Optimal configuration method of off-grid wind power hydrogen production system considering full life cycle income

Abstract

The invention provides an optimal configuration method of an off-grid wind power hydrogen production system considering full life cycle income, which comprises the following steps: receiving original data, carrying out abnormal value detection and missing value repair, obtaining multi-scale component data through empirical mode decomposition, and generating a time sequence characteristic matrix; extracting equipment performance characteristic parameters and establishing a performance prediction model; constructing a system dynamic model based on a Bond Graph method, and performing stability analysis; calculating all costs and system income of the whole life cycle, and generating a comprehensive evaluation matrix in combination with risk analysis; a multi-objective optimization function is constructed, a grey wolf algorithm is adopted for solving, and an optimal configuration scheme is generated; and performing simulation verification and sensitivity analysis on the optimal configuration scheme to generate a final configuration scheme. According to the method, the equipment performance attenuation law and the system dynamic characteristics are fully considered, and the accuracy and reliability of optimal configuration are improved.

Description

Technical Field

[0001] The present invention relates to power electronics technology, and in particular to an optimization configuration method for an off-grid wind power hydrogen production system taking full life cycle benefits into consideration. Background Art

[0002] With the advancement of energy transformation, the importance of hydrogen as a clean energy carrier has become increasingly prominent. The use of renewable energy to produce hydrogen, especially wind power hydrogen production systems, has become an important way to achieve large-scale production of green hydrogen energy. Off-grid wind power hydrogen production systems can not only improve the utilization efficiency of renewable energy by converting fluctuating wind energy into stable hydrogen energy, but also provide clean energy solutions for remote areas. Optimizing the configuration of such systems from a full life cycle perspective is of great significance for reducing hydrogen production costs, improving system efficiency and ensuring long-term reliable operation.

[0003] At present, the research on off-grid wind power hydrogen production system mainly focuses on the performance optimization of single equipment and simple system integration. Traditional optimization configuration methods mostly use static planning or simple heuristic algorithms, mainly considering the initial investment cost and short-term operating efficiency of the equipment. In terms of performance prediction, commonly used methods include time series prediction based on statistical models and simple neural network models, which often ignore the nonlinear characteristics of equipment performance degradation. In terms of system modeling, simplified steady-state models or traditional dynamic modeling methods are mainly used, which makes it difficult to accurately describe the multi-energy flow coupling relationship and dynamic characteristics in the system. In terms of benefit evaluation, existing research focuses more on direct economic benefits, and the consideration of environmental benefits and risk factors is relatively one-sided.

[0004] However, existing technologies still face the following specific challenges when dealing with the optimization configuration of off-grid wind power hydrogen production systems: at the data processing level, it is difficult to effectively handle the non-stationary characteristics and multi-scale variation characteristics of parameters such as wind speed, resulting in insufficient accuracy in feature extraction; the equipment performance prediction model fails to fully consider the performance degradation law of electrolyzers and energy storage systems under different working conditions, affecting the accuracy of life cycle cost assessment; traditional system modeling methods cannot uniformly describe the energy conversion process in various forms such as mechanical, electrical and electrochemical, restricting the accuracy of system dynamic characteristics analysis; in addition, existing risk assessment methods are difficult to effectively quantify the impact of uncertain factors such as market price fluctuations and equipment failures on system benefits; when dealing with multi-objective optimization problems, optimization algorithms are difficult to balance solution efficiency and solution diversity, and are prone to fall into local optimality. These technical problems have seriously affected the optimization configuration effect and engineering implementation effect of off-grid wind power hydrogen production systems. Summary of the invention

[0005] The purpose of the invention is to provide an optimization configuration method for an off-grid wind power hydrogen production system taking into account the full life cycle benefits, in order to solve at least some of the technical problems existing in the prior art.

[0006] The technical solution is an optimization configuration method of an off-grid wind power hydrogen production system considering the full life cycle benefits, including the following steps: Step S1, receiving original wind speed data, equipment operation status data, environmental monitoring data and electricity price data sampled at predetermined time intervals, performing preprocessing, and generating a time series feature matrix; Step S2, receiving the time series feature matrix and pre-stored historical operation data, performing time series analysis, and generating a performance prediction model and a prediction parameter set; Step S3, receiving the performance prediction model, the prediction parameter set and the pre-stored system parameter data, establishing a system dynamic model, and generating a system characteristic matrix; Step S4, receiving the system characteristic matrix and pre-stored cost data, calculating cost evaluation data and benefit evaluation data, and performing risk analysis to generate a comprehensive evaluation matrix; Step S5, receiving the comprehensive evaluation matrix and the pre-stored optimization parameter data, constructing an objective function set and a constraint condition set; using the gray wolf algorithm to solve the objective function set and generate an optimization configuration scheme; Step S6: receiving the optimized configuration scheme and the pre-stored verification data set, performing simulation verification and sensitivity analysis; optimizing and adjusting the configuration scheme according to the analysis results, and generating a final configuration scheme.

[0007] Beneficial effect: The present invention fully considers the equipment performance attenuation law and the system dynamic characteristics, and improves the accuracy and reliability of the optimized configuration. The relevant technical advantages will be described in detail in conjunction with the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 is a flow chart of an embodiment of the present invention.

[0009] Figure 2 This is a topological diagram of a wind hydrogen storage system according to another embodiment of the present invention. DETAILED DESCRIPTION

[0010] like Figure 1 As shown, an optimization configuration method for an off-grid wind power hydrogen production system considering the full life cycle benefits specifically includes the following steps: Step S1, receiving original wind speed data, equipment operation status data, environmental monitoring data and electricity price data sampled at predetermined time intervals, using statistical analysis methods to detect outliers and repair missing values ​​on the data to obtain a cleaned data set; performing empirical mode decomposition on the cleaned data set to obtain multi-scale component data; extracting statistical features based on the multi-scale component data to obtain a feature vector set; processing the feature vector set using a sliding time window method with a preset time step to generate a time series feature matrix.

[0011] Step S2, receiving the timing feature matrix output by step S1 and the pre-stored historical operation data, extracting the equipment performance feature parameters, and obtaining the equipment feature data set; performing timing analysis on the equipment feature data set to obtain a performance decay sequence; establishing a prediction model according to the performance decay sequence, and generating a performance prediction model and a prediction parameter set.

[0012] Step S3, receiving the performance prediction model, prediction parameter set and pre-stored system parameter data output by step S2, establishing a system dynamic model, and obtaining a dynamic response model; performing stability analysis on the dynamic response model to obtain stability index data; generating a system characteristic matrix according to the dynamic response model and the stability index data.

[0013] Step S4, receiving the system characteristic matrix output from step S3 and the pre-stored cost data, calculating various costs of the entire life cycle, and obtaining cost evaluation data; calculating the system benefits based on the pre-stored market price data and environmental benefit data to obtain benefit evaluation data; combining the cost evaluation data and benefit evaluation data to perform risk analysis and generate a comprehensive evaluation matrix.

[0014] Step S5, receiving the comprehensive evaluation matrix output by step S4, the dynamic response model output by step S3 and the pre-stored optimization parameter data, constructing a multi-objective optimization function, and obtaining an objective function set; establishing constraint conditions according to the dynamic response model, and obtaining a constraint condition set; using the gray wolf algorithm to solve the objective function set, and generating an optimized configuration scheme.

[0015] Step S6, receiving the optimized configuration scheme outputted from step S5 and the pre-stored verification data set, performing simulation verification, and obtaining verification result data; performing sensitivity analysis on the verification result data to obtain sensitivity data; optimizing and adjusting the configuration scheme according to the verification result data and the sensitivity data, and generating a final configuration scheme.

[0016] Through six main steps, namely data processing, performance prediction, dynamic modeling, benefit evaluation, multi-objective optimization and verification optimization, the full life cycle optimization of system configuration is achieved. Multi-scale data analysis and deep learning methods are used to establish a high-precision data processing and performance prediction model; secondly, the system dynamic model is constructed through the Bond Graph method to achieve a unified description and analysis of the multi-energy flow system; thirdly, a multi-dimensional evaluation system including economic benefits, technical performance and environmental impact is established to achieve a comprehensive quantification of the system value; the configuration scheme is optimized and verified through the Gray Wolf Algorithm and digital twin technology. This method not only improves the optimization effect of system configuration, but also significantly improves the reliability and practicality of the optimization results, providing effective technical support for the engineering practice of off-grid wind power hydrogen production systems.

[0017] According to one aspect of the present application, step S1 specifically comprises: Step S11, receiving original wind speed data, equipment operation status data, environmental monitoring data and electricity price data, identifying abnormal values ​​using a preset multiple standard deviation method, and marking the identified abnormal values ​​as missing values; repairing all missing values ​​using a time series interpolation method to obtain a cleaned data set; normalizing the cleaned data set to obtain a standardized data set; In another embodiment of the present application, the mathematical form of the time series interpolation method can be: Interpolation function I(t) = ω·W(t) + (1-ω)·H(t); where: W(t) is the wavelet basis function interpolation, W(t) = Σ(cj,k·ψj,k(t)); H(t) is the Hermite interpolation, H(t) = Σ(yi·hi(t) + y ’ i·h ’ i(t)); ω is the adaptive mixing coefficient, ω = exp(-μ·D(t)); D(t) is the data fluctuation, D(t) = |y ’’ (t)| / (1+|y ’ (t)|); ψj,k(t) is the wavelet basis function; cj,k is the wavelet coefficient; yi is the node value; y ’ i is the node derivative value; hi(t) is the Hermite basis function; μ is the smoothing parameter; j and k are the wavelet scale and translation parameters. Combining the advantages of wavelet interpolation and Hermite interpolation, the smoothness and accuracy of interpolation are improved.

[0018] In another embodiment of the present application, the outlier detection process of the standard deviation can be: outlier determination function D(x) = |x - μ(t)| / (σ(t)·δ(t)); wherein: μ(t) is the local mean, μ(t) = (1 / n)·Σx(i), i∈[tn / 2,t+n / 2]; σ(t) is the local standard deviation, σ(t) = sqrt((1 / n)·Σ(x(i)-μ(t)) 2 ); δ(t) is the adaptive threshold coefficient, δ(t) = β·(1+|S(t)| / σ(t)); S(t) is the data slope, S(t) = (x(t)-x(t-1)) / Δt; n is the local window length; x is the data sequence; β is the benchmark threshold coefficient; t is the time index; i is the data index in the window; Δt is the sampling time interval. By introducing the adaptive threshold coefficient, the accuracy and robustness of outlier detection are improved.

[0019] Step S12, inputting the standardized data set into the empirical mode decomposition algorithm, extracting each order of intrinsic mode functions through iterative screening, and obtaining an eigenmode function sequence; calculating the instantaneous frequency of the eigenmode function sequence to obtain frequency characteristic data; combining the eigenmode function sequence and the frequency characteristic data to generate multi-scale component data; In another embodiment of the present application, the mathematical process of empirical mode decomposition is as follows: envelope constructor E(t) = α·S(t) + (1-α)·C(t); where: S(t) is a cubic spline interpolation function, S(t) = Σ(ai + bi(t-ti) +ci(t-ti) 2 + di(t-ti) 3 ); C(t) is the ideal envelope, C(t) = max(x(t))·exp(-β|t-tp|); α is the adaptive mixing coefficient, α = exp(-γ·|R(t)| / σ); R(t) is the residual function, R(t) = x(t) - m(t); m(t) is the mean function, m(t) = (Eup(t) + Edown(t)) / 2; σ is the signal standard deviation; γ is the adjustment coefficient; tp is the peak time; x(t) is the original signal; t is the time variable; ai, bi, ci, di are the spline interpolation coefficients; ti is the node time. By introducing the adaptive mixing coefficient α, the adaptive optimization of the envelope construction is realized, and the accuracy of the decomposition is improved.

[0020] Step S13, calculate the mean, standard deviation, skewness and kurtosis of each component in the multi-scale component data to obtain statistical feature data; perform correlation analysis on the statistical feature data to obtain a feature correlation matrix; select important features according to the feature correlation matrix according to a preset threshold value to generate a feature vector set; use a sliding time window with a preset time step to process the feature vector set and output a time series feature matrix.

[0021] Through multi-level data preprocessing and feature extraction methods, comprehensive cleaning and analysis of wind speed, equipment status, environment and electricity price data are achieved. Statistical methods are used to detect outliers and repair missing values ​​to ensure data quality and continuity; the empirical mode decomposition method is used to perform multi-scale analysis of the data. Compared with traditional Fourier analysis, this method can better handle nonlinear and non-stationary signals, and can effectively extract the intrinsic mode functions of parameters such as wind speed, revealing the inherent characteristics and changing laws of the data. By processing the feature vectors through a sliding time window, a time series feature matrix reflecting the dynamic characteristics of the system is constructed. This method not only retains the time series correlation of the data, but also realizes the dimensionality reduction and feature extraction of large-scale data. The entire step establishes a complete processing chain from raw data to feature matrix, providing a high-quality data foundation for subsequent system modeling and optimization, significantly improving the availability and representativeness of the data, and reducing the uncertainty in the subsequent modeling and optimization process.

[0022] According to one aspect of the present application, step S2 is specifically: Step S21, extracting parameters such as the efficiency, temperature, pressure and operation time of the electrolytic cell from the pre-stored historical operation data to obtain electrolytic cell characteristic data; extracting parameters such as the capacity, number of cycles, temperature and discharge depth of the energy storage system from the historical operation data to obtain energy storage characteristic data; combining the electrolytic cell characteristic data and the energy storage characteristic data to form a device characteristic data set; Step S22, segmenting the device feature data set into time series to obtain a segmented feature sequence; calculating the performance change rate of each time period of the segmented feature sequence to obtain performance change data; combining the segmented feature sequence and the performance change data to generate a performance decay sequence; Step S23, input the performance decay sequence into a multi-layer extreme learning machine and train to obtain a performance prediction model; use the performance prediction model to predict the equipment performance to obtain prediction result data; calculate the prediction accuracy and confidence interval based on the prediction result data to generate a prediction parameter set.

[0023] In another embodiment of the present application, the mathematical form of the multilayer extreme learning machine can be: output function Y(x) = Σ(βi·G(wi·x + bi)); wherein: G(z) is the activation function, G(z) = 1 / (1+exp(-θ(z))); θ(z) is the adaptive slope, θ(z) = α·(1+|z| / σ); wi is the input weight, wi = wi_init + η·δE / δwi; bi is the bias parameter, bi = bi_init + η·δE / δbi; βi is the output weight; x is the input vector; z is the intermediate variable; α is the slope adjustment coefficient; σ is the data standard deviation; η is the learning rate; E is the error function; wi_init and bi_init are initial parameters. By introducing the slope of the adaptive activation function, the nonlinear fitting ability of the model is improved.

[0024] A performance prediction system based on the historical operation data of the equipment was constructed. Through multi-level feature extraction and model training, the performance degradation of key system equipment (such as electrolyzers and energy storage systems) was accurately predicted. This step extracts the multi-dimensional characteristic parameters of the equipment from the historical data, including operating parameters such as efficiency, temperature, and pressure, identifies the performance degradation trend through time series analysis methods, and constructs a prediction model using a multi-layer extreme learning machine. The uniqueness of this method is that by analyzing the historical operation data of the equipment, a prediction model that reflects the law of equipment performance changes over time is established. It can not only predict the performance degradation trend of the equipment, but also evaluate the reliability and confidence interval of the prediction results. The output of the prediction model contains a complete set of prediction parameters, which provides a reliable basis for subsequent system optimization and maintenance decisions, significantly improves the predictability and maintainability of system operation, and effectively reduces maintenance costs and equipment failure risks.

[0025] According to one aspect of the present application, step S3 is specifically: Step S31, inputting the performance prediction model, the prediction parameter set and the system parameter data into the Bond Graph modeling algorithm to obtain an energy transfer model; performing causal analysis on the energy transfer model to obtain causal relationship data; establishing a state equation according to the causal relationship data to generate a dynamic response model; In another embodiment of the present application, the mathematical form of the Bond Graph modeling algorithm can be: energy transfer function P(t) = Σ(ei(t)·fi(t)); wherein: ei(t) is a generalized potential variable, ei(t) = Ti·δqi(t) / δt + Ri·fi(t); fi(t) is a generalized flow variable, fi(t) = (1 / Ci)·qi(t) + Li·δfi(t) / δt; qi(t) is a generalized displacement; Ti is a transformation coefficient; Ri is a damping coefficient; Ci is a capacitive parameter; Li is an inertial parameter; t is a time variable; i is an energy port index. By introducing a dynamic parameter adjustment mechanism, the accuracy of the energy transfer model is improved.

[0026] Step S32, linearizing the dynamic response model to obtain a linearized model; calculating the eigenvalues ​​of the linearized model to obtain an eigenvalue sequence; analyzing system stability according to the eigenvalue sequence to generate stability index data; Step S33, calculating the system response time and overshoot based on the dynamic response model to obtain response parameter data; combining the stability index data and the response parameter data to obtain performance evaluation data; organizing the performance evaluation data into a matrix form to generate a system characteristic matrix.

[0027] Based on the Bond Graph modeling method, the dynamic response model of the system was constructed to achieve an accurate description and analysis of the dynamic characteristics of the system. In this step, the energy transfer model was established, the system state equation was determined through causal analysis, and the stability analysis and performance evaluation of the model were performed. The Bond Graph method can uniformly describe the energy conversion processes in various forms such as mechanical, electrical and chemical in the system, so that the model has better physical meaning and interpretability. By linearizing the state equation, the eigenvalues ​​and stability indicators of the system were obtained, and then the dynamic response characteristics of the system were evaluated. This modeling method can not only accurately reflect the dynamic behavior of the system, but also reveal the energy transfer relationship between the components of the system, providing a theoretical basis for system optimization. The system characteristic matrix finally generated comprehensively reflects the dynamic characteristics and performance indicators of the system, providing a reliable basis for subsequent optimization configuration.

[0028] According to one aspect of the present application, step S4 is specifically: Step S41, extracting the initial investment, operation and maintenance cost and replacement cost of the equipment from the pre-stored cost data to obtain basic cost data; calculating the maintenance cost during the life cycle of the equipment according to the prediction parameter set output in step S2 to obtain maintenance cost data; combining the basic cost data and the maintenance cost data to form cost assessment data; Step S42: Calculate the hydrogen sales revenue based on the pre-stored market price data to obtain sales revenue data; calculate the carbon emission reduction benefit based on the pre-stored environmental benefit data to obtain environmental benefit data; integrate the sales revenue data and the environmental benefit data to form benefit evaluation data; Step S43: Analyze the volatility of cost assessment data and benefit assessment data using the conditional risk value method to obtain risk indicator data; calculate the confidence interval based on the risk indicator data to obtain the risk assessment result; combine the cost assessment data, benefit assessment data and risk assessment result to generate a comprehensive assessment matrix.

[0029] In another embodiment of the present application, the mathematical form of the conditional risk value method can be: risk measurement function CVaR(X) = E[X|X≥VaR(X)] + κ·σ(X|X≥VaR(X)); wherein: VaR(X) is the risk value function, VaR(X) =inf{x|P(X≤x)≥α}; E[X|X≥VaR(X)] is the conditional expectation; σ(X|X≥VaR(X)) is the conditional standard deviation; κ is the risk aversion coefficient, κ=ln(1+|Sk(X)|); Sk(X) is the skewness coefficient; X is the loss random variable; α is the confidence level; P is the probability measure. By introducing the risk aversion coefficient adjusted by skewness, the accuracy of risk assessment is improved.

[0030] A full life cycle cost-benefit evaluation system was established. By comprehensively considering multiple dimensions such as initial investment, operation and maintenance costs, replacement costs, sales revenue and environmental benefits, a comprehensive evaluation of the system's economic efficiency was achieved. The conditional risk value method was used to analyze the volatility of costs and benefits, and the impact of uncertain factors such as market price fluctuations and carbon trading on the system's economic efficiency was considered. A complete risk assessment system was formed by calculating confidence intervals and risk assessment results. This method not only evaluates the direct economic benefits of the system, but also considers environmental benefits and carbon emission reduction benefits, achieving multi-dimensional quantification of the system's value. The final comprehensive evaluation matrix includes the evaluation results of the three dimensions of cost, benefit and risk, providing a comprehensive decision-making basis for the optimal configuration of the system, effectively balancing economic and environmental benefits, and improving the scientific nature of investment decisions.

[0031] According to one aspect of the present application, step S5 is specifically: Step S51: construct an economic objective function based on the comprehensive evaluation matrix to obtain economic objective data; construct a technical objective function based on the dynamic response model to obtain technical objective data; construct an environmental objective function in combination with the comprehensive evaluation matrix to generate environmental objective data; combine the economic objective data, technical objective data and environmental objective data to form an objective function set; Step S52: extracting equipment operation boundary conditions from the dynamic response model to obtain boundary constraint data; determining various parameter constraints according to pre-stored optimization parameter data to obtain parameter constraint data; combining the boundary constraint data and parameter constraint data to form a constraint condition set; Step S53, using the gray wolf algorithm to initialize the population position and obtain initial position data; calculating the fitness value according to the objective function set and the constraint condition set to obtain fitness data; updating the population position based on the fitness data to generate updated position data; repeating the execution until the preset termination condition is met, and outputting the optimized configuration plan.

[0032] In another embodiment of the present application, the mathematical form of the gray wolf algorithm can be: position update function X(t+1) =X(t) + C·A·D; where: D is the distance vector, D = |B·Xα - X(t)|; A is the convergence factor, A = 2·a·r1 -a; B is the fluctuation factor, B = 2·r2; a is the linear decreasing parameter, a = 2·(1-t / T) ρ ; C is the adaptive weight, C = exp(-φ·|f(X(t))-f(Xα)| / σ); X(t) is the current position; Xα is the position of α wolf; r1 and r2 are random numbers; t is the current iteration number; T is the maximum iteration number; ρ is the nonlinear index; φ is the adaptive coefficient; f(X) is the objective function; σ is the standard deviation of the objective function. By introducing adaptive weights and nonlinear convergence mechanisms, the convergence and diversity of optimization are improved.

[0033] A multi-objective optimization framework based on the gray wolf algorithm was constructed. By considering the objective functions of the three dimensions of economy, technology and environment at the same time, the global optimization of the system configuration was achieved. The gray wolf algorithm was used for multi-objective optimization. The convergence performance of the algorithm and the quality of the solution were improved through adaptive weight adjustment and local search strategy. In the processing of constraints, a complete set of constraints was established in combination with the dynamic response model to ensure that the optimization results meet the technical requirements and operating boundaries of the system. The optimization algorithm adopts a population evolution strategy, and finally obtains the Pareto optimal solution set by continuously updating individual positions and evaluating fitness. This optimization method can not only effectively handle multi-objective optimization problems, but also ensure the diversity and convergence of solutions, providing multiple options for system configuration.

[0034] According to one aspect of the present application, step S6 is specifically: Step S61: input the optimized configuration scheme into the digital twin simulation platform to obtain simulation sequence data; use the verification data set to perform comparative analysis on the simulation sequence data to obtain comparative analysis data; calculate the error index of the comparative analysis data to generate verification result data; Step S62, extracting key parameters from the optimized configuration scheme to obtain parameter sequence data; performing disturbance analysis on the parameter sequence data to obtain disturbance response data; calculating parameter sensitivity according to the disturbance response data to generate sensitivity data; Step S63, identifying the system performance deviation based on the verification result data to obtain deviation data; determining the parameter adjustment direction according to the sensitivity data to obtain adjustment strategy data; optimizing and adjusting the optimization configuration scheme in combination with the deviation data and the adjustment strategy data to generate a final configuration scheme.

[0035] A verification and optimization framework based on digital twins was established. Through simulation verification and sensitivity analysis, the reliability evaluation and optimization adjustment of the optimized configuration scheme were realized. A virtual test environment was constructed using digital twin technology, and the effectiveness of the optimization scheme was verified through comparative analysis and error evaluation. In the sensitivity analysis, the perturbation analysis method was used to identify key parameters, and the system performance was optimized through parameter adjustment. This verification and optimization method can not only evaluate the reliability of the optimization scheme, but also identify the key parameters and performance bottlenecks of the system, providing a scientific basis for scheme optimization. The final configuration scheme generated has been fully verified and optimized, and has high reliability and practicality.

[0036] According to one aspect of the present application, step S12 is specifically: Step S121, receiving a standardized data set, constructing a data envelope using a cubic spline interpolation method to obtain envelope data; calculating local extreme value points of the envelope data to obtain an extreme value point sequence; constructing upper and lower envelope curves based on the extreme value point sequence to generate envelope curve data; Step S122, using the envelope curve data to calculate the mean envelope to obtain mean function data; subtracting the mean function data from the original signal to obtain difference sequence data; performing IMF condition judgment on the difference sequence data to generate screening criterion data; Step S123, performing an iterative screening process according to the screening criterion data to obtain iterative result data; calculating the standard deviation of the iterative result data to obtain convergence index data; when the convergence index data meets a preset threshold, outputting an eigenmode function sequence; Step S124, performing Hilbert transform on the eigenmode function sequence to obtain analytical signal data; calculating the phase function based on the analytical signal data to obtain phase function data; taking a derivative of the phase function data to obtain an instantaneous frequency to generate frequency characteristic data; In another embodiment of the present application, the mathematical form of the Hilbert transform can be: analytical signal function A(t) = x(t) + j·H(x(t)); wherein: H(x(t)) is the Hilbert transform, H(x(t)) = (1 / π)·P∫(x(τ) / (t-τ))dτ; instantaneous phase φ(t) = arctan(H(x(t)) / x(t)) + π·k(t); instantaneous frequency f(t) = (1 / 2π)·dφ(t) / dt + γ(t)·d 2 φ(t) / dt 2 ; k(t) is the phase correction function, k(t) = floor((dφ(t) / dt) / (2π)); γ(t) is the frequency correction coefficient, γ(t) = α·exp(-β·|d 2 φ(t) / dt 2 |); x(t) is the original signal; t and τ are time variables; P is the Cauchy principal value integral; α and β are adjustment parameters. By introducing the adaptive frequency correction mechanism, the accuracy of instantaneous frequency estimation is improved.

[0037] Step S125, construct a time-frequency distribution matrix by combining the eigenmode function sequence and the frequency characteristic data to obtain time-frequency matrix data; perform energy distribution analysis on the time-frequency matrix data to obtain energy distribution data; combine the time-frequency matrix data and the energy distribution data to form multi-scale component data.

[0038] The adaptive decomposition of complex nonlinear signals is achieved through the empirical mode decomposition algorithm. The data envelope is constructed by cubic spline interpolation. The intrinsic mode function is extracted through an iterative screening process to achieve multi-scale analysis of the signal. In the process of frequency feature extraction, the instantaneous frequency is calculated in combination with the Hilbert transform to construct a complete time-frequency distribution matrix. Compared with traditional Fourier analysis, this method has better adaptability and local feature extraction capabilities, and can effectively handle transient features and local changes in non-stationary signals. Through energy distribution analysis, multi-scale characterization of signal features is achieved, providing a reliable data basis for subsequent feature extraction and pattern recognition. The entire processing process does not require pre-setting of basis functions and has strong adaptability and robustness.

[0039] According to one aspect of the present application, step S13 is specifically: Step S131, receiving multi-scale component data, performing data distribution inspection on each component to obtain distribution characteristic data; using an adaptive segmentation method to time segment the distribution characteristic data to obtain segment sequence data; calculating the mean, standard deviation, skewness and kurtosis of each segment of the segment sequence data to generate statistical characteristic data; In another embodiment of the present application, the mathematical form of the adaptive segmentation method can be: segmentation determination function P(t) = ω1·V(t)+ω2·C(t)+ω3·T(t); wherein: V(t) is the variance ratio, V(t) = var(x[t:t+w]) / var(x[tw:t]); C(t) is the autocorrelation coefficient, C(t) = corr(x[tw:t], x[t:t+w]); T(t) is the trend change rate, T(t) = |(dx / dt)t - (dx / dt)tw| / w; ω1, ω2, ω3 are adaptive weights, ωi = exp(-λi·|dP(t) / dt|); w is the sliding window length; x is the data sequence; t is the time index; λi is the weight adjustment coefficient; var is the variance operator; corr is the correlation coefficient operator. By comprehensively considering multiple statistical features, adaptive segmentation of data sequences is achieved.

[0040] Step S132, normalizing the statistical feature data to obtain normalized feature data; calculating the Pearson correlation coefficient between the features in the normalized feature data to obtain a correlation coefficient matrix; combining the correlation coefficient matrix and the statistical feature data, using a principal component analysis method to generate a feature correlation matrix; Step S133, performing eigenvalue decomposition on the feature correlation matrix to obtain an eigenvalue sequence; calculating the contribution of each feature according to the eigenvalue sequence to obtain feature weight data; performing feature screening based on the feature weight data and a preset threshold to generate a feature vector set; Step S134, using a preset time step to perform sliding window division on the feature vector set to obtain window sequence data; performing Fourier transform on each window in the window sequence data to obtain frequency domain feature data; combining the frequency domain feature data and the feature vector set to generate a time series feature matrix.

[0041] In another embodiment of the present application, the mathematical form of the sliding window can be: feature extraction function F(t) = λ·M(t) + (1-λ)·V(t); wherein: M(t) is the window mean function, M(t) = (1 / L)·Σx(i), i∈[tL,t]; V(t) is the variance function, V(t) = (1 / L)·Σ(x(i)-M(t)) 2 ; λ is the dynamic weight coefficient, λ = 1 / (1+exp(-k·S(t))); S(t) is the data slope, S(t) = (x(t)-x(tL)) / L; L is the window length; x(t) is the time series data; k is the sensitivity parameter; t is the time index; i is the data index in the window. By introducing the data slope to adaptively adjust the weight coefficient, the accuracy of feature extraction is improved.

[0042] A complete feature extraction and dimensionality reduction analysis framework was established. Through multi-level feature extraction and correlation analysis, effective dimensionality reduction and feature selection of high-dimensional data were achieved. The adaptive segmentation method was used to segment the data in time, calculate the statistical features and perform normalization, and perform feature correlation analysis and selection through principal component analysis. This method can not only extract the key statistical features of the data, but also remove redundant features through correlation analysis, thereby improving the representativeness and discrimination ability of the features. In particular, in the feature selection process, automatic feature screening is achieved through feature weights and preset thresholds, which significantly reduces the dimension of the data while retaining important feature information. The final generated time series feature matrix not only retains the time series characteristics of the data, but also achieves dimensionality reduction and compression of the data.

[0043] According to one aspect of the present application, step S22 is specifically: Step S221, receiving a device feature data set, aligning the data using a dynamic time warping algorithm to obtain time-aligned data; using wavelet transform to identify mutation points in the time-aligned data to obtain mutation feature data; dividing the time series into multiple sub-segments based on the mutation feature data to generate a segmented feature sequence.

[0044] In another embodiment of the present application, the mathematical form of the dynamic time warping algorithm can be: distance measurement function DTW(X,Y) = min{Σ(d(xi,yj)·W(k))} / N; wherein: d(xi,yj) is the local distance, d(xi,yj) = |xi-yj|·(1+α·|dxi / dt-dyj / dt|); W(k) is the path weight, W(k) = exp(-β·|ij| / max(M,N)); N is the normalization factor, N = Σ(W(k)); xi, yj are sequences to be aligned; i, j are sequence indexes; k is the path index; M, N are sequence lengths; α is the speed difference weight; β is the time difference weight. By introducing the speed difference term, the accuracy of sequence alignment is improved.

[0045] Step S222, performing trend analysis on each sub-segment of the segmented feature sequence to obtain trend feature data; fitting the trend feature data using an autoregressive moving average model to obtain fitting model data; extracting the slope information of each sub-segment based on the fitting model data to generate slope sequence data.

[0046] Step S223, calculating the performance jump amplitude between adjacent sub-segments of the segmented characteristic sequence to obtain jump amplitude data; combining the slope sequence data and the jump amplitude data to obtain change characteristic data; and performing standardization processing on the change characteristic data to generate performance change data.

[0047] Step S224, construct a performance decay prediction model based on exponential smoothing, input the performance change data into the model, and obtain prediction result data; calculate the confidence interval of the prediction result data to obtain confidence interval data; evaluate the prediction reliability based on the prediction result data and the confidence interval data, and generate reliability evaluation data.

[0048] Step S225, fusing the segmented feature sequence, the performance change data and the reliability assessment data to obtain fused sequence data; performing noise suppression on the fused sequence data using Kalman filtering to obtain filtered result data; and performing time series reconstruction on the filtered result data to generate a performance decay sequence.

[0049] In another embodiment of the present application, the mathematical form of Kalman filtering can be: state estimation function X(t|t) = X(t|t-1) + K(t)·(Z(t) - H·X(t|t-1)); wherein: K(t) is an adaptive gain matrix, K(t) = P(t|t-1)·H ’ ·(H·P(t|t-1)·H ’ + R(t))^(-1); P(t|t-1) is the prediction error covariance, P(t|t-1) = F·P(t-1|t-1)·F ’ + Q(t); R(t) is the adaptive measurement noise covariance, R(t) = λ·R(t-1) + (1-λ)·(v(t)·v(t) ’ ), Q(t) is the adaptive process noise covariance, Q(t) = η·Q(t-1) + (1-η)·(w(t)·w(t) ’ ), v(t) is the innovation sequence, w(t) is the residual sequence, λ and η are the forgetting factors, F is the state transfer matrix, and H is the observation matrix. The accuracy of state estimation is improved by introducing adaptive noise covariance.

[0050] An accurate performance degradation analysis and prediction framework was constructed. Through multi-level data processing and modeling analysis, accurate characterization and prediction of equipment performance changes were achieved. The dynamic time warping algorithm was used to align the data, and the mutation points were identified and segmented through wavelet transform to establish a performance degradation prediction model based on exponential smoothing. This method can accurately capture the gradual and sudden changes in equipment performance, and comprehensively reflect the performance change rules of the equipment through a combination of trend analysis and jump analysis. In particular, in the process of building the prediction model, Kalman filtering was used for noise suppression, which significantly improved the accuracy and reliability of the prediction. The performance degradation sequence finally generated provides a reliable data basis for subsequent performance prediction.

[0051] According to one aspect of the present application, step S23 is specifically: Step S231, receiving a performance decay sequence, using a sliding time window method to perform sequence segmentation to obtain a training sample set; performing a data stationarity test on the training sample set to obtain stationarity index data; performing differential processing on the training sample set according to the stationarity index data to generate preprocessed sample data.

[0052] Step S232, constructing a network structure of a multi-layer extreme learning machine, setting the number of hidden layer nodes based on the input dimension, and obtaining network structure data; randomly initializing the input weights and biases in the network structure data to obtain initial parameter data; inputting the preprocessed sample data into the network for forward propagation calculation to generate hidden layer output data.

[0053] Step S233, using the ridge regression method to calculate the output weight of the hidden layer output data to obtain weight optimization data; performing error analysis on the weight optimization data to obtain error index data; adjusting the network structure parameters according to the error index data, repeating the optimization process until the preset conditions are met, and outputting a performance prediction model.

[0054] Step S234, input the test data into the performance prediction model to perform performance prediction and obtain initial prediction data; perform reverse differential processing on the initial prediction data to obtain restored prediction data; use an integrated learning method to perform combination optimization on the restored prediction data to generate prediction result data.

[0055] Step S235, calculate the root mean square error and mean absolute percentage error of the prediction result data to obtain accuracy evaluation data; use the Bootstrap resampling method to estimate the confidence interval of the prediction to obtain interval estimation data; combine the accuracy evaluation data and the interval estimation data to form a prediction parameter set.

[0056] A high-precision performance prediction model based on a multi-layer extreme learning machine was constructed. Through scientific data preprocessing and model optimization, accurate prediction of equipment performance was achieved. The sliding time window and differential processing were used to achieve data stabilization. A deep learning model was established through a multi-layer extreme learning machine, and an integrated learning method was used to optimize the prediction results. In terms of network structure design, the output weights were optimized through the ridge regression method to avoid the overfitting problem. In particular, in the prediction accuracy evaluation link, the Bootstrap resampling method was used to estimate the confidence interval, providing uncertainty quantification of the prediction results. This prediction method not only has a fast training speed and good generalization ability, but also can improve the stability and reliability of the prediction through integrated learning. Through a complete set of prediction parameters, a reliable decision-making basis is provided for system optimization, which significantly improves the practicality of the prediction.

[0057] According to one aspect of the present application, step S31 is specifically: Step S311, receiving a performance prediction model, a prediction parameter set and system parameter data, extracting energy storage element parameters in the system, and obtaining storage element data; identifying energy conversion element parameters in the system, and obtaining conversion element data; establishing a connection relationship between the storage element data and the conversion element data, and generating element association data.

[0058] Step S312, constructing energy nodes based on the Bond Graph method, mapping the component-related data to power bonding points to obtain bonding point data; determining the energy flow direction between the bonding point data to obtain energy flow data; establishing a power connection relationship based on the energy flow data to generate an energy transfer model.

[0059] Step S313, perform priority analysis on each bonding point in the energy transfer model to obtain priority data; determine a calculation sequence based on the priority data to obtain sequence relationship data; and generate causal relationship data by combining the sequence relationship data and the action relationship between the bonding points.

[0060] Step S314, extracting generalized displacement variables from the causal relationship data to obtain displacement variable data; identifying generalized momentum variables in the system to obtain momentum variable data; establishing a state variable set in combination with the displacement variable data and momentum variable data to generate state variable data.

[0061] Step S315, constructing a system matrix of the state equation based on the state variable data to obtain the system matrix data; calculating the eigenvalues ​​and eigenvectors of the system matrix data to obtain characteristic analysis data; constructing a complete state space equation according to the characteristic analysis data and the state variable data to generate a dynamic response model.

[0062] Based on the Bond Graph method, the energy transfer model of the system was established. Through a unified modeling framework, the accurate description and analysis of the multi-energy flow system was achieved. The energy storage and conversion elements in the system were identified, the connection relationship between the elements was established, the energy transfer process was described through the power bonding points, and the complete state space equation was established. This modeling method can uniformly handle various forms of energy conversion processes in the system, such as mechanical, electrical, and chemical, and determine the calculation sequence through causal analysis, ensuring the physical feasibility of the model. In particular, in the process of state variable selection, a complete set of state variables was established by extracting generalized displacement and momentum variables, which provided a theoretical basis for the analysis of the dynamic characteristics of the system. The dynamic response model finally generated not only maintains energy conservation, but also reflects the dynamic characteristics of the system.

[0063] According to one aspect of the present application, step S32 is specifically: Step S321, receiving a dynamic response model, performing Taylor expansion on a nonlinear state equation near a preset working point to obtain expansion coefficient data; extracting first-order terms in the expansion coefficient data to obtain linear term data; constructing a Jacobian matrix based on the linear term data to generate a linearized model.

[0064] Step S322, calculate the determinant of the Jacobian matrix in the linearized model to obtain characteristic equation data; use the QR decomposition algorithm to solve the characteristic equation data to obtain characteristic root data; sort and classify the characteristic root data to generate an eigenvalue sequence.

[0065] Step S323, analyzing the real part distribution of the eigenvalue sequence to obtain stability criterion data; calculating the damping ratio corresponding to the eigenvalue sequence to obtain damping characteristic data; combining the stability criterion data and the damping characteristic data to generate stability evaluation data.

[0066] Step S324, calculating the controllability matrix of the linearized model to obtain controllability data; constructing the observability matrix of the linearized model to obtain observability data; evaluating the control characteristics of the system according to the controllability data and observability data to generate control characteristic data.

[0067] Step S325, calculating the stability margin of the system based on the stability evaluation data to obtain stability margin data; analyzing the control characteristic data to determine the robustness index of the system to obtain robustness data; combining the stability margin data and the robustness data to form stability index data.

[0068] A complete framework for system stability analysis and evaluation was established. Through linearization analysis and eigenvalue calculation, an in-depth study of the system's dynamic characteristics was achieved. The local linearization of the model was achieved through Taylor expansion, the characteristic equation was solved using the QR decomposition algorithm, and the stability of the system was evaluated through eigenvalue analysis. In the stability evaluation process, not only the real distribution of the eigenvalues ​​was analyzed, but also the damping characteristics of the system were calculated to comprehensively evaluate the dynamic performance of the system. In particular, in the control characteristic analysis link, the control performance of the system was evaluated by calculating the controllability and observability matrices, providing a theoretical basis for system control. The stability index data finally generated includes stability margin and robustness evaluation, providing reliable constraints for system optimization.

[0069] According to one aspect of the present application, step S43 is specifically: Step S431, receiving cost evaluation data and benefit evaluation data, using the Monte Carlo method to generate a random perturbation sequence to obtain perturbation sample data; applying perturbations to the cost evaluation data and benefit evaluation data to obtain simulated sequence data; using the kernel density estimation method to analyze the probability distribution characteristics of the simulated sequence data to generate distribution characteristic data.

[0070] In another embodiment of the present application, the mathematical form of the Monte Carlo simulation can be: the simulation sample generation function S(t) = μ(t) + σ(t)·Z(t)·exp(γ(t)); where: μ(t) is the conditional mean, μ(t) = Σ(wi·xi(t)); σ(t) is the conditional standard deviation, σ(t) = sqrt(Σ(wi 2 ·vi(t))); γ(t) is the jump intensity, γ(t) = λ·J(t)·I(U<p(t)); wi is the adaptive weight, wi = exp(-κ·|ri(t)|) / Σ(exp(-κ·|ri(t)|)); Z(t) is a standard normal random variable; J(t) is the jump amplitude; I(·) is the indicator function; U is a uniform random variable; p(t) is the jump probability; κ is the weight parameter; xi(t) is the historical data; vi(t) is the historical volatility; ri(t) is the rate of return. By introducing the adaptive weight and the jump mechanism, the authenticity of the simulation is improved.

[0071] Step S432: Calculate the conditional risk value of the simulation sequence data at a preset confidence level to obtain risk measurement data; perform quantile analysis on the risk measurement data to obtain extreme value feature data; combine the distribution feature data and the extreme value feature data to generate risk index data.

[0072] Step S433: Construct a multi-dimensional risk dependence model based on the Copula function, map the risk index data to the joint distribution space to obtain dependence structure data; calculate the correlation coefficient matrix of the dependence structure data to obtain correlation feature data; analyze the risk propagation path based on the correlation feature data to generate risk propagation data.

[0073] In another embodiment of the present application, the Copula function risk dependence model can be: the joint distribution function C(u1,...,un) = exp(-[((-lnu1)^θ +... + (-lnun)^θ)^(1 / θ)]·η(ρ)); where: η(ρ) is the dynamic correlation function, η(ρ) = 1 + α·(1-exp(-β·|ρt - ρt-1|)); ρt is the dynamic correlation coefficient, ρt = (1-λ)·ρt-1 + λ·(εt·εt ’ )); θ is the dependence parameter; ui is the marginal distribution; n is the dimension; α, β are adjustment parameters; λ is the update coefficient; εt is the standardized residual. By introducing the dynamic correlation mechanism, the description accuracy of the risk dependence structure is improved.

[0074] Step S434: Use the ARIMA-GARCH hybrid model to perform time series analysis on the risk indicator data to obtain fluctuation characteristic data; extract the periodicity and trend components in the fluctuation characteristic data to obtain time-varying characteristic data; combine the risk propagation data and the time-varying characteristic data to generate a risk assessment result.

[0075] In another embodiment of the present application, the mathematical form of the ARIMA-GARCH mixed model can be: conditional mean equation μt = Σ(φi·yt-i) + Σ(θj·εt-j) + γ·σt; conditional variance equation σt 2 = ω + Σ(αi·εt-i 2 ) + Σ(βj·σt-j 2 ) + δ·|rt-1|; where: εt is the residual sequence, εt = yt - μt; rt is the rate of return, rt = ln(yt / yt-1); φi is the autoregressive coefficient; θj is the moving average coefficient; γ is the risk premium coefficient; αi and βj are GARCH coefficients; ω is the constant term; δ is the leverage effect coefficient; yt is the observation sequence; t is the time index; i and j are the lag orders. By introducing the risk premium term and leverage effect, the accuracy of volatility prediction is improved.

[0076] Step S435: standardize the cost assessment data, benefit assessment data and risk assessment results to obtain standardized assessment data; construct a hierarchical analysis model to calculate the weight of each assessment indicator to obtain weight coefficient data; perform weighted combination of the standardized assessment data according to the weight coefficient data to generate a comprehensive assessment matrix.

[0077] A multidimensional risk assessment system based on conditional risk value was constructed. Through statistical analysis and model simulation, the accurate quantification and assessment of system operation risks were achieved. The Monte Carlo method was used to generate random disturbance sequences, the Copula function was used to build a multidimensional risk dependency model, and the ARIMA-GARCH mixed model was used for time series analysis. This method can not only evaluate the impact of a single risk factor, but also study the interaction between multiple risk factors through risk propagation analysis. In particular, in the process of risk quantification, the probability characteristics and extreme value characteristics of risks are provided through kernel density estimation and conditional risk value analysis. The final generated comprehensive assessment matrix fully reflects the risk status of the system and provides a reliable risk assessment basis for decision optimization.

[0078] According to one aspect of the present application, step S51 is specifically: Step S511, receiving a comprehensive evaluation matrix, extracting cost and benefit related indicators, and obtaining economic indicator data; constructing a net present value calculation model, converting the economic indicator data into a time series, and obtaining cash flow data; using fuzzy mathematics methods to process uncertain factors in the cash flow data and generate economic target data.

[0079] Step S512, extracting system performance parameters from the dynamic response model to obtain performance index data; establishing a reliability evaluation model for the performance index data to obtain reliability data; and constructing a comprehensive technical evaluation function in combination with the performance index data and reliability data to generate technical target data.

[0080] Step S513, extract the environmental benefit indicators in the comprehensive evaluation matrix to obtain environmental basic data; establish a carbon emission model based on the life cycle assessment method, perform quantitative analysis on the environmental basic data, and obtain carbon emission data; construct an environmental impact assessment function based on the carbon emission data to generate environmental target data.

[0081] Step S514: normalize the economic target data to obtain normalized economic data; perform standardization conversion on the technical target data to obtain standardized technical data; perform dimensionality consistency processing on the environmental target data to generate consistent environmental data.

[0082] Step S515, construct a target weight calculation model based on the entropy weight method, evaluate the importance of the normalized economic data, standardized technical data and consistent environmental data, and obtain weight evaluation data; use the Chebyshev approximation method to construct a compromise function for each target and obtain target combination data; combine the target combination data and the weight evaluation data to form a target function set.

[0083] An objective function system for multi-objective optimization was established. Through comprehensive evaluation of the three dimensions of economy, technology and environment, the scientific quantification of the system optimization objectives was achieved. An economic objective evaluation model based on fuzzy mathematics was constructed, a reliability evaluation model was established to measure technical objectives, and environmental objectives were quantified through life cycle assessment methods. In the process of objective function construction, the entropy weight method was used to determine the weight of each objective, and the compromise optimization of the objectives was achieved through the Chebyshev approximation method. This method not only takes into account the dimensional differences of each objective, but also reflects the relative importance of different objectives through weight calculation. The final generated objective function set not only ensures the comprehensiveness of the optimization objectives, but also achieves an effective balance between multiple objectives.

[0084] According to one aspect of the present application, step S53 is specifically: Step S531, using the Latin hypercube sampling method to generate an initial population in the solution space to obtain candidate position data; performing a boundary constraint check on the candidate position data to obtain feasible position data; screening the feasible position data according to a population diversity evaluation criterion to generate initial position data.

[0085] In another embodiment of the present application, the mathematical form of Latin hypercube sampling can be: sample generation function X(i,j) = F^(-1)((π(i,j)-U(i,j)) / n); wherein: π(i,j) is a random permutation matrix; U(i,j) is a disturbance term, U(i,j) = α·V(i,j)·(1-exp(-β·D(i,j))); D(i,j) is the distance between samples, D(i,j) = |X(i,j)-X(i-1,j)| / σj; F^(-1) is an inverse distribution function; V(i,j) is a uniform random number; n is the number of samples; i is a sample index; j is a variable index; α and β are adjustment parameters; σj is the standard deviation. By introducing an adaptive perturbation mechanism, the representativeness and uniformity of the sample are improved.

[0086] Step S532, substitute the initial position data into the objective function set to calculate the objective function value and obtain the target evaluation data; check whether the initial position data satisfies the constraint condition set to obtain the constraint evaluation data; use the penalty function method to convert the target evaluation data and the constraint evaluation data into fitness to generate fitness data.

[0087] Step S533, sorting the population individuals based on the fitness data to obtain rank sequence data; determining the positions of α wolf, β wolf and δ wolf according to the rank sequence data to obtain the leader position data; calculating the encirclement coefficient of the leader position data and other individuals to generate encirclement feature data.

[0088] Step S534, construct an adaptive weight update mechanism, adjust the convergence factor according to the surrounding feature data, and obtain weight coefficient data; calculate the position update vector in combination with the weight coefficient data and the leader position data to obtain update vector data; perform local search optimization on the update vector data to generate update position data.

[0089] In another embodiment of the present application, the adaptive weight update mechanism may be: weight update function W(t+1) =W(t) + η(t)·▽f(X(t))·M(t); wherein: η(t) is the adaptive learning rate, η(t) = η0 / (1+λ·||▽f(X(t))||); M(t) is the momentum term, M(t) = β·M(t-1) + (1-β)·▽f(X(t)); ▽f(X(t)) is the objective function gradient; W(t) is the weight vector; X(t) is the current solution; η0 is the initial learning rate; λ is the decay coefficient; β is the momentum coefficient; t is the number of iterations. By introducing the adaptive learning rate and momentum mechanism, the efficiency and stability of weight update are improved.

[0090] Step S535, perform difference analysis on the updated position data to obtain difference feature data; determine the degree of population convergence based on the difference feature data to obtain convergence criterion data; when the convergence criterion data meets the preset termination condition, extract the optimal solution from the updated position data to generate an optimized configuration plan.

[0091] An optimization solution framework based on the gray wolf algorithm was constructed, and the efficient solution of multi-objective optimization problems was achieved through the intelligent optimization algorithm. The Latin hypercube sampling method was used to generate the initial population, the convergence factor was adjusted through the adaptive weight update mechanism, and the local search strategy was used to optimize the quality of the solution. In the population update process, the leader position was determined by the hierarchical sequence, and an effective group learning mechanism was established. Especially in terms of convergence control, the search strategy was dynamically adjusted through difference analysis, which not only ensured the convergence of the algorithm but also maintained the diversity of the population. The final generated optimization configuration scheme has good convergence and diversity, providing a reliable optimization scheme for system configuration.

[0092] According to one aspect of the present application, step S62 is specifically: Step S621, receiving the optimized configuration plan, using principal component analysis to identify important parameter dimensions, and obtaining parameter importance data; setting a screening threshold according to the parameter importance data to obtain threshold screening data; extracting a key parameter set based on the threshold screening data to generate parameter sequence data.

[0093] Step S622, construct an orthogonal experimental design scheme, divide the parameter sequence data into parameter levels, and obtain orthogonal design data; generate parameter perturbation combinations according to the orthogonal design data to obtain perturbation scheme data; use interval analysis to determine the perturbation amplitude in the perturbation scheme data to generate perturbation amplitude data.

[0094] Step S623, applying the disturbance amplitude data to the system model to obtain system response data; performing statistical analysis on the system response data to obtain response characteristic data; combining the disturbance scheme data and the response characteristic data to generate disturbance response data.

[0095] Step S624, using the Sobol method to calculate the first-order sensitivity index of the disturbance response data to obtain first-order index data; calculating the global sensitivity index of the disturbance response data to obtain global index data; analyzing the interaction effect in the disturbance response data to generate interaction index data.

[0096] Step S625, normalize the first-order indicator data, the global indicator data and the interactive indicator data to obtain standardized indicator data; construct a comprehensive evaluation model based on the fuzzy hierarchical analysis method to obtain evaluation result data; combine the standardized indicator data and the evaluation result data to form sensitivity data.

[0097] A complete framework for parameter sensitivity analysis was established, and the identification and quantification of key system parameters were achieved through system disturbance analysis and response evaluation. Principal component analysis was used to identify important parameter dimensions, parameter disturbance analysis was performed through orthogonal experimental design, and the Sobol method was used to calculate global sensitivity indicators. In the sensitivity analysis process, not only the first-order effects of parameters were considered, but also the interaction effects between parameters were analyzed, and the impact of parameters on system performance was comprehensively evaluated. In particular, in the evaluation process, the fuzzy hierarchical analysis method was used for comprehensive evaluation, providing quantitative evaluation results of parameter importance. The sensitivity data finally generated provides a scientific basis for system parameter optimization and adjustment, and significantly improves the reliability and practicality of the optimization results.

[0098] like Figure 2 As shown, in one embodiment, the brief process is: The AC power generated by the direct-drive permanent magnet synchronous wind turbine is connected to the DC bus after AC / DC rectification, and the energy storage battery pack and each alkaline electrolyzer are connected to the DC bus after DC / DC control. When the wind turbine input power is excessive, the electrolyzer cannot absorb all the electricity due to the limitation of installed capacity. In addition to powering the electrolyzer, the bus also charges the energy storage battery pack. When the wind power is insufficient, the energy storage battery pack discharges to maintain the minimum power operation of the electrolyzer to ensure that the electrolyzer can be quickly hot-started.

[0099] The economic model of the system is as follows: The average annual life cycle income of the electrolyzer is: C l =X l c l / h; Among them, X l is the installed capacity of the electrolyzer, c lis the lifetime income of each kilowatt electrolyzer installed, and h is the operating years of the wind power hydrogen production system. The average annual lifetime income of lithium batteries is: C b =X b c b / h; where X b is the installed capacity of lithium battery, c b The system's annual average life cycle income is: C all =Cw+C l +C b ; Where Cw is the annual operation and maintenance life cycle income of the direct-drive permanent magnet synchronous fan. The annual sales price of hydrogen produced by the alkaline electrolyzer is: R H2 =S H2 ×Q H2 ; Among them, S H2 is the selling price per kg of hydrogen, Q H2 is the amount of hydrogen produced by the system each year.

[0100] Assuming that all hydrogen sold is used in hydrogen fuel vehicles, the annual environmental benefit generated by the system is: Reb = Rec + Rpe + Rg + Rr; Among them, Rec and Rpe are the equivalent environmental benefits of reduced electricity costs and pollutant emissions from wind power generation compared with hydrogen production from coal-fired power plants, Rg is the annual government subsidy for wind power generation, and Rr is the equivalent environmental benefit of reduced emissions from hydrogen fuel cell vehicles compared with traditional fuel vehicles.

[0101] The environmental benefits of coal-fired hydrogen production are: Rhp=∑ i=1 N (Δmi×μi+Δmi×λi); Among them, N is the type of pollutant emitted, Δmi is the annual emission of the ith pollutant reduced through wind power generation, μi is the equivalent environmental benefit of the ith pollutant, and λi is the environmental penalty order of magnitude of the ith pollutant.

[0102] The calculation formula of the equivalent environmental benefit of hydrogen fuel cell vehicles is: Re = (Ee / μ H2 )×Q H2 ×(μ CO2 +λ CO2 ); Where Ee is the greenhouse gas emissions per kilometer using gasoline as the reference fuel, μ H2 is the hydrogen fuel consumed per kilometer by hydrogen fuel cell vehicles, μ CO2 is the equivalent environmental benefit of carbon dioxide, λ CO2 The environmental penalty for carbon dioxide is of the order of magnitude.

[0103] The annual income of the system is: Ry=R H2+R eb -C all ; With the goal of maximizing the revenue of the wind hydrogen storage system, the capacity configuration optimization process is established through the gray wolf optimization algorithm, taking into account the system cost, hydrogen sales revenue and environmental benefits. Initialize the GWO algorithm parameters, set the wolf group position information to [electrolyzer installed capacity, lithium battery installed capacity], and set the prey information to system profit; in the first iteration, randomly select the dominant wolf; in non-first iterations, select the dominant wolf according to the profit size, determine the coefficient vector through the installed capacity information of the previous iteration, update the installed capacity information of this iteration, and calculate the system profit value; repeat the calculation, and output the optimal installed capacity solution and the highest profit value after the termination condition is met.

[0104] Construct constraints, including: electrolytic cell capacity constraint 0<X l <Pmax; Lithium battery capacity constraint 0<X b <Pmax; where X l and X b are the installed capacity of electrolyzer and lithium battery respectively, and Pmax is the maximum power of wind power input to the system. l <X l ; Among them, P l It is the real-time working power of the electrolytic cell, and its value should be less than the installed capacity of the electrolytic cell.

[0105] Lithium battery power limit 0<P bo <X b , 0<P bil <min[P bo , P bmax ]; where P bo The remaining energy in the lithium battery, its value should be less than the installed capacity of the lithium battery and greater than 0; P bil The energy used by the lithium battery when it is working, and its value should be less than the remaining energy P in the lithium battery bo And the maximum battery charge and discharge power P bmax .

[0106] 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.

[0107] 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. An optimization configuration method for an off-grid wind power hydrogen production system considering the full life cycle benefits, characterized in that: The steps include: Step S1, receiving original wind speed data, equipment operation status data, environmental monitoring data and electricity price data sampled at predetermined time intervals, performing preprocessing, and generating a time series feature matrix; Step S2, receiving the time series feature matrix and pre-stored historical operation data, performing time series analysis, and generating a performance prediction model and a prediction parameter set; Step S3, receiving the performance prediction model, the prediction parameter set and the pre-stored system parameter data, establishing a system dynamic model, and generating a system characteristic matrix; Step S4, receiving the system characteristic matrix and pre-stored cost data, calculating cost evaluation data and benefit evaluation data, and performing risk analysis to generate a comprehensive evaluation matrix; Step S5, receiving the comprehensive evaluation matrix and the pre-stored optimization parameter data, constructing an objective function set and a constraint condition set; using the gray wolf algorithm to solve the objective function set and generate an optimization configuration scheme; Step S6: receiving the optimized configuration scheme and the pre-stored verification data set, performing simulation verification and sensitivity analysis; optimizing and adjusting the configuration scheme according to the analysis results, and generating a final configuration scheme.

2. According to the optimization configuration method of the off-grid wind power hydrogen production system considering the full life cycle benefits according to claim 1, step S1 is specifically: Step S11, receiving original wind speed data, equipment operation status data, environmental monitoring data and electricity price data, identifying abnormal values ​​using a preset multiple standard deviation method, and marking the identified abnormal values ​​as missing values; repairing all missing values ​​using a time series interpolation method to obtain a cleaned data set; normalizing the cleaned data set to obtain a standardized data set; Step S12, inputting the standardized data set into the empirical mode decomposition algorithm, extracting each order of intrinsic mode functions through iterative screening, and obtaining an eigenmode function sequence; calculating the instantaneous frequency of the eigenmode function sequence to obtain frequency characteristic data; combining the eigenmode function sequence and the frequency characteristic data to generate multi-scale component data; Step S13, calculate the mean, standard deviation, skewness and kurtosis of each component in the multi-scale component data to obtain statistical feature data; perform correlation analysis on the statistical feature data to obtain a feature correlation matrix; select important features according to the feature correlation matrix according to a preset threshold value to generate a feature vector set; use a sliding time window with a preset time step to process the feature vector set and output a time series feature matrix.

3. According to the optimization configuration method of the off-grid wind power hydrogen production system considering the full life cycle benefits according to claim 1, step S2 is specifically: Step S21, extracting parameters including the efficiency, temperature, pressure and operation time of the electrolytic cell from the pre-stored historical operation data to obtain electrolytic cell characteristic data; Extracting parameters including capacity, number of cycles, temperature and discharge depth of the energy storage system from the historical operation data to obtain energy storage characteristic data; Combining the electrolytic cell characteristic data and the energy storage characteristic data to form an equipment characteristic data set; Step S22, segmenting the device feature data set into time series to obtain a segmented feature sequence; calculating the performance change rate of each time period of the segmented feature sequence to obtain performance change data; combining the segmented feature sequence and the performance change data to generate a performance decay sequence; Step S23, input the performance decay sequence into a multi-layer extreme learning machine and train to obtain a performance prediction model; use the performance prediction model to predict the equipment performance to obtain prediction result data; calculate the prediction accuracy and confidence interval based on the prediction result data to generate a prediction parameter set.

4. According to the optimization configuration method of the off-grid wind power hydrogen production system considering the full life cycle benefits according to claim 1, step S3 is specifically: Step S31, inputting the performance prediction model, the prediction parameter set and the system parameter data into the Bond Graph modeling algorithm to obtain an energy transfer model; performing causal analysis on the energy transfer model to obtain causal relationship data; Establishing a state equation based on the causal relationship data to generate a dynamic response model; Step S32, linearizing the dynamic response model to obtain a linearized model; Calculating the eigenvalues ​​of the linearized model to obtain an eigenvalue sequence; Analyze system stability according to the characteristic value sequence and generate stability index data; Step S33, calculating the system response time and overshoot based on the dynamic response model to obtain response parameter data; Combining stability index data and response parameter data to obtain performance evaluation data; The performance evaluation data is organized into a matrix form to generate a system characteristic matrix.

5. According to the optimization configuration method of the off-grid wind power hydrogen production system considering the full life cycle benefits according to claim 1, step S4 is specifically: Step S41, extracting the initial investment, operation and maintenance cost and replacement cost of the equipment from the pre-stored cost data to obtain basic cost data; calculating the maintenance cost within the life cycle of the equipment according to the prediction parameter set output in step S2 to obtain maintenance cost data; combining the basic cost data and the maintenance cost data to form cost assessment data; Step S42: Calculate the hydrogen sales revenue based on the pre-stored market price data to obtain sales revenue data; calculate the carbon emission reduction benefit based on the pre-stored environmental benefit data to obtain environmental benefit data; integrate the sales revenue data and the environmental benefit data to form benefit evaluation data; Step S43: Analyze the volatility of cost assessment data and benefit assessment data using the conditional risk value method to obtain risk indicator data; calculate the confidence interval based on the risk indicator data to obtain the risk assessment result; combine the cost assessment data, benefit assessment data and risk assessment result to generate a comprehensive assessment matrix.

6. According to the optimization configuration method of the off-grid wind power hydrogen production system considering the full life cycle benefits according to claim 1, step S5 is specifically: Step S51: construct an economic objective function based on the comprehensive evaluation matrix to obtain economic objective data; construct a technical objective function based on the dynamic response model to obtain technical objective data; construct an environmental objective function in combination with the comprehensive evaluation matrix to generate environmental objective data; combine the economic objective data, technical objective data and environmental objective data to form an objective function set; Step S52: extracting equipment operation boundary conditions from the dynamic response model to obtain boundary constraint data; determining various parameter constraints according to pre-stored optimization parameter data to obtain parameter constraint data; combining the boundary constraint data and parameter constraint data to form a constraint condition set; Step S53, using the grey wolf algorithm to initialize the population position to obtain initial position data; The fitness value is calculated according to the objective function set and the constraint condition set to obtain fitness data; the population position is updated based on the fitness data to generate updated position data; the execution is repeated until a preset termination condition is met, and an optimized configuration scheme is output.

7. According to the optimization configuration method of the off-grid wind power hydrogen production system considering the full life cycle benefits according to claim 1, step S6 is specifically: Step S61: input the optimized configuration scheme into the digital twin simulation platform to obtain simulation sequence data; use the verification data set to perform comparative analysis on the simulation sequence data to obtain comparative analysis data; Calculating the error index of the comparative analysis data to generate verification result data; Step S62: extract key parameters from the optimized configuration scheme to obtain parameter sequence data; Performing disturbance analysis on the parameter sequence data to obtain disturbance response data; Calculating parameter sensitivity according to the disturbance response data to generate sensitivity data; Step S63, identifying the system performance deviation based on the verification result data to obtain deviation data; determining the parameter adjustment direction according to the sensitivity data to obtain adjustment strategy data; optimizing and adjusting the optimization configuration scheme in combination with the deviation data and the adjustment strategy data to generate a final configuration scheme.

8. According to the method for optimizing the configuration of an off-grid wind power hydrogen production system taking into account the full life cycle benefits according to claim 1, step S12 is specifically: Step S121, receiving a standardized data set, constructing a data envelope using a cubic spline interpolation method to obtain envelope data; calculating local extreme value points of the envelope data to obtain an extreme value point sequence; constructing upper and lower envelope curves based on the extreme value point sequence to generate envelope curve data; Step S122, using the envelope curve data to calculate the mean envelope to obtain mean function data; subtracting the mean function data from the original signal to obtain difference sequence data; Performing IMF condition judgment on the difference sequence data to generate screening criterion data; Step S123, performing an iterative screening process according to the screening criterion data to obtain iterative result data; Calculating the standard deviation of the iteration result data to obtain convergence index data; when the convergence index data meets a preset threshold, outputting an eigenmode function sequence; Step S124, performing Hilbert transform on the eigenmode function sequence to obtain analytical signal data; calculating the phase function based on the analytical signal data to obtain phase function data; taking a derivative of the phase function data to obtain an instantaneous frequency to generate frequency characteristic data; Step S125, construct a time-frequency distribution matrix by combining the eigenmode function sequence and the frequency characteristic data to obtain time-frequency matrix data; perform energy distribution analysis on the time-frequency matrix data to obtain energy distribution data; combine the time-frequency matrix data and the energy distribution data to form multi-scale component data.

9. According to the method for optimizing the configuration of an off-grid wind power hydrogen production system taking into account the full life cycle benefits according to claim 1, step S13 is specifically: Step S131, receiving multi-scale component data, performing data distribution test on each component to obtain distribution characteristic data; using an adaptive segmentation method to time segment the distribution characteristic data to obtain segment sequence data; calculating the mean, standard deviation, skewness and kurtosis of each segment of the segment sequence data to generate statistical characteristic data; Step S132, normalizing the statistical feature data to obtain normalized feature data; Calculating the Pearson correlation coefficient between each feature in the normalized feature data to obtain a correlation coefficient matrix; combining the correlation coefficient matrix and the statistical feature data to generate a feature correlation matrix using a principal component analysis method; Step S133, performing eigenvalue decomposition on the feature correlation matrix to obtain an eigenvalue sequence; calculating the contribution of each feature according to the eigenvalue sequence to obtain feature weight data; Perform feature screening based on the feature weight data and a preset threshold to generate a feature vector set; Step S134, using a preset time step to perform sliding window division on the feature vector set to obtain window sequence data; performing Fourier transform on each window in the window sequence data to obtain frequency domain feature data; combining the frequency domain feature data and the feature vector set to generate a time series feature matrix.

Images