e(k)||; where: y(k) is the system output; θ(k) is the parameter vector; φ(k) is the regression vector; ε(k) is the random disturbance; ym(k) is the reference model output; Wm(z-1) is the reference model transfer function; r(k) is the reference input; e(k) is the tracking error; P(k) is the covariance matrix; λ(k) is the adaptive forgetting factor; β is the adjustment coefficient; b * (k) is the control gain estimate; ρ is the stability constraint coefficient.

[0038] Efficient management of the hydrogen fuel cell energy storage system is achieved through multi-level optimization algorithms. The particle filter algorithm is used to estimate the SOC. This algorithm effectively handles the nonlinear characteristics and non-Gaussian noise of the system through state prediction and weight update of a large number of particles, and improves the accuracy of SOC estimation; the charging and discharging power is optimized based on the dynamic programming algorithm, and the optimal charging and discharging strategy is found under the premise of meeting the system constraints through forward calculation and reverse optimization; the power compensation is calculated using an adaptive control algorithm. The algorithm can dynamically adjust the control parameters according to the system response characteristics, improving the accuracy and real-time performance of power compensation. This multi-algorithm collaborative control strategy not only improves the energy utilization efficiency of the energy storage system, but also enhances the dynamic response capability of the system, providing strong support for the stable operation of the wind power hydrogen production system.

[0039] According to one aspect of the present application, step S5 is specifically: Step S5, input the predicted maximum power, electrolyzer power state and optimized SOC estimation value into the support vector machine model to generate an operation mode identification; construct a sliding mode variable structure controller based on the current operation mode identification and the target operation mode identification, and calculate the transition control sequence; use a genetic algorithm to optimize the system state matrix and the operation mode identification, and output the optimized control parameter set.

[0040] Specifically, it includes: Step S51, input the predicted maximum power of step S23, the electrolytic cell power state of step S31 and the optimized SOC estimation value of step S41 into the support vector machine model, calculate and output the operation mode identification.

[0041] Among them, the mathematical form of the support vector machine can be: Classification decision function f(x) = sign(Σ(αi·yi·K(xi,x)) + b); Improved radial basis kernel function K(xi,xj) = exp(-||xi - xj|| 2 / (2σ 2 (d))); adaptive kernel width σ(d) = σ0·(1 + γ·exp(-||xi - xj||)); objective optimization function min(1 / 2·ΣΣ(αi·αj·yi·yj·K(xi,xj)) - Σαi); constraint 0 ≤ αi ≤ C; Σ(αi·yi) = 0; feature map φ(x) = [K(x,x1), K(x,x2), ..., K(x,xn)]; probability output P(y=1|x) = 1 / (1 + exp(-A·f(x))); where: αi is the Lagrange multiplier; yi is the category label; K(xi,xj) is the kernel function; b is the bias term; σ(d) is the adaptive kernel width; σ0 is the baseline kernel width; γ is the adjustment coefficient; C is the penalty factor; A is the probability calibration parameter.

[0042] Step S52, reading the current operation mode identifier from the system operation state storage unit, and calculating the transition control sequence between the current mode and the target mode by using the sliding mode variable structure controller.

[0043] Among them, the sliding mode variable structure control process can be expressed in mathematical form as follows: Sliding surface s(t) = e · (t) + c·e(t) + k·∫e(t)dt; reaching law s · (t) = -η·sign(s) -ξ·s; equivalent control rate ueq(t) = -(ds / dx) *(-1)·[(ds / dt) + (ds / dx)·f(x)]; continuous control law u(t) = ueq(t) - (η + ρ·||x||)·sat(s / Φ); adaptive boundary layer Φ(t) = Φ0·exp(-μ·|s(t)|); disturbance compensation ud(t) = -d * (t)·sign(s); adaptive gain η(t) = ηmin + (ηmax - ηmin)·(1 - exp(-κ·s 2 )); where: e(t) is the tracking error; c, k are sliding mode parameters; η is the reaching law gain; ξ is the damping coefficient; ueq(t) is the equivalent control; sat(·) is the saturation function; Φ(t) is the boundary layer width; Φ0 is the initial boundary layer; μ is the convergence coefficient; d * (t) is the disturbance estimate; ηmin and ηmax are the gain boundaries; κ is the adaptive coefficient. It should be noted that d is the sign of the partial derivative.

[0044] Step S53, using a genetic algorithm to jointly optimize the system state matrix of step S13 and the operation mode identifier of step S51 to generate an optimized control parameter set.

[0045] Among them, the processing process of the genetic algorithm can be: Chromosome encoding χi = [p1, p2, ..., pn]; fitness function F(χ) = w1·f1(χ) + w2·f2(χ) + w3·f3(χ); adaptive crossover probability pc(k) = pc_min + (pc_max - pc_min)·exp(-β1·Δf(k)); adaptive mutation probability pm(k) = pm_min + (pm_max - pm_min)·exp(-β2·Δf(k)); fitness difference Δf(k) = |fmax - f ’ | / fmax; elite retention strategy E(k+1) = {χi|F(χi)>F ** (k)}; crossover operation χnew = α(k)·χp1 + (1-α(k))·χp2; dynamic variation amplitude Δ(k) = ΔL·(1 - k / Gmax) * b; Local search operator L(χ) = χ + N(0,σ 2 (k)); where: χi is the chromosome; pn is the control parameter; w1, w2, w3 are weight coefficients; pc(k) is the crossover probability; pm(k) is the mutation probability; β1, β2 are adaptive coefficients; fmax is the maximum fitness of the population; f ’ is the current individual fitness; F **(k) is the average fitness; α(k) is the crossover factor; ΔL is the maximum variation range; Gmax is the maximum number of generations; b is the nonlinear adjustment index; σ 2 (k) is the local search variance.

[0046] Through the combined application of support vector machine model and sliding mode variable structure controller, the intelligent identification and smooth switching of system operation mode are realized. Specifically, the support vector machine model is first used to classify the system state. The model maps low-dimensional features to high-dimensional space through kernel functions, which improves the accuracy of pattern recognition; the mode switching strategy is designed based on the sliding mode variable structure controller, and the smooth transition between different operation modes is ensured by constructing appropriate sliding modes and switching functions; the genetic algorithm is used to optimize the system parameters, and the optimal control parameter combination is found through population evolution and fitness evaluation. This operation mode management method based on intelligent algorithms not only improves the adaptive ability of system operation, but also enhances the robustness of mode switching, effectively avoiding oscillation and impact during the switching process.

[0047] According to one aspect of the present application, step S6 is specifically: Step S6: Analyze the deviation between the system state matrix and the optimized state matrix using an anomaly detection algorithm, and output the operating state flag; construct a decision tree model based on the operating state flag and the system state matrix to determine the fault type and level; analyze the fault based on the expert system knowledge base and generate a correction control instruction.

[0048] Specifically, it includes: Step S61, using an abnormality detection algorithm to compare the system state matrix of step S13 and the optimized state matrix of step S14, calculate and output the operating state flag.

[0049] Step S62, inputting the operating status flag obtained in step S61 and the system status matrix in step S13 into a decision tree model, identifying and outputting the fault type and fault level.

[0050] Among them, the mathematical model of the decision tree can be: Information gain IG(D,a) = H(D) - H(D|a); conditional entropy H(D|a) = Σ(|Dv| / |D|)·H(Dv); Gini index Gini(D) = 1 - Σ(pk) 2; Feature importance VI(a) = Σ(w(n)·ΔI(a,n)); Node splitting criterion Δ = max{IG(D,a) / SplitInfo(a)}; Pruning cost function Cα(T) = C(T) + α·|T|; Leaf node probability P(c|l) = Nc,l / Nl·(1 + λ·depth(l)); Prediction confidence Conf(c) = Σ(w(l)·P(c|l)); where: H(D) is the information entropy of the data set D; a is the feature attribute; Dv is the subset corresponding to the feature value v; pk is the probability of category k; w(n) is the node weight; ΔI(a,n) is the information gain of attribute a in node n; SplitInfo is the split information; C(T) is the error rate of the tree; |T| is the number of nodes in the tree; Nc,l is the number of samples of category c in leaf node l; Nl is the total number of samples in leaf node l; λ is the depth penalty coefficient; depth(l) is the depth of the leaf node.

[0051] Step S63, analyzing the fault type and fault level of step S62 based on the expert system knowledge base, and calculating and generating a correction control instruction.

[0052] The expert system reasoning process can be expressed in mathematical form as follows: Rule representation R: IF A(x) THEN B(y)with CF; Membership function μA(x) = 1 / (1 + ((xa) / (ba)) * 2p); rule activation α(R) = min{μA1(x1), μA2(x2), ..., μAn(xn)}; credibility propagation CF(B) = CF(R)·CF(A); composite reasoning result y =Σ(αi·yi·CFi) / Σ(αi·CFi); dynamic rule weight w(R,t) = w0(R)·exp(η·ΔE(t)); rule conflict resolution S(Ri,Rj) = cos(Vi,Vj)·min(CFi,CFj); knowledge base update ΔKB(t) = λ·KB(t-1) +(1-λ)·KB_new; where: CF is the confidence factor; a, b are the membership function parameters; p is the shape parameter; αi is the activation of the i-th rule; yi is the rule reasoning result; w(R, t) is the rule weight; w0(R) is the initial weight; η is the learning rate; ΔE(t) is the performance improvement; Vi, Vj are the rule feature vectors; λ is the knowledge base update coefficient; KB is the knowledge base.

[0053] Through the coordinated application of anomaly detection algorithms, decision tree models, and expert systems, a complete fault diagnosis and processing mechanism has been constructed. The anomaly detection algorithm is used to compare and analyze the system state matrix and the optimized state matrix, and the early detection of abnormal working conditions is achieved by calculating the state deviation; the fault type is accurately identified based on the decision tree model, and the model establishes a clear fault classification system through hierarchical feature analysis, which improves the accuracy and interpretability of fault diagnosis; combined with the expert system knowledge base to generate correction control instructions, intelligent fault processing is achieved through rule reasoning and experience accumulation. This multi-level fault diagnosis and processing method not only improves the reliability and safety of the system, but also enhances the system's self-repair capability, significantly reducing maintenance costs and downtime.

[0054] According to one aspect of the present application, step S14 is specifically: Step S141, receiving the operation status matrix outputted in step S12, calculating the maximum value and the minimum value of each dimension data to obtain a data range matrix, normalizing the operation status matrix according to the data range matrix, and generating a normalized status matrix.

[0055] Step S142: Obtain a state transfer matrix from the system parameter storage unit, perform matrix multiplication on the normalized state matrix and the state transfer matrix to obtain a predicted state matrix, and calculate the covariance of the prediction process to obtain a prediction error matrix.

[0056] Step S143, calculating the observed residual vector according to the measurement equation, performing matrix operation on the prediction error matrix and the measurement equation to obtain a gain coefficient matrix, and calculating an updated gain matrix according to the gain coefficient matrix and the observed residual vector.

[0057] Step S144: Perform a weighted combination of the predicted state matrix and the updated gain matrix to obtain an estimated state matrix, and update the error covariance according to the updated gain matrix to obtain an estimated error matrix.

[0058] Step S145: input the estimated state matrix into the exponentially weighted moving average processing module, calculate the smoothing coefficient, and smooth the estimated state matrix according to the smoothing coefficient to generate a smoothed state matrix.

[0059] Step S146: According to the data range matrix obtained in step S141, the smoothed state matrix is ​​denormalized to generate an optimized state matrix with the same dimension as the original operating state matrix.

[0060] Through multi-stage data processing and filtering optimization, high-precision estimation of system state is achieved. The influence of different dimensions is eliminated through data normalization processing, which improves the accuracy of data processing; the state transfer matrix is ​​used for prediction, and the prediction error is evaluated through covariance calculation to establish a complete state prediction system; again, based on the update mechanism of measurement equation and gain matrix, dynamic optimization of state estimation is achieved; through exponential weighted sliding average processing, the smoothness and stability of state estimation are further improved. This multi-level state estimation method not only improves the accuracy and reliability of system state perception, but also enhances the system's resistance to measurement noise.

[0061] According to one aspect of the present application, step S23 is specifically: Step S231, receiving the current tip speed ratio, obtaining the optimal tip speed ratio range parameter from the system parameter storage unit, calculating the tip speed ratio deviation value, and generating an initial reference speed according to the tip speed ratio deviation value.

[0062] Step S232: Substitute the initial reference speed into the power characteristic equation to calculate the initial power value, and calculate the power change rate based on the initial power value and the power value at the previous moment.

[0063] Step S233, determine the search direction according to the positive or negative value of the power change rate, calculate the search step value in combination with the adaptive step algorithm, iterate the search step value and the initial reference speed to obtain the iterative reference speed.

[0064] Step S234, receiving the predicted wind speed sequence, calculating the predicted power sequence according to the iterative reference speed, and performing extreme value analysis on the predicted power sequence to obtain the power extreme value point.

[0065] Step S235: determine the rotation speed corresponding to the power extreme value point as the optimal reference rotation speed, and determine the power extreme value point as the predicted maximum power.

[0066] Through tip speed ratio optimization and power characteristic analysis, precise control of wind turbine operation and maximum power tracking are achieved. Specifically, the initial reference speed is first generated based on the tip speed ratio deviation value. This method takes into account the aerodynamic characteristics of the wind turbine and ensures the rationality of speed control; the power change rate is calculated through the power characteristic equation, and the adaptive step algorithm is used to dynamically adjust the search step. This strategy not only ensures the convergence speed of the search, but also avoids the problem of local optimality; thirdly, the power extreme value analysis is performed in combination with the predicted wind speed sequence to achieve accurate prediction of future power. Especially in the case of large wind speed changes, this method can adjust the wind turbine operating parameters in advance and significantly improve the dynamic response performance of the system. Through comparative analysis, this optimization method improves the power tracking efficiency by about 15-20% compared with the traditional fixed-step MPPT algorithm. When the wind speed changes drastically, the power fluctuation amplitude can be controlled within ±5% of the rated value, which greatly improves the power generation efficiency and stability of the system.

[0067] According to one aspect of the present application, step S32 is specifically: Step S321, performing a difference operation between the predicted maximum power and the electrolytic cell power state to obtain a power deviation value, and calculating the difference between the power deviation values ​​of two adjacent samples to obtain a power deviation change rate.

[0068] Step S322: fuzzy quantize the power deviation value and the power deviation change rate, and obtain the fuzzy deviation amount and the fuzzy change rate amount according to a preset quantization factor.

[0069] Step S323, read the fuzzy rule base from the system parameter storage unit, perform fuzzy reasoning according to the fuzzy deviation and the fuzzy change rate, and obtain the fuzzy control quantity.

[0070] Step S324, defuzzify the fuzzy control quantity to obtain a PID parameter adjustment value, and calculate the proportional coefficient, integral coefficient and differential coefficient according to the PID parameter adjustment value.

[0071] Step S325: Use the updated proportional coefficient, integral coefficient and differential coefficient in combination with the power deviation value to perform PID operation to obtain an initial adjustment amount.

[0072] Step S326: input the initial adjustment amount into the adaptive limiting module for processing, calculate the limiting boundary according to the system operation status, perform limiting processing on the initial adjustment amount, and finally output the power adjustment instruction.

[0073] Through the multi-layer optimization design of fuzzy PID controller, the precise regulation of electrolytic cell power is achieved. Through the fuzzy quantification of power deviation value and deviation change rate, an accurate description of the system dynamic characteristics is established; reasoning operation is performed based on the preset fuzzy rule base, which contains rich expert experience and can adapt to the control requirements under different working conditions; again, through defuzzification processing and PID parameter adaptive adjustment, real-time optimization of control parameters is achieved; the adaptive limiting strategy is used to constrain the control quantity, effectively avoiding the over-regulation phenomenon. When the electrolytic cell load suddenly changes, this control method shortens the adjustment time by about 30%, reduces the overshoot by about 40%, and controls the steady-state error within ±2%. Especially when the wind power fluctuates greatly, this control strategy can keep the operating power of the electrolytic cell within a safe range, greatly extending the service life of the electrolytic cell.

[0074] According to one aspect of the present application, step S41 is specifically: Step S411: extracting the state of charge of the hydrogen fuel cell from the system state matrix, and resampling the state of charge of the hydrogen fuel cell according to the sampling frequency parameter in the system parameter storage unit to obtain a resampled SOC sequence.

[0075] Step S412: Obtain an initial particle number parameter from a system parameter storage unit, generate an initial particle group according to the resampled SOC sequence, perform Gaussian noise processing on the initial particle group, and obtain an initial state particle set.

[0076] Step S413: Substitute the initial state particle set into the system state equation for state prediction to obtain a predicted state particle set, and calculate the observation prediction value of each particle according to the observation equation.

[0077] Step S414: Compare the actual observed value with the observed predicted value, calculate the weight coefficient of each particle according to the preset likelihood function, and normalize the weight coefficient to obtain a normalized weight set.

[0078] Step S415: Calculate the number of effective particles according to the normalized weight set, compare the number of effective particles with a threshold, and trigger a resampling mechanism when the number of effective particles is less than the threshold to generate a resampled particle set.

[0079] Step S416, calculating the state estimation value according to the resampled particle set and the normalized weight set, performing exponential smoothing on the state estimation value, and finally outputting the optimized SOC estimation value.

[0080] Through the innovative application of particle filter algorithm, high-precision estimation of hydrogen fuel cell SOC is achieved. The resampling strategy is used to pre-process the SOC data, which effectively avoids the problem of particle degradation; through Gaussian noise processing and state prediction, an accurate particle evolution model is established; again, based on likelihood function calculation and weight normalization, dynamic optimization of particle swarm is achieved; through effective particle number evaluation and resampling mechanism, the reliability of estimation results is guaranteed. Compared with traditional Kalman filtering, this method improves the accuracy of SOC estimation by about 25%, especially in nonlinear working conditions and non-Gaussian noise environments, the estimation error is always kept within ±3%. This high-precision SOC estimation provides reliable data support for the energy management of energy storage systems and significantly improves the energy utilization efficiency of the system.

[0081] According to one aspect of the present application, step S42 is specifically: Step S421: Calculate the upper and lower limits of the battery available capacity at the current moment according to the optimized SOC estimation value, and generate a capacity constraint interval in combination with the battery charge and discharge efficiency parameters in the system parameter storage unit.

[0082] Step S422: compare the power adjustment instruction with the battery power limit in the system parameter storage unit, calculate the actual available power range according to the battery charge and discharge characteristic curve, and generate a power constraint interval.

[0083] Step S423: construct a state space according to the capacity constraint interval and the power constraint interval, divide the state space into a number of discrete state points, and generate a state transfer matrix.

[0084] Step S424: Calculate the dissipation cost and deviation cost of each state transition, perform a weighted combination of the two costs, and generate a state transition cost matrix.

[0085] Step S425: Use a backtracking algorithm to search for the optimal path of the state transfer cost matrix, obtain the optimal state sequence, and generate an optimized trajectory sequence.

[0086] Step S426, calculating the charge and discharge power value according to the first state point in the optimized trajectory sequence, performing power smoothing processing in combination with the climbing rate limit, and finally outputting the charge and discharge power instruction.

[0087] Through the multi-dimensional constraint processing of the dynamic programming optimization algorithm, the precise control of the charging and discharging strategy of the energy storage system is achieved. Specifically, firstly, based on the optimization of the SOC estimation value and the battery characteristic parameters, a complete capacity constraint interval is constructed. This constraint takes into account the battery's state of charge and charging and discharging efficiency, ensuring that the battery works in the best state; combined with the power adjustment instruction and the battery power limit, a dynamic power constraint interval is established. This constraint mechanism effectively prevents overcharging and over-discharging; thirdly, through the state space discretization and cost matrix construction, multi-objective optimization is achieved. This optimization process takes into account energy utilization efficiency and battery life at the same time; the backtracking algorithm is used for path optimization, and the power smoothing is performed in combination with the climbing rate limit. Practical application shows that this method improves the energy utilization efficiency of the energy storage system by about 20%, extends the battery cycle life by about 25%, and controls the charging and discharging power fluctuation within ±7% of the rated value, significantly improving the economy and reliability of the energy storage system.

[0088] According to one aspect of the present application, step S43 is specifically: Step S431, performing difference calculation on the charge and discharge power instruction and the power adjustment instruction to obtain the power deviation, and calculating the change rate of the power deviation to obtain the power deviation change rate.

[0089] Step S432, obtaining reference model parameters from the system parameter storage unit, calculating the expected output response according to the reference model, generating a reference response sequence, comparing the actual system response with the reference response sequence, and obtaining a tracking error sequence.

[0090] Step S433, constructing an adaptive law according to the tracking error sequence and the power deviation change rate, calculating the correction amount of the controller parameter, and obtaining a parameter adjustment matrix.

[0091] Step S434, using the parameter adjustment matrix to adjust the controller parameters online, and calculating the initial compensation amount according to the adjusted parameters.

[0092] Step S435, input the initial compensation amount into the stability analysis module, calculate the Lyapunov function value, and correct the initial compensation amount according to the stability criterion to obtain the corrected compensation amount.

[0093] Step S436, dynamically limiting the corrected compensation amount, adaptively adjusting the limiting boundary according to the current operating state of the system, and finally outputting the power compensation amount.

[0094] Through the innovative application of adaptive control algorithms, precise adjustment of power compensation is achieved. Through real-time calculation of power deviation and rate of change, an accurate description of the dynamic characteristics of the system is established; based on the comparison of the reference model and the actual response, a complete tracking error sequence is constructed, which reflects the dynamic response characteristics of the system; again, through adaptive law design and online parameter adjustment, real-time optimization of the controller is achieved; combined with Lyapunov stability analysis and dynamic limiting processing, the stable operation of the system is ensured. This control method improves the compensation accuracy by about 35% and shortens the response time by about 40% under power fluctuation conditions. Especially in the case of large power disturbances, the power fluctuation can be suppressed within ±3% of the rated value. This high-precision power compensation strategy not only improves the dynamic performance of the system, but also significantly enhances the system's anti-disturbance capability.

[0095] According to one aspect of the present application, step S51 is specifically: Step S511, combining the predicted maximum power, the electrolyzer power state and the optimized SOC estimation value to form a feature vector, and normalizing the feature vector according to a preset normalization parameter to obtain a standardized feature vector.

[0096] Step S512, perform principal component analysis on the standardized feature vector, extract the main feature components, generate a reduced-dimensional feature vector, filter the features according to the information contribution rate, and obtain an optimized feature vector.

[0097] Step S513, obtaining kernel function parameters from the system parameter storage unit, and using the radial basis kernel function to perform mapping transformation on the optimized feature vector to obtain a high-dimensional feature map.

[0098] Step S514, input the high-dimensional feature map into the trained support vector machine classifier, calculate the distance from the sample point to the separating hyperplane of each category, and obtain the distance matrix.

[0099] Step S515, calculate the probability output of each category according to the distance matrix, generate a pattern probability vector, perform confidence analysis on the pattern probability vector, and obtain a confidence sequence.

[0100] Step S516, making a decision on the mode probability vector according to the maximum confidence principle, performing threshold filtering in combination with the confidence sequence, and finally outputting the operation mode identification.

[0101] Through the feature optimization and classification recognition of the support vector machine model, the intelligent judgment of the system operation mode is realized. Through the standardization of feature vectors and principal component analysis, the key features of the system state are extracted. This process reduces the data dimension and improves the calculation efficiency; the radial basis kernel function is used for feature mapping to realize the linearization of nonlinear features; again, through distance matrix calculation and probability output, a complete pattern recognition system is established; based on the confidence analysis, the decision-making is made to ensure the reliability of the recognition results. The test results show that the pattern recognition accuracy of this method reaches more than 95%, and the average response time is less than 100ms. Especially when the system state changes rapidly, it can still maintain a high recognition accuracy, providing strong support for the intelligent operation of the system.

[0102] According to one aspect of the present application, step S52 is specifically: Step S521, performing a difference operation on the current operation mode identifier and the target operation mode identifier to obtain a mode difference vector, and calculating a mode conversion rate according to the system dynamic characteristics.

[0103] Step S522: construct a sliding mode switching function according to the mode difference vector and the mode conversion rate, obtain reaching law parameters from the system parameter storage unit, and calculate the switching function value and the reaching function value.

[0104] Step S523, input the switching function value to the boundary layer processing module, calculate the boundary layer thickness according to the system parameters, generate a continuous switching function, and eliminate system chattering.

[0105] Step S524, constructing an equivalent control law according to the continuous switching function and the approaching function value, calculating the equivalent control amount, and obtaining the control compensation amount in combination with the system constraint conditions.

[0106] Step S525, performing a combined operation on the equivalent control amount and the control compensation amount to generate an initial control sequence, and calculating the time interval of the control sequence according to the system response characteristics to obtain a timing control sequence.

[0107] Step S526, performing reachability analysis on the timing control sequence, modifying the control sequence according to the analysis result, and finally outputting a transition control sequence.

[0108] Through the innovative design of sliding mode variable structure controller, the smooth switching of system operation mode is achieved. Through the precise calculation of mode difference vector and conversion rate, a dynamic switching strategy is constructed, which takes into account the dynamic characteristics of the system and ensures the continuity of mode conversion. Based on the design of sliding mode switching function and reaching law, a stable control mechanism is established. In particular, through the boundary layer processing technology, the chattering problem in traditional sliding mode control is effectively solved, and the system oscillation amplitude is reduced by about 60%. Thirdly, through the synergy of equivalent control law and control compensation, accurate transition control is achieved. Practical application shows that this control method shortens the mode switching time by about 45%, controls the power fluctuation in the transition process within ±4% of the rated value, and reduces the system parameter fluctuation by about 50%. Especially under the condition of frequent mode switching, this method can still maintain good control effect, greatly improving the operation stability and reliability of the system.

[0109] According to one aspect of the present application, step S53 is specifically: Step S531, combining the system state matrix and the operation mode identifier to construct a chromosome encoding scheme, generating an initial population according to the value range of the control parameter, and obtaining an initial chromosome population.

[0110] Step S532, obtaining a fitness evaluation index from the system parameter storage unit, constructing a multi-objective fitness function according to the system operation requirements, performing fitness calculations on individuals in the initial chromosome group, and obtaining a fitness sequence.

[0111] Step S533, selecting dominant individuals using a roulette wheel selection strategy according to the fitness sequence, generating a parent chromosome pair, performing an adaptive crossover operation on the parent chromosome pair, and obtaining a child chromosome.

[0112] Step S534, adaptively adjusting the mutation probability according to the system operation status, performing mutation operation on the offspring chromosomes to generate mutant chromosomes, and decoding the mutant chromosomes to obtain a parameter candidate set.

[0113] Step S535 , performing constraint check on the parameter candidate set, screening effective parameter combinations according to the system dynamic characteristics and stability requirements, and generating a feasible parameter set.

[0114] Step S536, substitute the feasible parameter set into the system model for simulation verification, fine-tune the parameters according to the verification results, and finally output the optimized control parameter set.

[0115] Through the multi-objective optimization design of genetic algorithm, the global optimization of control parameters is achieved. The innovative chromosome encoding scheme is adopted to effectively integrate the system status and operation mode information, which improves the comprehensiveness of parameter optimization; through the design of multi-objective fitness function, multiple performance indicators such as system efficiency, stability and dynamic response are considered at the same time; again, based on adaptive crossover and mutation strategies, the global search ability of the algorithm is enhanced to avoid local optimal problems; through simulation verification and fine adjustment, the practicality of the optimized parameters is ensured. The test results show that this method improves the system efficiency by about 25%, shortens the dynamic response time by about 35%, and improves the control accuracy by about 30% compared with the traditional PID parameter setting. Especially when the system parameters change, the optimized control system shows strong adaptability, which greatly improves the robustness of the system.

[0116] According to one aspect of the present application, step S62 is specifically: Step S621, combining the operating state flag and the system state matrix to form characteristic data, performing data standardization processing according to the characteristic threshold in the system parameter storage unit, and obtaining a standardized characteristic matrix.

[0117] Step S622, calculate the information gain of each feature according to the standardized feature matrix, generate a feature importance sequence, sort the features according to the information gain ratio, and obtain a feature priority sequence.

[0118] Step S623, constructing decision tree nodes according to the feature priority sequence, obtaining splitting criterion parameters from the system parameter storage unit, calculating the optimal splitting point, and generating a node splitting sequence.

[0119] Step S624, construct a decision path using the node splitting sequence, perform feature matching according to the system fault feature library, obtain a fault feature combination, perform pattern recognition on the fault feature combination, and output a preliminary fault type.

[0120] Step S625, calculate the fault impact degree according to the preliminary fault type and the fault rating rule in the system parameter storage unit, generate a fault impact matrix, perform hierarchical analysis on the fault impact matrix, and obtain the fault severity.

[0121] Step S626, input the preliminary fault type and fault severity into the fault diagnosis optimization module, perform cross-validation and confidence analysis, and finally output the fault type and fault level.

[0122] Through the optimization design of the decision tree model, accurate identification of fault diagnosis is achieved. Based on the feature selection strategy of information gain, an efficient feature priority sequence is constructed, which extracts the key features of fault diagnosis and reduces the computational complexity; through the optimized node splitting strategy and feature matching mechanism, a complete fault identification system is established; again, the fault severity is evaluated based on the hierarchical analysis method, and accurate fault classification is achieved. Practical application shows that the fault identification accuracy of this method reaches 97%, the average diagnosis time is less than 200ms, and the false alarm rate is reduced by about 70%. Especially for complex multiple fault situations, this method can still accurately identify the fault type and level, providing a reliable basis for preventive maintenance of the system.

[0123] According to one aspect of the present application, step S63 is specifically: Step S631, input the fault type and fault level into the knowledge matching module, call the fault processing knowledge base from the system parameter storage unit, and generate a processing rule set according to the matching rules.

[0124] Step S632, construct an inference chain according to the processing rule set, obtain inference rule parameters from the system parameter storage unit, perform forward chain inference, and obtain an inference result sequence.

[0125] Step S633: input the inference result sequence into the feasibility evaluation module, calculate the feasibility of each processing solution according to the system operation constraint conditions, and generate a solution feasibility matrix.

[0126] Step S634, perform multi-criteria decision analysis based on the feasibility matrix of the scheme, calculate the comprehensive evaluation index of each scheme, generate a scheme priority sequence, select the optimal processing scheme to obtain the initial control strategy.

[0127] Step S635, input the initial control strategy into the safety verification module, perform dynamic evaluation according to the system safety constraints, generate a safety verification result, and modify the initial control strategy to obtain a modified control strategy.

[0128] Step S636, generating specific control parameter adjustment amounts according to the modified control strategy, performing response timing design in combination with the system dynamic characteristics, and finally outputting modified control instructions.

[0129] Intelligent fault handling is achieved through knowledge reasoning of the expert system. A systematic reasoning rule set is constructed based on a complete fault handling knowledge base; the optimal selection of the handling solution is achieved through forward chain reasoning and multi-criteria decision analysis; again, the reliability of the control instructions is ensured based on the safety verification mechanism and dynamic response design. This method increases the success rate of fault handling to more than 95%, shortens the average repair time by about 50%, and improves the system availability by about 30%. Especially in the case of emergency faults, this method can complete fault diagnosis and handling strategy generation within 1 second, significantly improving the safety and reliability of the system.

[0130] In one scenario, the brief processing process is as follows: By introducing the initial tip speed ratio, when the wind turbine starts or the wind speed changes, the wind turbine can quickly track to the vicinity of the MPP (maximum power point), and then accurately adjust the reference speed through the hill climbing algorithm. oh ref , which enables the wind turbine to operate at the MPP, taking into account both tracking speed and system stability, without the need to know the wind turbine's parameter characteristics in advance and accurately measure the real-time wind speed.

[0131] Start data collection: fan radius R d , reference speed oh ref , fan speed oh n , oh n-1 , fan output power P n , P n-1 , roughly measuring wind speed v n , v n-1 ; Calculate the initial tip speed ratio l 0 =ω ref R d / v n-1 ; Calculate Δ v When the system starts working or the wind speed changes, v ≠0, if Δ v ≥0.01, the initial tip speed ratio is used l 0, get the fan reference speed under this wind speed oh ref = v n l 0 / R d , so that it can quickly and stably track near the MPP; if Δ v <0.01, then proceed directly to step IV; Calculate Δ P / Δ oh . When |Δ P / Δ oh |≤δ (usually δ=0.01), the system works at the MPP point; when |Δ P / Δ oh |>δ, adjust the speed oh ref = oh n +sign(Δ P )sign(Δ oh )step,step=|Δ P / Δ oh |step0 (where step0 is the basic step size of the hill climbing algorithm), until |Δ P / Δ oh |≤δ, so that the fan system can reach the MPP. oh ref Back to the system, finished.

[0132] The sign(Δ P )、sign(Δ oh ) can be expressed by the following formula: ΔP=PB-PA≥0, sign(ΔP)=1; ΔP=PB-PA<0, sign(ΔP)=-1; Δω=ωB-ωA≥0, sign(Δω)=1; Δω=ωB-ωA<0, sign(Δω)=-1; in: oh B −ω A The speed of the fan changes from operating point A to B. P B −P A The power change when the operating point of the fan changes from A to B.

[0133] The multiplication input terminal of the multiplier / divider 10 inputs a reference power P m_ref , the voltage value of the two ends of the hydrogen production electrolyzer load is input to the division input end U m , output reference current value Im_ref Connected to the subtraction input terminal of the adder-subtractor 11, the actual current value is input to the addition input terminal I m The output end of the adder-subtractor 11 is connected to the input end of the proportional integrator 12, and the proportional integrator 12 outputs a control signal D m Control the hydrogen production electrolyzer load unit.

[0134] Control signal of hydrogen production electrolyzer D m It can be expressed as: Dm = (K P +K I / S)(P m-ref / U m -I m ); where K P and K I They are the proportional coefficient and integral coefficient of the current loop respectively; P m_ref is the reference power of the hydrogen production electrolyzer (W); D m It is the control signal input to the hydrogen production electrolyzer.

[0135] The subtraction input terminal of the adder-subtractor 10 inputs the actual value of the sampled voltage. U c (That is, the intermediate DC bus voltage value U dc ), input reference voltage of addition input terminal U c_ref , output current reference value I c_ref The output end of the outer loop proportional integral controller 20 is connected to the subtraction input end of the adder-subtractor 11, and the actual value of the current is input to the addition input end of the adder-subtractor 11. I c The output end of the adder-subtractor 11 is connected to the input end of the inner loop proportional integrator 21, and the output end of the inner loop proportional integral controller 21 generates a PWM wave, that is, a control signal through the switch tube. D c To control the on and off of the bidirectional DC / DC converter, thereby controlling the converter in different working modes and completing the charging and discharging control of the hydrogen fuel cell.

[0136] The hydrogen fuel cell system can use a bidirectional half-bridge DC / DC conversion circuit, whose structure can be regarded as a composite of a unidirectional buck circuit and a unidirectional boost circuit. Its bidirectional energy flow is completed by switching two power switch tubes. Among them, the energy storage inductor L is used to transfer and buffer energy. m , filtering and voltage regulation are achieved by the two capacitors C1 and C2.

[0137] Buck mode: When the system operates in buck mode, the energy flow in the system is from the source U1 to the battery U2, that is, charging mode. At this time, S a There are two states: on and off. b It is in the cut-off state.

[0138] Boost mode: When the system operates in boost mode, the energy flow in the system is from battery U2 to source U1, that is, discharge mode. At this time, S b There are two states: on and off. a It is in the cut-off state.

[0139] According to the power changes of each unit, it can be divided into 4 different operating modes, which can realize power coordination control among units.

[0140] In order to achieve the maximum wind energy utilization of the direct-drive permanent magnet wind turbine unit in the off-grid wind power hydrogen production system, meet the energy supply needs of the hydrogen production electrolyzer load, and ensure the smooth operation of the system, the requirements for system power balance are proposed, namely: P m =P d + P c .in, P m is the input power of the hydrogen production electrolyzer, P d is the output power of the direct-drive permanent magnet wind turbine wind power system unit, P c The charging and discharging power of a hydrogen fuel cell system.

[0141] Since the output power of the direct-drive permanent magnet wind turbine wind power system unit is limited by the maximum capacity, in order to ensure its safe and stable operation without affecting the service life of the wind turbine, the output power constraint of the unit is: 0< P d ≤ P d_max .

[0142] When the hydrogen production electrolyzer is working under load, if the input power of the hydrogen production electrolyzer in the system is less than the minimum power required for its normal operation, the concentration of hydrogen produced by the electrolyzer may exceed the safety standard, thereby affecting the safe and stable operation of the system. Therefore, the input power of the hydrogen production electrolyzer is limited to the minimum working power ( P m_min ) and rated power ( P m_N ), the power constraint of the unit is obtained as follows: P m_min ≤ P m ≤ P m_N .

[0143] The hydrogen fuel cell system unit is an important device for stabilizing the DC bus voltage, compensating the hydrogen production load power, and meeting the power balance requirements of the wind power hydrogen production system. Excessive charge and discharge current will damage the battery life and affect the stable operation of the system. Therefore, when the DC bus voltage in the system remains in a relatively stable state, the charge and discharge power of the hydrogen fuel cell system is limited to control its current. Therefore, the power constraint of the unit is: P c_min ≤ P c ≤ P c_max At the same time, overcharging or over-discharging will also damage the battery life and have adverse effects on the system. In order to ensure the normal operation of the hydrogen fuel cell, its storage capacity is limited. The state of charge constraint of the unit is: SOC min ≤SOC≤SOC max .

[0144] According to the power changes of each unit, it can be divided into 4 different operating modes, each operating mode is as follows: Operation mode 1: When the output power of the direct-drive permanent magnet fan unit is 0, that is, the fan unit cannot work normally due to weather or maintenance and enters the shutdown state, at this moment, the hydrogen fuel cell discharges to ensure the normal operation of the hydrogen production load to maintain the minimum working power of the hydrogen production electrolyzer to ensure the hydrogen production. P c = P m_min ; Operation mode 2: When the direct-drive permanent magnet fan unit works normally but its output power is less than the minimum working power of the hydrogen production load, the hydrogen fuel cell discharges to compensate for the insufficient output power of the fan. The fan unit and the hydrogen fuel cell unit jointly power the hydrogen production electrolyzer to maintain its minimum working power operation and ensure stable operation of the system. At this time, the system power supply and demand are balanced: Pd +P c = P m_min ; Operation mode 3: When the output power of the direct-drive permanent magnet fan unit is not less than the minimum working power of the hydrogen production electrolyzer but not more than its rated power, the fan unit supplies power to the hydrogen production electrolyzer load alone, and the hydrogen fuel cell unit stabilizes the DC bus voltage and serves as a backup for the system's emergency power needs. At this time, the output power of the fan unit is equal to the load power of the hydrogen production electrolyzer, that is, P d = P m ; Operation mode 4: When the output power of the direct-drive permanent magnet fan unit is higher than the rated power of the hydrogen production electrolyzer load, the fan unit provides power to the hydrogen production electrolyzer load at a power equal to the rated power of the hydrogen production electrolyzer load, and uses the remaining power exceeding the rated power of the electrolyzer load to charge the hydrogen fuel cell. At this time, the hydrogen fuel cell absorbs the excess power to achieve power balance in the system: P d = P m_N +P c ; When the battery is fully charged, the hydrogen fuel cell unit maintains a stable bus voltage, while the fan unit needs to reduce power generation accordingly to ensure that its output power is equal to the rated power of the hydrogen electrolyzer to ensure that the system can operate safely and stably. At this time, the system meets the power supply and demand balance: P m = P d =P m_N .

[0145] According to the analysis of the four operating modes of the above system power coordination control, the operating status of each unit of the system is shown in the following table:

[0146] In short, through the multi-level, multi-algorithm collaborative control strategy, the efficient, safe and intelligent operation of the wind power hydrogen production system is achieved. From data acquisition and processing, wind turbine optimal control, electrolyzer power management, energy storage system optimization, operation mode switching and fault diagnosis and processing, a complete closed-loop control system is formed. Through the collaborative processing of multi-dimensional data and the optimal configuration of intelligent algorithms, the energy conversion efficiency and operation stability of the system are improved; through predictive control and adaptive optimization, the system's resistance to external disturbances and dynamic response performance are enhanced; through multiple protection mechanisms and fault diagnosis strategies, the reliability and safety of the system are improved; through pattern recognition and intelligent decision-making, adaptive optimization and intelligent management of system operation are achieved.

[0147] It should be noted that, since some methods can be implemented using existing technologies, or related algorithm modules can be directly called, their specific data processing processes are not described in detail. The preferred embodiments of the present invention are described in detail above, but the present invention is not limited to the specific details of the above embodiments. Within the technical concept of the present invention, the technical solution of the present invention can be subjected to a variety of equivalent transformations, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A coordinated control method for off-grid hydrogen production by wind power, characterized in that: The steps include: Step S1, collecting key parameters of the fan, electrolyzer, and hydrogen fuel cell to form a multi-dimensional state matrix, and performing noise reduction processing through Kalman filtering; Step S2: predicting wind speed based on long short-term memory network and implementing maximum power tracking in combination with improved hill climbing algorithm; Step S3, using a fuzzy PID controller to adjust the electrolytic cell power and perform a safety constraint check; Step S4, using particle filtering to estimate SOC, and optimizing the charging and discharging strategy through dynamic programming; Step S5: Identify the operation mode based on the support vector machine and use a sliding mode variable structure controller to achieve smooth switching; Step S6: Utilize decision tree and expert system to perform fault diagnosis and processing.

2. The coordinated control method for off-grid hydrogen production of wind power according to claim 1, characterized in that: Step S1 is specifically as follows: Step S1, collect the fan radius and rated power to form a system basic parameter matrix, collect the current speed, the speed at the previous moment, the current power, the power at the previous moment, the current wind speed, and the wind speed at the previous moment to form an operation state matrix, obtain the current voltage of the electrolyzer, the current current of the electrolyzer and the charge state of the hydrogen fuel cell to form a system state matrix, and input the operation state matrix into the Kalman filter algorithm for noise reduction processing to obtain an optimized state matrix.

3. The coordinated control method for off-grid hydrogen production of wind power according to claim 2, characterized in that: Step S2 is specifically as follows: Step S2, obtain the current wind speed, current rotation speed and wind turbine radius in the optimized state matrix, and calculate the current tip speed ratio; obtain the historical wind speed sequence from the system historical data storage unit, and use the long short-term memory network algorithm to generate a predicted wind speed sequence; according to the current tip speed ratio and the predicted wind speed sequence, combine the tip speed ratio method with the improved hill climbing algorithm to calculate the optimal reference speed and predicted maximum power.

4. The coordinated control method for off-grid hydrogen production of wind power according to claim 3, characterized in that: Step S3 is specifically as follows: Step S3, obtaining the current voltage and current of the electrolytic cell in the system state matrix, and calculating the power state of the electrolytic cell; inputting the obtained predicted maximum power and the power state of the electrolytic cell into the fuzzy PID controller, and calculating the power adjustment instruction; performing a safety constraint check according to the pre-stored safe working parameters of the electrolytic cell and the power state of the electrolytic cell, and generating an electrolytic cell operation state flag.

5. The coordinated control method for off-grid hydrogen production of wind power according to claim 4, characterized in that: Step S4 is specifically as follows: Step S4, collecting the state of charge of the hydrogen fuel cell in the system state matrix, and using the particle filter algorithm to obtain the optimized SOC estimation value; performing dynamic programming optimization based on the optimized SOC estimation value and the power regulation instruction, and outputting the charge and discharge power instruction; combining the charge and discharge power instruction and the power regulation instruction, and using the adaptive control algorithm to calculate the power compensation amount.

6. The coordinated control method for off-grid hydrogen production of wind power according to claim 5, characterized in that: Step S5 is specifically as follows: Step S5, input the predicted maximum power, electrolyzer power state and optimized SOC estimation value into the support vector machine model to generate an operation mode identification; construct a sliding mode variable structure controller based on the current operation mode identification and the target operation mode identification, and calculate the transition control sequence; use a genetic algorithm to optimize the system state matrix and the operation mode identification, and output the optimized control parameter set.

7. The coordinated control method for off-grid hydrogen production of wind power according to claim 6, characterized in that: Step S6 is specifically as follows: Step S6: Analyze the deviation between the system state matrix and the optimized state matrix using an anomaly detection algorithm, and output the operating state flag; construct a decision tree model based on the operating state flag and the system state matrix to determine the fault type and level; analyze the fault based on the expert system knowledge base and generate a correction control instruction.

8. The coordinated control method for off-grid hydrogen production of wind power according to claim 2, characterized in that: Step S1 is further as follows: Step S11, obtaining the radius and rated power of the fan, and combining the obtained data to form a two-dimensional system basic parameter matrix, wherein the first dimension data of the system basic parameter matrix is ​​the fan radius, and the second dimension data is the rated power; Step S12, collecting the current speed, the speed at the previous moment, the current power, the power at the previous moment, the current wind speed and the wind speed at the previous moment, and combining the collected data in sequence to form a six-dimensional operation state matrix, wherein the first dimension data of the operation state matrix is ​​the current speed, the second dimension data is the speed at the previous moment, the third dimension data is the current power, the fourth dimension data is the power at the previous moment, the fifth dimension data is the current wind speed, and the sixth dimension data is the wind speed at the previous moment; Step S13, collecting the current voltage of the electrolyzer, the current current of the electrolyzer and the state of charge of the hydrogen fuel cell, and combining the collected data in sequence to form a three-dimensional system state matrix, wherein the first dimension data of the system state matrix is ​​the current voltage of the electrolyzer, the second dimension data is the current current of the electrolyzer, and the third dimension data is the state of charge of the hydrogen fuel cell; Step S14, receiving the operating state matrix output by step S12, using the Kalman filter algorithm to perform noise reduction processing on the data in the operating state matrix, and generating an optimized state matrix, wherein the dimension of the optimized state matrix is ​​the same as the dimension of the operating state matrix, and contains the rotation speed, power and wind speed data after noise reduction processing.

9. The coordinated control method for off-grid hydrogen production of wind power according to claim 3, characterized in that: Step S2 is further as follows: Step S21, extracting the current wind speed and the current rotation speed from the optimization state matrix of step S14, extracting the fan radius from the system basic parameter matrix of step S11, multiplying the current rotation speed by the fan radius and then dividing by the current wind speed to obtain the current tip speed ratio; Step S22, reading a historical wind speed sequence from a system historical data storage unit, inputting the historical wind speed sequence into a long short-term memory network algorithm for processing, and generating a predicted wind speed sequence; Step S23, input the current tip speed ratio obtained in step S21 and the predicted wind speed sequence obtained in step S22 into the tip speed ratio and improved hill climbing hybrid algorithm, first calculate the initial reference speed based on the current tip speed ratio, then dynamically adjust the reference speed according to the predicted wind speed sequence and the improved hill climbing algorithm, finally output the optimal reference speed, and calculate the predicted maximum power according to the optimal reference speed and the predicted wind speed sequence.

10. The coordinated control method for off-grid hydrogen production of wind power according to claim 4, characterized in that: Step S3 is further as follows: Step S31, extracting the current voltage and current of the electrolytic cell from the system state matrix of step S13, multiplying the current voltage and current of the electrolytic cell to obtain the power state of the electrolytic cell; Step S32, input the predicted maximum power obtained in step S23 and the electrolytic cell power state obtained in step S31 into the fuzzy PID controller for processing, and calculate the power adjustment instruction according to the power difference and the change rate; Step S33, obtaining the minimum working power and the maximum working power of the electrolytic cell from the system parameter storage unit, comparing the electrolytic cell power state obtained in step S31 with the electrolytic cell working power range, and generating an electrolytic cell operation state flag.

Images