Multi-energy coordination control method for off-grid system

Abstract

The invention relates to the technical field of off-grid energy control, and discloses a multi-energy coordination control method for an off-grid system. The method comprises the following steps: firstly, synchronously collecting operation state data of a photovoltaic array, a wind driven generator and an energy storage unit, and constructing a multi-energy operation state data set; establishing a power distribution characteristic space by combining power fluctuation historical data under different load working conditions; constructing and training an energy coordination decision model on the basis, extracting a system stability index, defining an energy coordination control level and establishing a mapping relation between the energy coordination control level and the stability index; acquiring load fluctuation parameters in real time, generating a working condition self-adaptive coordination instruction and determining a target control level; and when the target control level exceeds a preset threshold range, activating a multi-energy redistribution mechanism. According to the method, multi-energy dynamic coordination can be realized, the operation performance of an off-grid system under a complex working condition is improved, and the adaptability of the system to load fluctuation and energy supply change is enhanced.

Description

Technical Field

[0001] This invention relates to the field of off-grid energy control technology, specifically to a multi-energy coordinated control method for off-grid systems. Background Technology

[0002] Against the backdrop of energy structure transformation, off-grid systems are seeing their application scope continuously expand due to their unique advantages in remote areas, isolated locations, and emergency power supply scenarios. Off-grid systems typically integrate renewable energy sources such as photovoltaics and wind power, as well as energy storage units, to achieve self-sufficiency in energy supply. However, the output of photovoltaic arrays is significantly affected by natural conditions such as sunlight intensity and temperature, while the power output of wind turbines is closely related to wind speed variations; both exhibit strong intermittency and volatility. Meanwhile, the load demand of off-grid systems is not constant; load conditions vary significantly across different times and scenarios. For example, residential electricity load curves differ drastically between day and night, and equipment start-ups and shutdowns in industrial production can cause sudden increases and decreases in load. This instability in energy supply, coupled with the dynamic changes in load demand, presents numerous challenges to the operation of off-grid systems. Currently, most off-grid systems employ relatively simple energy coordination and control methods, often using fixed power allocation strategies that are difficult to flexibly adjust based on real-time energy supply and load demand. When photovoltaic or wind power output suddenly decreases, if energy storage units fail to compensate in time, the system is highly susceptible to power imbalance, leading to voltage fluctuations, frequency deviations, and other problems that affect power quality. Conversely, when energy supply is sufficient, the lack of an effective coordination mechanism may result in some energy being wasted, failing to fully leverage the advantages of multi-energy complementarity. Existing control methods rely heavily on single indicators to assess system stability, making it difficult to comprehensively reflect the system's operating status and accurately adjust control strategies based on stability changes. When load fluctuations exceed expectations, the system often resorts to simple power-limiting or shutdown measures, lacking an effective multi-energy redistribution mechanism. This not only reduces system reliability but also limits the application of off-grid systems in complex scenarios. Summary of the Invention

[0003] The purpose of this invention is to provide a multi-energy coordinated control method for off-grid systems to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides a multi-energy coordinated control method for off-grid systems, the method comprising: Simultaneously collect operational status data from photovoltaic arrays, wind turbines, and energy storage units in off-grid systems to construct a multi-energy operational status dataset; Acquire historical power fluctuation data of off-grid systems under different load conditions, and establish a power distribution feature space based on multi-energy operating status dataset and historical power fluctuation data; Based on the power allocation feature space, an energy coordination decision model is constructed and trained to extract system stability indicators; energy coordination control levels are defined, and a mapping relationship between system stability indicators and energy coordination control levels is established. Real-time acquisition of off-grid system load fluctuation parameters; generation of condition-adaptive coordination instructions based on load fluctuation parameters and energy coordination decision model; determination of target control level based on condition-adaptive coordination instructions and the aforementioned mapping relationship. When the target control level exceeds the preset threshold range, the multi-energy redistribution mechanism is activated.

[0005] Preferably, the operating status data includes photovoltaic output power fluctuation characteristics, wind turbine speed change characteristics, and energy storage state of charge change rate; the establishment of the power allocation feature space includes: performing a dynamic time warping algorithm on the multi-energy operating status dataset to generate a power fluctuation feature matrix; performing mode decomposition processing on historical power fluctuation data to extract historical power mode feature sets; and fusing the power fluctuation feature matrix and the historical power mode feature sets to construct the power allocation feature space.

[0006] Preferably, the generation of the power fluctuation feature matrix includes: performing Hilbert-Huang transform on the photovoltaic output power fluctuation characteristics to extract the photovoltaic intrinsic mode function energy distribution; performing phase space reconstruction on the wind turbine speed change characteristics to extract the speed chaotic feature vector; performing variational mode decomposition on the energy storage state of charge rate of change to obtain the state of charge multi-scale components; and constructing the power fluctuation feature matrix based on the photovoltaic intrinsic mode function energy distribution, the speed chaotic feature vector, and the state of charge multi-scale components.

[0007] Preferably, the extraction of photovoltaic intrinsic mode function energy distribution includes: using an empirical mode decomposition algorithm to process the photovoltaic output power fluctuation characteristics and separate multiple intrinsic mode function components; calculating the energy proportion of each intrinsic mode function component in a predetermined frequency band to form a photovoltaic energy distribution vector.

[0008] Preferably, the extraction of the rotational speed chaotic feature vector includes: inheriting the wind turbine rotational speed change features from the power fluctuation feature matrix, reconstructing the rotational speed change phase space; calculating the Lyapunov exponent of adjacent trajectory points in the rotational speed change phase space; and generating the rotational speed chaotic feature vector based on the Lyapunov exponent distribution.

[0009] Preferably, the extraction of historical power mode feature set includes: establishing an equivalent inertial response model for historical power fluctuation data; analyzing the frequency modulation depth coefficient and damping response time constant from the equivalent inertial response model; and combining the frequency modulation depth coefficient and damping response time constant to form a historical power mode feature set.

[0010] Preferably, the fusion of the power fluctuation feature matrix and the historical power mode feature set includes: performing maximum and minimum value normalization on the power fluctuation feature matrix and the historical power mode feature set respectively; superimposing the normalized power fluctuation feature matrix row by row onto the normalized historical power mode feature set to form a multi-source feature fusion matrix; processing the multi-source feature fusion matrix using a recursive feature elimination mechanism; and reducing the dimensionality of the recursive feature elimination result using a kernel principal component analysis algorithm to obtain a power allocation feature space, which contains N power coordination feature dimensions.

[0011] Preferably, the construction and training of the energy coordination decision model includes: constructing a training sample library using N power coordination feature dimensions of the power allocation feature space as input variables and manually labeled system stability indicators as output labels; initializing the energy coordination decision model using a Gaussian process regression algorithm; iteratively updating the hyperparameters of the Gaussian process regression using the training sample library; and outputting the final energy coordination decision model based on the hyperparameter optimization results.

[0012] Preferably, the generation of adaptive coordination instructions for operating conditions includes: real-time monitoring of changes in the network topology of the off-grid system; updating the energy priority coefficients according to the changes in the network topology; injecting the updated energy priority coefficients into the kernel function of the energy coordination decision model; and calculating the current adaptive coordination instructions using the adjusted kernel function.

[0013] Preferably, the activation of the multi-energy redistribution mechanism includes: creating a communication link based on chaotic carrier modulation; sending power limiting commands to the photovoltaic array, issuing speed compensation commands to the wind turbine, and transmitting charge and discharge rate adjustment commands to the energy storage unit through the communication link; and synchronously collecting the execution feedback data of each energy unit and inputting it into the energy coordination decision model for command calibration.

[0014] Compared with the prior art, the beneficial effects of the present invention are: By synchronously collecting operational status data from photovoltaic arrays, wind turbines, and energy storage units, a multi-energy operational status dataset is constructed. This allows for a comprehensive understanding of the real-time status of each energy component within the system, providing detailed foundational information for subsequent coordinated control. Based on this multi-energy operational status dataset and historical power fluctuation data under different load conditions, a power allocation feature space is established. This provides a more realistic reference framework for power allocation, giving energy coordination a more scientific basis. By constructing and training an energy coordination decision-making model, extracting system stability indicators, defining energy coordination control levels, and establishing a mapping relationship between these levels and system stability indicators, the system can more accurately assess its own operating state. This accurate assessment provides a reliable direction for generating subsequent coordination commands, ensuring that control strategies are no longer formulated blindly. By acquiring load fluctuation parameters in real time, combining them with an energy coordination decision model to generate adaptive coordination commands for different operating conditions, and determining the target control level based on mapping relationships, the system can flexibly adjust its control methods according to real-time load changes. This flexibility allows the system to have control strategies that match different load conditions, avoiding the limitations of fixed strategies under complex operating conditions. When the target control level exceeds a preset threshold, a multi-energy redistribution mechanism is activated. This mechanism can alleviate power imbalances by reallocating the output of various energy sources when the system faces significant fluctuations. This mechanism makes the system more resilient to emergencies and reduces adverse consequences caused by power imbalances. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the working principle of the off-grid system multi-energy coordinated control method described in this invention. Figure 2 A flowchart for constructing the power allocation feature space; Figure 3 The flowchart for extracting chaotic feature vectors of rotational speed; Figure 4 A flowchart for multi-source feature fusion and dimensionality reduction; Figure 5 This is a flowchart for training an energy coordination decision-making model. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 This invention provides a multi-energy coordinated control method for off-grid systems, the method comprising: Simultaneously collect operational status data from photovoltaic arrays, wind turbines, and energy storage units in the off-grid system to construct a multi-energy operational status dataset. Photovoltaic array data includes instantaneous output power values ​​and fluctuation trends; wind turbine data covers dynamic speed change curves; and energy storage unit data involves real-time rate of change of state of charge. Data acquisition is completed through a distributed sensor network, with a sampling frequency of no less than 10Hz.

[0018] Obtain historical load power fluctuation data of the off-grid system and combine it with multi-energy operating status datasets to establish a power distribution feature space. The historical data covers typical load abrupt changes, periodic fluctuations, and random disturbances, with a time span of no less than three complete seasonal cycles.

[0019] Based on the power allocation feature space, a Gaussian process regression algorithm is used to construct an energy coordination decision model. The model input is the N-dimensional coordination features extracted from the feature space, and the output is the quantified value of the system stability index. The stability index is generated by weighting the voltage deviation rate, frequency offset, and power deficit rate.

[0020] A five-level energy coordination and control system is defined, with level 1 representing the lowest intervention intensity and level 5 representing the highest emergency control. A mapping relationship between stability indicators and control levels is established using fuzzy logic rules, and the mapping threshold is dynamically adjusted based on the off-grid system capacity.

[0021] Load fluctuation parameters are monitored in real time and input into the energy coordination decision model to generate adaptive coordination commands. When the target control level corresponding to the command exceeds a preset threshold (e.g., level ≥ 4), a multi-energy redistribution mechanism is triggered.

[0022] Example 1: See Figure 2 Operational status data is acquired through a distributed sensor network, with a sampling frequency set to 10Hz. The output power fluctuation characteristics of the photovoltaic array are transmitted to the central processor in real time, including instantaneous power values ​​and the gradient of change between adjacent sampling points. The wind turbine's rotational speed signal is captured by a magnetoelectric encoder, recording the three averages of the rate of change of rotational speed per second. The state of charge (SOC) of the energy storage unit is provided by the battery management system, calculating the percentage change in SOC every 5 seconds. These three sets of data are stored aligned with a unified timestamp, forming a multidimensional time-series dataset.

[0023] The processing of photovoltaic output power fluctuation characteristics follows a specific procedure. The raw power signal is input to the empirical mode decomposition module, which adaptively decomposes it into 6 to 8 intrinsic mode function components. Hilbert spectrum analysis is performed on each component, and the energy integral value of each component is calculated within a preset frequency band of 0.1Hz to 10Hz. After normalization of the energy values ​​in each frequency band, they are arranged in descending order of energy to form an 8-dimensional vector. This vector reflects the frequency domain distribution characteristics of photovoltaic power fluctuations, where low-frequency components reflect the gradual influence of irradiance changes, and high-frequency components correspond to abrupt changes in cloud cover.

[0024] The processing of wind turbine rotational speed characteristics includes a phase space reconstruction operation. The rotational speed change sequence is imported into the reconstruction algorithm, with the embedding dimension fixed at 5 dimensions. The time delay parameter is calculated using a mutual information function: traversing the 1-50ms delay interval, τ=12ms is determined when the mutual information function first drops to 1 / e of its peak value. The reconstructed phase space trajectory point set is imported into the chaos analysis module, and the nearest neighbor of each trajectory point is selected to calculate the trajectory divergence rate. The maximum Lyapunov exponent is solved by least-squares fitting of the logarithmic time series of distances between trajectory points. This process is repeated three times in different time windows, taking data from the load abrupt change period, steady-state operation period, and shutdown transition period, respectively, to generate a feature vector containing three chaotic exponent values.

[0025] The rate of change of state of charge (POC) of energy storage is processed using variational mode decomposition (MODE). The POC signal is input into a constrained variational model with a preset decomposition level of 4. Iterative solutions using the alternating direction multiplier method are obtained, yielding modal components with center frequencies at 0.01Hz, 0.05Hz, 0.2Hz, and 1Hz. Each component characterizes four timescale characteristics: long-term capacity decay, intraday cycling fluctuations, minute-level compensation response, and second-level instantaneous adjustment. The component amplitudes form a 4-dimensional eigenvector, describing the dynamic response capability of the energy storage system.

[0026] Historical load power fluctuation data collection covers a three-year operating cycle, including typical operating conditions during the spring snowmelt season, summer air conditioning load season, and winter heating season. The equivalent inertia response model uses second-order differential equations to represent the dynamics of the mechanical-electrical system. Parameter estimation is optimized using a particle swarm optimization algorithm: 50 sets of parameter combinations for inertia J, damping coefficient D, and stiffness coefficient K are initialized, and the mean square error between the model output and historical data is minimized in 100 iterations. The optimal parameter set yields two physical quantities: the frequency modulation depth coefficient and the damping response time constant. The former is equal to the ratio of inertia to the square of the natural frequency, and the latter is the ratio of twice the inertia value to the damping coefficient. The historical modal eigenvector formed by these two parameters has a clear physical meaning.

[0027] The construction of the power fluctuation feature matrix follows the time alignment principle. The photovoltaic energy distribution vector, wind turbine chaotic feature vector, and energy storage multi-scale components are resampled at second intervals, forming a 15-row × T-column matrix (the row vectors contain 8 dimensions for photovoltaic, 3 dimensions for wind turbine, and 4 dimensions for energy storage). The historical modal feature set undergoes time-scale matching processing, linearly interpolating the originally minute-level sampled frequency modulation coefficients and damping coefficients to the second-level data points to form a 2×T feature matrix.

[0028] Data normalization is performed before feature fusion. Each feature row within the matrix is ​​independently normalized to its maximum and minimum values, mapping the feature values ​​to the [0,1] interval. The row vectors of the normalized feature matrix are then stacked column-wise: the first two rows represent historical modal features, and the last 15 rows represent real-time running features, forming a 17×T dimensional fusion matrix. In the recursive feature elimination stage, a random forest is used to evaluate feature importance. One hundred decision trees are constructed for feature perturbation testing, and the average importance score for each feature in ten-fold cross-validation is calculated. After deleting redundant feature columns with scores below 0.7, the remaining feature subset is input into the kernel principal component analysis module.

[0029] Kernel principal component analysis (KPCA) uses radial basis function kernels to perform nonlinear transformations. The kernel width parameter is automatically set based on the number of remaining features. After eigenvalue decomposition, the top 5 principal components with a cumulative contribution rate ≥ 95% are extracted. The resulting 5-dimensional feature space represents the interaction intensity of different energy types: Dimension 1 primarily reflects the phase relationship between new energy fluctuations and system inertia; Dimension 2 reflects the time-delay characteristics of energy storage response and load fluctuations; Dimension 3 characterizes the energy transfer intensity between different frequency components; Dimension 4 describes the weight of chaotic characteristics on system damping; and Dimension 5 shows the degree of constraint of historical frequency regulation characteristics on current power allocation. This feature space can simultaneously express the coupling relationship between real-time operating status and historical operating condition characteristics.

[0030] In practical applications of off-grid systems, the identification of the frequency characteristics of photovoltaic energy distribution vectors is particularly important. When the proportion of high-frequency components (>5Hz) exceeds 35%, it indicates that rapid cloud movement leads to sudden power changes. The setting of a 12ms delay parameter in phase space reconstruction effectively captures the response hysteresis phenomenon of wind turbine pitch control, and its positive Lyapunov exponent confirms the existence of chaotic characteristics in the speed system. The design of four fixed center frequencies in variational mode decomposition enables clear separation of energy storage response characteristics at different time scales, avoiding mode aliasing effects. The particle swarm optimization algorithm used for historical data fitting can complete parameter optimization within 8 seconds, meeting the requirements of real-time calculation. The key features retained by recursive feature elimination mainly include: the low-frequency energy ratio of photovoltaic components, the median value of the Lyapunov exponent, the second-level fluctuation of the state of charge, the offset of the frequency modulation depth coefficient, and the rate of change of the damping time constant. The five-dimensional feature vector generated by kernel principal component analysis ultimately constitutes the basic input data space for energy coordination decision-making.

[0031] Example 2: See Figure 3The processing of photovoltaic output power fluctuation characteristics begins with empirical mode decomposition (EMD). The original power signal is input into an adaptive decomposition module, which uses mirror continuation to handle signal boundary issues. The decomposition process iteratively performs a screening operation: identifying local maxima and minima of the signal, forming upper and lower envelopes through cubic spline interpolation, calculating the envelope mean, and subtracting the mean from the original signal to obtain candidate components, until the two basic conditions of intrinsic mode functions are met. The final decomposition produces 6 to 8 intrinsic mode function components and residual terms, arranged in descending order of frequency. Each intrinsic mode function component is input into a Hilbert transform unit to calculate instantaneous amplitude and phase information. The instantaneous frequency is obtained through numerical differentiation, establishing a complete time-frequency distribution representation. Within a preset frequency band of 0.1Hz to 10Hz, full-band energy integration is performed on each component. The integration results are normalized so that the sum of the component energies is 1, and an 8-dimensional distribution vector is generated in descending order of frequency. The element distribution characteristics of this vector reveal the essence of photovoltaic power fluctuations: the 0.1-0.5Hz range corresponds to the influence of slow changes in atmospheric transmittance; the 0.5-2Hz band reflects the effect of local cloud movement; and the components above 2Hz characterize the coupling disturbance between the inverter switching frequency and the power grid.

[0032] The wind turbine rotational speed characteristic processing employs phase space reconstruction technology. Continuously acquired rotational speed sequences are truncated into analysis segments of 500 sampling points each using a sliding window, with a window sliding step size of 50 sampling points. The reconstruction process is based on the following mathematical relationships:

[0033] in This represents the measured value of the rotational speed at time t. For the embedding dimension, For delay time parameters, Represents phase space The nearest neighbor point. This formula calculates the trajectory point. Its nearest neighbor The Euclidean distance between them. Distance sequence Import the chaotic feature analysis module and select the option that satisfies the initial distance. The set of nearest neighbor pairs ( (The phase space diameter). The logarithmic distance is fitted using least squares. The slope of the time-varying curve is calculated using a second-order difference method to reduce noise interference. Each fitting step includes 80 consecutive sampling points, and the median slope value is taken as the Lyapunov exponent estimate for that time period. The eigenvector construction follows specific time-series rules: the first exponent is calculated using a 60-second data window before the load abrupt change, reflecting the system's pre-disturbance state; the second exponent is taken from a 20-second window after the start of the load disturbance, characterizing the transient response; the third exponent is calculated based on steady-state recovery data, with a window length of 120 seconds. These three calculations form a three-dimensional chaotic eigenvector. The order of elements reflects the dynamic characteristics of the system throughout the entire process from steady state to disturbance recovery.

[0034] The engineering implementation involves several key technological configurations. The speed sensor employs an incremental photoelectric encoder with a resolution of 1024 pulses per revolution. Signal transmission utilizes shielded twisted-pair cabling and a π-type filter to eliminate electromagnetic interference. The phase space reconstruction module is integrated into the real-time processor, equipped with a dedicated nearest neighbor search algorithm that achieves fast neighbor point matching based on a kd-tree structure. The Lyapunov exponent calculation unit includes a nonlinear fitting coprocessor and incorporates an improved RANSAC algorithm to remove outliers. Time parameter settings adhere to physical constraints: the load mutation detection window uses a 4-second moving average filter, and the steady-state threshold is a speed fluctuation rate that remains below the rated value by 0.2% for 30 consecutive seconds.

[0035] The engineering processing of intrinsic modulus function components employs a hierarchical computational architecture. Signal decomposition is performed on a separate FPGA chip, processing eight component channels in parallel. The Hilbert transform is implemented using an FIR digital filter with a coefficient length of 101st order, and a windowing design is used to reduce spectral leakage. Instantaneous frequency calculation uses a fourth-order center-difference scheme to avoid the abrupt transition problem in phase differentiation processing. The band integrator is designed as a digital bandpass filter bank containing eight sixth-order Butterworth filters, with a transition band attenuation of over 60dB. Energy normalization is performed on the microcontroller unit, with a protection mechanism configured to prevent division-by-zero anomalies.

[0036] The vector generation process follows standardized timing control. Energy integral values ​​for each frequency band are cached in a circular queue, triggering vector updates when new data overwrites older data. Frequency sorting is performed using a high-speed comparator array, achieving microsecond-level sorting latency. The final photovoltaic energy vector is stored in ascending frequency order: elements 1-4 correspond to components below 0.5Hz (quarterly / daily scale fluctuations); elements 5-6 are in the 0.5-2Hz range (minute-level fluctuations); elements 7-8 represent high-frequency components of 2-10Hz (second-level transients). The real-time synchronization mechanism for constructing chaotic feature vectors includes precise timestamp management. Rotation speed data frames are accompanied by a GPS synchronization clock, with errors controlled within ±100μs. The nearest neighbor search window is constrained within ±5 minutes to avoid long-distance pseudo-correlation. The Lyapunov index is divided into three time periods using an event-driven model: load mutation events are triggered by a power change rate threshold (≥10% / second), and steady-state determination is based on the rotation speed standard deviation continuously falling below the rated value by 0.05%. Vector elements are stored in 32-bit floating-point format, with fixed memory addresses allocated to ensure real-time read efficiency.

[0037] An anomaly handling mechanism during implementation ensures system robustness. When abnormal modes occur during photovoltaic signal decomposition (e.g., amplitude abrupt changes exceeding 300%), the system automatically switches to a piecewise stationary decomposition mode. The phase space reconstruction module monitors the singular attractor exponent; if the Lyapunov exponent calculation diverges, it switches to an alternative algorithm: first performing singular value decomposition for dimensionality reduction, then executing local linear fitting. In nearest neighbor search, when the spatial distribution density of candidate points is below a threshold, the search radius is automatically expanded and a weighted average algorithm is enabled. All intermediate calculation results are appended with confidence labels for subsequent modules to perform confidence-weighted processing.

[0038] The system is equipped with a dedicated calibration mode to cope with environmental changes. Every 24 hours, standard test signals are automatically injected: a 0.2Hz / 2Hz dual-frequency sine wave is input to the photovoltaic channel to detect the energy integration accuracy of each frequency band; a Lorenz system chaotic signal is injected into the wind turbine channel to verify the Lyapunov exponent calculation error. Calibration results are saved to a log file, and calculation parameters are dynamically corrected using an adaptive algorithm. A temperature sensor monitors the processor junction temperature in real time; when the temperature exceeds 85℃, the sampling frequency is automatically reduced to 5Hz to ensure the system's long-term continuous operational stability.

[0039] Example 3: See Figure 4 The historical power fluctuation data is processed using an equivalent inertia response model. The raw load power data undergoes a preprocessing stage, including outlier removal and data alignment. Outlier detection employs the sliding interquartile range method, with a window width of 60 sampling points, defining data exceeding 1.5 times the interquartile range as outlier data. Data alignment compensates for missing points using linear interpolation to ensure the continuity of the time series. The preprocessed power data is then input into a second-order differential equation modeling module, which describes the dynamic response characteristics of the mechanical-electrical system.

[0040] The parameter estimation process for the equivalent inertia model employs an improved particle swarm optimization algorithm. During the algorithm initialization phase, 50 sets of parameter combinations are randomly generated. ,in Represents the equivalent inertia parameter. The damping coefficient is... This represents the stiffness coefficient. The parameter search range is set as follows: , , The fitness function is defined as the normalized mean square error between the model output and the actual power data:

[0041] in Let be the model output power at time t. This is the actual measured power. This represents the average measured power. To determine the analysis window length, the optimization process executes 100 iterations. In each iteration, the particle updates its velocity vector based on its individual and swarm optimal positions. An adaptive inertia weight is introduced for velocity updates, initially set to 0.9 and linearly decreasing to 0.4. An early stopping mechanism is triggered when the fitness improvement is less than 0.1% after 10 consecutive iterations. The final output is the optimal parameter set. Used to calculate frequency modulation depth coefficient and damping response time constant ,in This represents the system's natural frequency. These two parameters constitute a 2D historical power mode eigenvector. .

[0042] The data normalization process before feature fusion employs an improved maximum-minimum scaling method. Power fluctuation feature matrix. Each row is processed independently, and the scaling formula is adjusted as follows:

[0043] in and They represent the first The mean and standard deviation of row features. This processing preserves the data distribution characteristics while mapping 99.7% of the values ​​to the [0,1] interval. Historical modal feature set. The same normalization method is used, but the extreme values ​​in the first and last 5% are excluded when calculating the standard deviation to improve robustness. The normalized matrix... and Merge rows to form a 17×T dimensional fusion matrix. .

[0044] The recursive feature elimination stage employs a two-layer evaluation strategy. The first round uses a random forest classifier for initial screening, configuring 100 decision trees of depth 8, and calculating the average importance of features in 10-fold cross-validation. The second round introduces a nonparametric test based on mutual information to calculate the maximum information coefficient between each feature and the target variable. Finally, feature columns that simultaneously satisfy a random forest importance > 0.7 and a maximum information coefficient > 0.5 are retained. This strategy effectively balances the identification of linear and nonlinear relationships in the feature selection process.

[0045] Kernel principal component analysis employs a hybrid kernel function design, combining the advantages of linear kernels and Laplace kernels:

[0046] in Controlling the mixing ratio of kernel functions, The scaling parameter is adaptive. The eigenvalue decomposition process is accelerated using the Lanczos iterative method, calculating only the first 20 eigenvectors. The principal component selection criteria are improved to: cumulative contribution rate ≥ 90% and individual principal component contribution rate > 2%. The final five retained principal components are subjected to Varimax rotation optimization for interpretability, and the rotated component loading matrix is ​​used for interpreting the physical meaning of the feature space.

[0047] The real-time architecture employs a pipelined processing design. The data acquisition layer receives sensor data via industrial Ethernet, with time synchronization accuracy controlled within ±1ms. The preprocessing module is deployed on edge computing nodes, equipped with a double buffering mechanism to ensure data continuity. Parameter optimization tasks are assigned to a GPU-accelerated cluster, with each worker node handling 10 sets of parallel particle swarm optimization. Feature fusion and dimensionality reduction operations are executed on the real-time processor, with memory access patterns optimized to reduce cache misses. Principal component analysis uses a block-based computation strategy, decomposing large matrices into multiple 256×256 sub-matrices for processing.

[0048] A dynamic update mechanism ensures model adaptability. The equivalent inertia parameter is re-estimated hourly, using a sliding window approach to update the historical dataset. The feature selection module performs a complete evaluation daily, adjusting the selection threshold based on the latest data distribution. Kernel function parameters... and Automatic calibration is performed weekly, using representative sample data from the most recent 7 days. The principal component rotation matrix is ​​updated monthly, recalculating the optimal rotation angle based on accumulated data.

[0049] The anomaly handling system incorporates multi-level protection measures. A triple redundancy mechanism is triggered in response to data acquisition anomalies: when the primary channel data becomes invalid, it automatically switches to the backup sensor channel; if the backup channel also fails, LSTM-based predictions are used to fill the gap. The parameter optimization process monitors particle swarm diversity indicators, injecting random perturbation particles when the effective particle count falls below 30%. A divergence detection mechanism is implemented in the feature selection phase, initiating expert review when the difference between two evaluation results exceeds 15%. Principal component analysis performs numerical stability checks, addressing issues where the condition number exceeds a certain threshold. The matrix is ​​automatically enabled with regularization.

[0050] The system maintenance mode includes a complete diagnostic test suite. Monthly full-process verification tests are performed: standard test signals are injected to verify preprocessing functions, including step changes, sinusoidal sweeps, and pulse sequence tests. The parameter optimization module is tested using a benchmark function with known analytical solutions to verify convergence speed and accuracy. Feature selection tests construct artificial datasets to verify the ability to identify true and false features. Dimensionality reduction performance tests use high-dimensional chaotic system data to evaluate the compression effect of principal components on the system's degrees of freedom. All test results generate detailed reports for system health assessment and parameter tuning.

[0051] An environmental adaptation mechanism addresses changes in operating conditions. The temperature compensation module adjusts sampling parameters in real time, automatically reducing the sampling rate by 10% when the ambient temperature exceeds 40°C. The voltage monitoring circuit detects power supply fluctuations and switches to energy-saving operation mode when the voltage drop exceeds 15%. The electromagnetic compatibility design incorporates multi-layered shielding to ensure a bit error rate below [value missing] even under strong interference in industrial environments. The mechanical vibration sensor triggers an anti-resonance algorithm that automatically adjusts the processing timing when vibration at a characteristic frequency is detected.

[0052] The data traceability system records the entire processing chain. Each data packet is appended with a unique timestamp and version marker, recording the complete transformation history from initial data collection to feature space generation. Changes to processing parameters are logged, supporting process replay at any point in time. Quality metrics are calculated and visualized in real time, including data integrity rate, processing latency, and computing resource utilization. The audit trail function records all manual interventions, forming encrypted and secure log files.

[0053] The interface design with other embodiments follows standardized protocols. Photovoltaic feature vector reception uses a high-speed serial interface with a transmission rate ≥1Mbps. Wind turbine chaotic features are transmitted via time-triggered Ethernet, with jitter controlled within ±50μs. Energy storage status information uses a compressed transmission format to reduce communication bandwidth usage. Control command output conforms to the IEC61850 standard and supports both multicast and unicast transmission modes. The status feedback channel is designed for full-duplex communication, simultaneously transmitting execution results and health status information.

[0054] Computational resource management employs a dynamic allocation strategy. In normal operation, 70% of resources are allocated to real-time processing tasks, and 30% is used for background learning. Sudden load spikes trigger resource reallocation, increasing the priority of real-time tasks to 90%. During maintenance periods, 50% of computing resources are released for self-checking and optimization. Energy management adjusts computational intensity based on system power supply status; when powered by battery, a simplified algorithm version is automatically activated.

[0055] A timing synchronization system constructs a precision clock network. The master clock source uses a GPS-disciplined rubidium atomic clock to maintain long-term stability. Regional clock nodes are synchronized via the PTP protocol, with offset compensation accuracy reaching 100ns. The processing unit is equipped with a local temperature-compensated crystal oscillator to maintain short-term stability when the clock signal is interrupted. Key data acquisition triggers hardware timestamps, eliminating uncertainties introduced by the software stack.

[0056] The security protection system implements a defense-in-depth strategy. Data transmission uses the AES-256 encryption algorithm, with the key rotating hourly. Identity authentication is based on two-factor authentication using digital certificates and biometrics. Firmware integrity verification uses the SHA-3 hash algorithm, automatically verifying the signature upon startup. Physical interfaces are equipped with anti-tampering detection circuits; sensitive data is immediately erased upon abnormal startup. Network security monitoring analyzes traffic patterns in real time, automatically isolating affected nodes upon detecting abnormal behavior.

[0057] Example 4: See Figure 5 The training process of the energy coordination decision-making model begins with the construction of a sample library. 500 valid samples are extracted from the power allocation feature space, each containing data across five feature dimensions. Feature values ​​are collected evenly across three typical operating periods: the photovoltaic output ramp-up period (6:00-8:00 AM), the peak load transition period (5:00-7:00 PM), and the wind power-dominated period at night (11:00-1:00 AM). Each sample is labeled with a corresponding system stability index value for that period, which is a weighted composite of three basic parameters: voltage deviation rate (weight 0.4), frequency offset (weight 0.35), and power deficit rate (weight 0.25). All labeling work was completed by an engineering team with over ten years of operational experience, and the labeled data underwent three rounds of cross-validation, with the coefficient of variation controlled within 0.05.

[0058] The kernel structure configured for Gaussian process regression modeling is as follows: The initial kernel function adopts the quadratic exponential form, and the three hyperparameters are set as follows: signal variance σ_f = 1.0, feature scale l = 1.0, and noise variance σ_n = 0.1. The training process consists of three stages: the first stage uses 200 samples for initial hyperparameter tuning, with the number of iterations limited to 50; the second stage incorporates the remaining samples for fine-tuning; and finally, 10% of the samples are retained as a validation set to prevent overfitting. Hyperparameter optimization employs the conjugate gradient algorithm, and the marginal likelihood function is calculated using Cholesky decomposition to improve numerical stability. Convergence is determined when the hyperparameter changes in two consecutive iterations are both less than 0.001, with a typical training cycle of approximately 12 minutes.

[0059] During real-time operation, the network topology monitoring system scans connection status at 100ms intervals. Topology identification is based on adjacency matrix analysis: when a branch impedance change exceeds a set threshold (e.g., ≥15%), it is determined to be a topology change event. The initial energy priority coefficients are set according to system design principles: 0.6 for photovoltaic, 0.3 for wind power, and 0.8 for energy storage. When a specific energy unit failure is detected, the coefficient update rule is automatically triggered. The updated coefficients are immediately injected into the model kernel function, forming a dynamically adjusted decision-making mechanism.

[0060] The adaptive coordination command generation adopts a dual-channel architecture. The main channel performs standard Gaussian process prediction, refreshing the output value every 5 seconds; the auxiliary channel triggers emergency recalculation within 50ms after detecting a topology change. The command value conversion module maps the predicted output to control quantities, including photovoltaic power regulation (unit: kW), wind turbine speed compensation (unit: rpm), and energy storage charge / discharge rate (unit: kW / s). The final output coordination command includes three types of metadata: timestamp, command validity period, and confidence factor, which are shared with the execution unit through a real-time database.

[0061] The following table shows an example of energy priority adjustment during network topology changes:

[0062] Model parameters are maintained using a regular rolling update mechanism. The incremental learning process starts every Monday at 2:00 AM: collecting the latest 168 hours of running data, re-labeling stability metrics, and generating a supplementary training set. Incremental training retains the original kernel structure, only fine-tuning hyperparameters locally. To prevent model drift, a stratified retention strategy is used for historical samples: retaining 30% of historical core samples (covering extreme conditions) + 70% of recent samples. After each incremental update, a consistency check is performed, comparing the output differences between the old and new models on the benchmark dataset, with a maximum allowable deviation threshold of 5%.

[0063] The real-time instruction generation process includes a three-stage verification process. The first stage verifies the range of input feature values ​​(e.g., the value range of each element in the photovoltaic energy vector [0,1]); the second stage analyzes the rationality of the output stability index (theoretical range 0.5-8.0); and the third stage compares the change rate of instructions between adjacent cycles (maximum allowable step change 10%). When any stage fails, a backup strategy is activated: linear extrapolation compensation is used for earlier-stage anomalies, a simplified model is switched for intermediate-stage anomalies, and the instructions from the previous cycle are retained for later-stage anomalies. All anomalies are logged in a detailed diagnostic log, including the anomaly type, occurrence time, and handling method.

[0064] The hardware platform employs a heterogeneous computing architecture. Gaussian process inference is deployed on a dedicated math accelerator card equipped with 128 parallel floating-point units. Topology analysis tasks are assigned to the network processor, supporting real-time parsing of OSPF protocol packets. The energy priority database runs on a highly available storage cluster, achieving active-active data synchronization. The control command output channel is configured with redundancy, with a primary / backup channel switchover time of less than 20ms.

[0065] A timing coordination mechanism ensures the determinism of the processing flow. The system initiates a precise time synchronization protocol, with clock deviations at each processing node less than 1ms. Data acquisition, model inference, and command transmission constitute a three-stage pipeline, with each stage's processing latency controlled within 80ms. Critical path monitoring employs hardware timestamps, and a real-time oscilloscope is deployed in the command generation channel to capture end-to-end latency.

[0066] The fault-tolerant design covers scenarios involving data anomalies and equipment failures. When input data is missing, multiple imputation strategies are triggered: for single-point missing data, linear interpolation of nearest neighbors is used; for more than five consecutive missing points, ARIMA prediction imputation is activated. The model computation unit is equipped with dual DSPs for hot standby, maintaining output continuity during fault switching. A communication interruption protection mechanism automatically activates a locally cached safe instruction set upon detecting an instruction transmission timeout.

[0067] The system maintenance interface supports online diagnostics and parameter adjustment. Maintenance personnel can view the feature input buffer status, model prediction confidence intervals, and historical command records in real time. Debug mode allows the injection of test feature vectors to observe intermediate calculation results without affecting actual control. Parameter adjustments require two-level authorization authentication, and all change operations are recorded in the security audit log.

[0068] The energy efficiency management module dynamically adjusts calculation accuracy. During periods of power shortage, it automatically reduces model computational complexity: feature dimensions are reduced from 5 to 3 core features, and floating-point precision is switched from 64-bit to 32-bit. In battery-powered mode, energy-saving mode is activated: the sampling period is extended to 150% of the standard value, and incremental learning is paused. After power is restored, the system gradually reverts to the standard configuration to avoid abrupt changes in control parameters.

[0069] Example 5: The activation of the multi-energy redistribution mechanism begins during the communication link establishment process. During system initialization, the Lorenz chaotic equation parameters are configured, with initialization variables set to x0=0.1, y0=0, and z0=0. The carrier signal is generated using fourth-order Runge-Kutta numerical integration with a fixed step size of 0.01 seconds. The RF front-end operates at 4.08 GHz, with a dynamic transmit power adjustment range of 10-30 dBm, automatically adjusted based on the receiver signal strength indication. The carrier signal is split into three independent channels via a directional coupler, corresponding to the photovoltaic, wind turbine, and energy storage control channels respectively, with each channel maintaining an isolation of over 60 dB.

[0070] The power limiting command encoding employs phase-off modulation technology. The command value range of ±500kW is quantized into 1024 discrete values, mapped to the phase offset interval [-π / 2, π / 2]. The modulation process is implemented through a voltage-controlled phase shifter: the DC control voltage is linearly correlated with the command value, achieving a voltage resolution of 0.5mV. The carrier signal phase change rate is limited to within 0.1 radians per microsecond to avoid signal distortion. The final modulated signal is fed into the photovoltaic array's dedicated communication interface via a power amplifier, with the interface impedance matching network set to a 50-ohm balanced structure. The modulation depth monitoring circuit detects the sideband power ratio in real time, and automatic gain control is activated when the ratio exceeds a threshold.

[0071] The wind turbine speed compensation command employs a dynamic amplitude modulation scheme. The compensation range of ±50 rpm is converted into a normalized proportional coefficient, and the fundamental carrier amplitude is set to 2Vpp. The voltage-controlled amplifier adjusts the output amplitude according to the command value, with the change slope limited to within 5V per second. A specially designed envelope shaping filter suppresses high-frequency harmonics, with a passband set to 3-6MHz and a stopband attenuation of 80dB. The modulated signal is transmitted to the wind turbine main controller through a current isolator, and the transformer coupling ratio is designed to be 1:1.25. The channel has a built-in amplitude calibration source, and a 1kHz test signal is periodically injected to verify the transmission linearity.

[0072] Energy storage charge / discharge rate commands are transmitted via chaotic keying. The command states are divided into five levels: deep discharge, shallow discharge, sustain, shallow charge, and deep charge. A keying threshold is defined for the Lorenz variable z(t): a high-level pulse is defined when z > 25, with the duration being a function of the command state. A pulse width encoder maps discharge commands to pulse widths of 100-500 μs, charging commands to 50-200 μs, and no pulse is generated during the sustain state. The keying signal drives a high-speed MOSFET switch, with a rise time controlled within 5 ns. The transmission line uses a shielded twisted-pair cable structure, and an RC network is configured at the termination to absorb reflected interference.

[0073] Feedback data collection establishes a synchronous demodulation system. The execution results of each energy unit are uploaded via the same chaotic carrier wave. Photovoltaic feedback includes actual power limiting values, execution delay, and error codes; wind turbine feedback includes speed adjustment, blade angle, and vibration amplitude; energy storage feedback records charge / discharge rates, state of charge, and temperature parameters. The demodulator employs a phase-locked loop (PLL) structure, with a phase-locked time set to 10 carrier cycles. A generalized synchronization algorithm based on the drive response principle uses an adaptive controller to synchronize the response system with the drive system, achieving a synchronization convergence time of no more than 80ms. The data sampling window is set to a moving 20 carrier cycles, and median filtering eliminates transient interference.

[0074] A precise timing chain is constructed using a time control architecture. The carrier generator is locked to a GPS-disciplined clock source, with clock pulse interval jitter less than 200ps. Transmit time slot allocation employs a time-division multiplexing mechanism: 0-5ms for photovoltaic commands, 5-10ms for wind turbine commands, 10-15ms for energy storage commands, and 15-20ms for reserved time slots. The feedback reception window is arranged in the 20-35ms period, with a 15ms protection interval reserved. The real-time end-to-end transmission delay compensation algorithm is based on historical delay statistics: 100 path delay models are established, and the optimal compensation value is automatically matched according to the current signal strength. The timing monitor continuously measures the end-to-end delay, triggering priority scheduling adjustments when limits are exceeded.

[0075] The fault-tolerance mechanism employs multiple security safeguards. Signal integrity verification utilizes a cyclic redundancy check (CRC-16-CCITT) generator polynomial. Transmission interruption handling follows an automatic retransmission rule: after two consecutive failed transmissions, three retries are initiated; if the transmission still fails, a backup frequency is switched. Command validity verification is achieved through a dual-buffer structure: the old command is maintained during the reception of a new command, and execution is switched only after successful verification. A safety interlock design prevents conflicting commands; when a conflict between power limiting and charging acceleration is detected, the energy storage is maintained and an alarm signal is sent.

[0076] The physical layer protection features environmental adaptability. A temperature compensation circuit monitors the RF module junction temperature, automatically reducing transmit power by 1dB for every 10°C increase. A humidity sensor triggers an anti-condensation heater to maintain cavity humidity at 40%-60%. The power monitoring unit activates linear regulation mode when voltage fluctuations exceed ±15%. Electromagnetic compatibility design includes a waveguide filter and ferrite core, ensuring radiated emissions comply with FCC Class B standards. A mechanical vibration sensor detects abnormal frequencies and activates the shock-absorbing base for active control.

[0077] The system integrates online testing capabilities for maintenance and diagnostics. A periodic self-test program executes hourly: first, a pseudo-random test sequence is sent to verify the bit error rate, then a standard triangular wave command is injected to test response linearity. The fault diagnosis tree includes 32 preset scenarios, automatically isolating damaged modules and replanning signal paths. Performance logs record key parameters such as carrier phase error mean square, demodulation signal-to-noise ratio, and command delivery success rate; data is retained for 90 days. The remote maintenance interface supports encrypted connections and is configured with dual authentication and command signing mechanisms.

[0078] The execution unit interface design fully considers compatibility. The photovoltaic controller is compatible with three communication protocols: Modbus RTU, CAN 2.0B, and Ethernet IP interface. The wind turbine interface is equipped with dual-channel analog voltage input and RS485. The energy storage system uses a fiber optic isolated interface to prevent ground loop interference. The terminal blocks are equipped with a keyway structure to prevent mis-mating, and the cable colors strictly adhere to international coding standards: red for photovoltaic control, blue for wind turbine control, and yellow for energy storage control. The connector mating life has undergone 5000 cycles of testing, with contact resistance change not exceeding 5% of the initial value.

[0079] The feedback calibration system implements closed-loop verification. The photovoltaic channel is equipped with a standard power sensor with a measurement accuracy of 0.5 class, automatically comparing the measured value with the command value every 24 hours. The wind turbine feedback loop integrates a high-precision rotary transformer with a 14-bit angular position resolution. The energy storage feedback uses four-wire resistance measurement to eliminate conductor voltage drop errors. The calibration reference source is periodically sent to the metrology institution for traceability, establishing a complete value transfer chain. The calibration process requires no manual intervention; triggered reports are automatically uploaded to the central database.

[0080] System state transitions follow deterministic logic. In standby mode, basic carrier transmission is maintained, but no control commands are transmitted. Upon activation, the first command is transmitted within 10ms. During stable operation, incremental updates are used: only command items with changes exceeding 5% are transmitted. Timeout management employs a triple-judgment system: a hardware watchdog resets every 500ms, message confirmation times out and retransmits after 300ms, and a control failure safety timer switches to backup control mode if there is no response after 400ms. During the sleep transition, controlled power is gradually reduced, with transmit power dropping to microampere-level standby mode within 10 seconds.

[0081] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0082] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-energy coordinated control method for an off-grid system, characterized in that, Includes the following steps: Simultaneously collect operational status data from photovoltaic arrays, wind turbines, and energy storage units in off-grid systems to construct a multi-energy operational status dataset; Acquire historical power fluctuation data of off-grid systems under different load conditions, and establish a power distribution feature space based on multi-energy operating status dataset and historical power fluctuation data; Based on the power allocation feature space, an energy coordination decision model is constructed and trained to extract system stability indicators; energy coordination control levels are defined, and a mapping relationship between system stability indicators and energy coordination control levels is established. Real-time acquisition of off-grid system load fluctuation parameters; generation of condition-adaptive coordination instructions based on load fluctuation parameters and energy coordination decision model; determination of target control level based on condition-adaptive coordination instructions and the aforementioned mapping relationship. When the target control level exceeds the preset threshold range, the multi-energy redistribution mechanism is activated.

2. The off-grid system multi-energy coordinated control method according to claim 1, characterized in that, The operational status data includes photovoltaic output power fluctuation characteristics, wind turbine speed variation characteristics, and energy storage state of charge variation rate; the establishment of the power allocation feature space includes: performing a dynamic time warping algorithm on the multi-energy operational status dataset to generate a power fluctuation feature matrix; performing mode decomposition processing on historical power fluctuation data to extract historical power mode feature sets; and fusing the power fluctuation feature matrix and historical power mode feature sets to construct the power allocation feature space.

3. The off-grid system multi-energy coordinated control method according to claim 2, characterized in that, The generated power fluctuation feature matrix includes: performing Hilbert-Huang transform on the photovoltaic output power fluctuation characteristics to extract the photovoltaic intrinsic mode function energy distribution; performing phase space reconstruction on the wind turbine speed change characteristics to extract the speed chaotic feature vector; performing variational mode decomposition on the energy storage state of charge rate of change to obtain the state of charge multi-scale components; and constructing the power fluctuation feature matrix based on the photovoltaic intrinsic mode function energy distribution, the speed chaotic feature vector, and the state of charge multi-scale components.

4. The off-grid system multi-energy coordinated control method according to claim 3, characterized in that, The extraction of photovoltaic intrinsic mode function energy distribution includes: using empirical mode decomposition algorithm to process photovoltaic output power fluctuation characteristics and separate multiple intrinsic mode function components; calculating the energy proportion of each intrinsic mode function component in a predetermined frequency band to form a photovoltaic energy distribution vector.

5. The off-grid system multi-energy coordinated control method according to claim 3, characterized in that, The extraction of the chaotic feature vector of rotational speed includes: inheriting the wind turbine rotational speed change features from the power fluctuation feature matrix, reconstructing the rotational speed change phase space; calculating the Lyapunov exponent of adjacent trajectory points in the rotational speed change phase space; and generating the chaotic feature vector of rotational speed based on the distribution of the Lyapunov exponent.

6. The off-grid system multi-energy coordinated control method according to claim 2, characterized in that, The extraction of historical power mode feature set includes: establishing an equivalent inertial response model for historical power fluctuation data; analyzing the frequency modulation depth coefficient and damping response time constant from the equivalent inertial response model; and combining the frequency modulation depth coefficient and damping response time constant to form a historical power mode feature set.

7. The off-grid system multi-energy coordinated control method according to claim 2, characterized in that, The fusion of the power fluctuation feature matrix and the historical power mode feature set includes: performing maximum and minimum value normalization on the power fluctuation feature matrix and the historical power mode feature set respectively; superimposing the normalized power fluctuation feature matrix row by row onto the normalized historical power mode feature set to form a multi-source feature fusion matrix; processing the multi-source feature fusion matrix using a recursive feature elimination mechanism; and reducing the dimensionality of the recursive feature elimination result using a kernel principal component analysis algorithm to obtain a power allocation feature space, which contains N power coordination feature dimensions.

8. The off-grid system multi-energy coordinated control method according to claim 7, characterized in that, The construction and training of the energy coordination decision model includes: using N power coordination feature dimensions of the power allocation feature space as input variables and manually labeled system stability indicators as output labels to build a training sample library; initializing the energy coordination decision model using a Gaussian process regression algorithm; iteratively updating the hyperparameters of the Gaussian process regression using the training sample library; and outputting the final energy coordination decision model based on the hyperparameter optimization results.

9. The off-grid system multi-energy coordinated control method according to claim 8, characterized in that, The generated adaptive coordination instructions include: real-time monitoring of changes in the network topology of the off-grid system; updating the energy priority coefficients according to the changes in the network topology; injecting the updated energy priority coefficients into the kernel function of the energy coordination decision model; and calculating the current adaptive coordination instructions using the adjusted kernel function.

10. The off-grid system multi-energy coordinated control method according to claim 1, characterized in that, The activation of the multi-energy redistribution mechanism includes: creating a communication link based on chaotic carrier modulation; sending power limiting commands to the photovoltaic array, issuing speed compensation commands to the wind turbine, and transmitting charge and discharge rate adjustment commands to the energy storage unit through the communication link; and synchronously collecting execution feedback data from each energy unit and inputting it into the energy coordination decision model for command calibration.

Images