Port multi-energy-supply system control method based on automatic inertia response

Abstract

The invention relates to the technical field of port multi-energy-supply systems. The invention provides a port multi-energy-supply system control method based on automatic inertia response. The method comprises the following steps of collecting data and performing data preprocessing; network load interaction mode recognition and feature extraction are carried out through construction of a time-varying network load interaction diagram and multi-level fusion; constructing a response characteristic tensor and time sequence prediction model and performing prediction; a multi-scale feature matrix is obtained through wavelet transformation and optimized, and an optimized control sequence and a response strategy matrix are obtained; and generating a control instruction matrix based on the optimization control sequence and the response strategy matrix, and performing stability analysis and real-time correction. By constructing a complete data processing, analysis and control closed loop, the intelligent control of the port multi-energy-supply system participating in the network load interaction is realized, the technical problem of the port multi-energy-supply system participating in the power grid regulation is solved, and the system can perform inertia response control more intelligently and efficiently.

Description

Technical Field

[0001] The present invention relates to a multi-energy supply system for ports, and particularly to a control method for a multi-energy supply system for ports based on automatic inertia response. Background Art

[0002] As an important hub of the marine economy, the energy demand of ports exhibits the characteristics of high power and large fluctuations. With the acceleration of the electrification and intelligentization processes of ports, as well as the continuous increase in the proportion of green energy, the multi-energy supply system for ports has become increasingly prominent in the power system. By participating in the frequency regulation and peak shaving of the power grid, such a system can not only provide important auxiliary services but also improve its own economic benefits. Especially in the context of the continuous reduction of the inertia of the current power system, using the fast response ability of the multi-energy supply system for ports to participate in the system inertia response is of great significance for maintaining the frequency stability of the power grid.

[0003] At present, the research on the participation of the multi-energy supply system for ports in power grid regulation mainly focuses on the traditional primary frequency regulation and secondary frequency regulation fields. Common control methods include classical control based on PID (Proportional-Integral-Derivative), model predictive control, and adaptive control, etc. These methods usually adopt a single data processing model and achieve control objectives through simple feature extraction and state estimation. In terms of data analysis, it mainly relies on basic statistical methods and traditional machine learning algorithms such as regression analysis and support vector machines.

[0004] However, these solutions have deficiencies in multiple aspects when dealing with the inertia response problem of the multi-energy supply system for ports: First, the extraction of the network-load interaction characteristics mostly uses static graph structures or simple dynamic graph models, making it difficult to accurately capture the time-varying characteristics of the system; Second, the existing network-load interaction modeling methods fail to fully consider the influence of time decay, resulting in long-term dependence problems in long-sequence modeling; Third, in the process of feature fusion, there is a lack of an effective integration mechanism for different-level features, affecting the accurate expression of the system state; Finally, the optimization algorithm has low computational efficiency when dealing with large-scale constrained optimization problems and is difficult to meet the requirements of real-time control. These technical problems seriously restrict the effect of the multi-energy supply system for ports participating in the power grid inertia response. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides a control method for a multi-energy supply system for ports based on automatic inertia response.

[0006] The technical solution adopted by the present invention is as follows: A control method for a port multi-energy supply system based on automatic inertia response, comprising the following steps: S1. Collect time series data of the port multi-energy supply system, construct a synchronous time series tensor based on the time series data, and perform dynamic characteristic extraction according to the synchronous time series tensor to obtain a standardized system characteristic matrix and a dynamic response ability index, wherein the time series data includes system real-time power data, port equipment operation status data, historical operation data, and feedback data; S2. Construct a time-varying network-load interaction graph based on the standardized system characteristic matrix, the dynamic response ability index, pre-stored power grid operation parameters, and load fluctuation data, perform feature extraction on the time-varying network-load interaction graph to generate a spatio-temporal feature matrix, use the spatio-temporal feature matrix for pattern recognition to obtain a network-load interaction feature vector and a key pattern matrix, and obtain a fusion feature vector and a feature mapping matrix through multi-level fusion; S3. Construct a response characteristic tensor based on the dynamic response ability index, the network-load interaction feature vector, and the fusion feature vector, construct a time series prediction model based on the response characteristic tensor, pre-stored system operation constraints, and historical response data and perform prediction to obtain an inertia response ability matrix; S4. Based on the inertia response ability matrix and the key pattern matrix, obtain a multi-scale feature matrix through wavelet transform and perform optimization to obtain an optimized control sequence and a response strategy matrix; S5. Generate a control instruction matrix for the system based on the optimized control sequence and the response strategy matrix, and perform stability analysis and real-time correction on the control instruction matrix to obtain a corrected control instruction sequence and execution feedback data.

[0007] According to one aspect of the present application, step S1 includes: collecting time series data of the port multi-energy supply system and constructing a synchronous time series tensor based on the time series data; calculating a data quality evaluation matrix based on the synchronous time series tensor; using the data quality evaluation matrix to perform anomaly detection and correction on the synchronous time series tensor to obtain a cleaned data tensor; and performing dynamic characteristic extraction on the cleaned data tensor by using a sliding window method to obtain a standardized system characteristic matrix and a dynamic response ability index.

[0008] According to one aspect of the present application, the steps of anomaly detection and correction specifically include: obtaining the synchronous time series tensor and the data quality evaluation matrix, calculating the first-order difference and the second-order difference values of the time series to generate a time series feature vector; using the time series feature vector to perform a sliding window statistical operation to generate a time domain anomaly score; performing multi-scale wavelet decomposition on the synchronous time series tensor, calculating the energy distribution characteristics of the decomposition coefficients to generate a frequency domain anomaly score; calculating the local density distribution and the distance distribution of sample points based on the synchronous time series tensor to generate a spatial domain anomaly score; inputting the time domain anomaly score, the frequency domain anomaly score, and the spatial domain anomaly score into an evidence theory fusion framework to calculate a comprehensive confidence level and generate an anomaly comprehensive score; and using the anomaly comprehensive score to correct the abnormal data in the synchronous time series tensor to generate a cleaned data tensor.

[0009] According to one aspect of the present application, step S2 specifically includes: constructing a time-varying weight function based on a standardized system characteristic matrix, a dynamic response ability index, power grid operation parameters, and load fluctuation data; calculating the edge weights between network nodes based on the time-varying weight function; constructing a time-varying network-load interaction graph using the edge weights; and calculating an interaction intensity matrix based on the time-varying network-load interaction graph; constructing a spatio-temporal attention model based on the time-varying network-load interaction graph and the interaction intensity matrix; using the spatio-temporal attention model to extract features from the time-varying network-load interaction graph to generate a spatio-temporal feature matrix and an attention weight matrix; obtaining the spatio-temporal feature matrix and the attention weight matrix, and constructing a dynamic density clustering model; using the dynamic density clustering model to perform clustering analysis on the spatio-temporal feature matrix, tracking the time evolution characteristics of the clustering results, and obtaining a network-load interaction feature vector, a key pattern matrix, and a pattern evolution matrix; constructing a three-layer feature fusion network based on the network-load interaction feature vector, the key pattern matrix, and the pattern evolution matrix, performing multi-level fusion, and obtaining a fusion feature vector and a feature mapping matrix.

[0010] According to one aspect of the present application, the steps of generating the spatio-temporal feature matrix and the attention weight matrix specifically include: constructing a multi-head attention calculation unit including a time decay factor based on the time-varying network-load interaction graph and the interaction intensity matrix to generate a decay attention matrix; performing graph convolution operations using the decay attention matrix to extract local features of nodes and generate a graph feature matrix; inputting the graph feature matrix into a bidirectional gated recurrent unit to extract time series dependence features and generate a time series feature vector; and performing feature fusion operations on the graph feature matrix and the time series feature vector to generate the spatio-temporal feature matrix and the attention weight matrix.

[0011] According to one aspect of the present application, the steps of multi-level fusion specifically include: obtaining the network-load interaction feature vector, constructing an autoencoder network, performing feature dimensionality reduction to obtain compressed features, reconstructing features to obtain reconstructed features, calculating a reconstruction error, and outputting low-level fusion features; obtaining the key pattern matrix and the low-level fusion features, constructing a dynamic convolution kernel, performing multi-scale convolution operations, performing pooling processing to obtain an intermediate feature map, extracting significant features, and outputting middle-level fusion features; obtaining the pattern evolution matrix and the middle-level fusion features, constructing an evolution feature extractor, calculating the state transition probability of a time series, extracting evolution pattern features, and fusing static and dynamic features to output high-level fusion features; obtaining the low-level fusion features, the middle-level fusion features, and the high-level fusion features, constructing a multi-layer optimization objective function, calculating the loss function values of each layer of features, performing gradient descent optimization, updating the fusion weights, and outputting the fusion feature vector and the feature mapping matrix.

[0012] According to one aspect of the present application, step S3 includes: obtaining dynamic response ability indicators, network-load interaction feature vectors, and fusion feature vectors, constructing a non-linear response model, calculating response characteristic parameters using the non-linear response model, and constructing a response characteristic tensor based on the response characteristic parameters; constructing a multi-dimensional evaluation function based on the response characteristic tensor and pre-stored system operation constraints, calculating sub-indicators using the multi-dimensional evaluation function, performing weighted calculation on the sub-indicators, and outputting a comprehensive evaluation index matrix; constructing a time series prediction model using the comprehensive evaluation index matrix and pre-stored historical response data, and performing time series prediction based on the time series prediction model to output an inertia response ability matrix.

[0013] According to one aspect of the present application, the steps of constructing the response characteristic tensor specifically include: obtaining dynamic response ability indicators and network-load interaction feature vectors, calculating a hidden layer state transition matrix and initial neuron weights based on a recurrent neural network structure to generate a set of basic model parameters; constructing a piecewise function model using the set of basic model parameters and pre-stored system operation historical data, calculating a forward dead zone threshold and a reverse dead zone threshold to generate a dead zone characteristic matrix; calculating a response amplitude attenuation coefficient and a phase shift parameter based on the set of basic model parameters and the dead zone characteristic matrix according to a non-linear response function to generate a set of initial characteristic parameters; constructing a three-dimensional characteristic mapping framework based on the set of initial characteristic parameters, calculating a time dimension feature mapping value, a frequency dimension transformation coefficient, and a power dimension mapping parameter to generate a response characteristic tensor.

[0014] According to one aspect of the present application, step S4 is specifically: constructing a wavelet transform model, decomposing the inertia response ability matrix and the key mode matrix at different time scales using the wavelet transform model, extracting features from the decomposition results to obtain a multi-scale feature matrix; constructing a multi-level objective function, and performing weight assignment on the multi-level objective function based on the multi-scale feature matrix and pre-stored system constraint conditions; obtaining a set of optimized objective functions according to the multi-level objective function; using a distributed optimization algorithm to solve the multi-objective optimization problem between the set of optimized objective functions and the pre-stored system operation boundary to obtain an optimized control sequence and a response strategy matrix.

[0015] According to one aspect of the present application, step S5 is specifically as follows: Obtain the optimized control sequence and the response strategy matrix, construct the adaptive control law, obtain the control instruction based on the optimized control sequence and the response strategy matrix according to the adaptive control law, and generate the control instruction matrix of the system; Obtain the control instruction matrix and the pre-stored system state vector, construct the Lyapunov function, analyze the system stability according to the control instruction matrix and the system state vector by using the Lyapunov function, and obtain the stability evaluation index; Obtain the control instruction matrix, the stability evaluation index and the pre-stored real-time system state, construct the control instruction correction function, input the control instruction matrix, the stability evaluation index and the real-time system state into the control instruction correction function to perform real-time correction on the control instruction, and obtain the corrected control instruction sequence and the execution feedback data.

[0016] Advantages of the present invention:

[0017] The present invention realizes the intelligent control of the port multi-energy supply system participating in the grid-load interaction by constructing a complete data processing, analysis and control closed-loop. First, through multi-dimensional data preprocessing, the standardized system characteristic matrix and the dynamic response ability index are obtained; Secondly, through the construction and analysis of the time-varying grid-load interaction graph, the dynamic characteristics of the system are accurately characterized; Thirdly, through the construction of the time series prediction model, the future response state of the system is accurately predicted; Finally, through the multi-level optimization algorithm and the real-time correction mechanism, the reliable execution of the control instruction is ensured. The whole solution takes into account the characteristics of the port multi-energy supply system, such as large load fluctuations and high response requirements, and through innovative algorithm design and optimization strategies, significantly improves the grid-load interaction ability of the system and realizes the efficient and reliable inertia response control. Brief Description of the Drawings

[0018] Figure 1 is the flowchart of the control method for the port multi-energy supply system based on automatic inertia response according to the embodiment of the present invention;

[0019] Figure 2 is the flowchart of step S1 according to an embodiment of the present invention;

[0020] Figure 3 is the flowchart of step S2 according to an embodiment of the present invention;

[0021] Figure 4 is the flowchart of step S3 according to an embodiment of the present invention;

[0022] Figure 5 is the flowchart of step S4 according to an embodiment of the present invention;

[0023] Figure 6 is the flowchart of step S5 according to an embodiment of the present invention. Detailed Embodiments

[0024] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0025] As Figure 1 shown, an embodiment of the present invention provides a control method for a port multi-energy supply system based on automatic inertia response, including the following steps:

[0026] S1. Collect the time series data of the port multi-energy supply system, construct a synchronous time series tensor based on the time series data, and perform dynamic characteristic extraction according to the synchronous time series tensor to obtain a standardized system characteristic matrix and a dynamic response ability index. The time series data includes system real-time power data, port equipment operation status data, historical operation data, and feedback data;

[0027] S2. Construct a time-varying network-load interaction graph based on the standardized system characteristic matrix, dynamic response ability index, pre-stored power grid operation parameters, and load fluctuation data, perform feature extraction on the time-varying network-load interaction graph to generate a spatio-temporal feature matrix, use the spatio-temporal feature matrix for pattern recognition to obtain a network-load interaction feature vector and a key pattern matrix, and obtain a fusion feature vector and a feature mapping matrix through multi-level fusion;

[0028] S3. Construct a response characteristic tensor based on the dynamic response ability index, network-load interaction feature vector, and fusion feature vector, construct a time series prediction model based on the response characteristic tensor, pre-stored system operation constraints, and historical response data to obtain an inertia response ability matrix;

[0029] S4. Based on the inertia response ability matrix and the key pattern matrix, obtain a multi-scale feature matrix through wavelet transform and optimize it to obtain an optimized control sequence and a response strategy matrix;

[0030] S5. Generate a control instruction matrix for the system based on the optimized control sequence and the response strategy matrix, and perform stability analysis and real-time correction on the control instruction matrix to obtain a corrected control instruction sequence and execution feedback data.

[0031] By constructing a complete closed-loop of data processing, analysis, and control, the intelligent control of the multi-energy supply system in the port participating in network-load interaction is achieved. First, through multi-dimensional data preprocessing, a standardized system characteristic matrix and dynamic response ability indicators are obtained; second, through the construction and analysis of a time-varying network-load interaction diagram, the dynamic characteristics of the system are accurately characterized; third, through the construction of a time series prediction model, the future response state of the system is accurately predicted; finally, through a multi-level optimization algorithm and real-time correction mechanism, the reliable execution of control instructions is ensured. The entire solution takes into account the characteristics of the multi-energy supply system in the port, such as large load fluctuations and high response requirements. Through innovative algorithm design and optimization strategies, the network-load interaction ability of the system is significantly improved, and the efficient and reliable inertia response control is achieved.

[0032] As Figure 2 shown, according to one aspect of the present application, step S1 is specifically as follows:

[0033] S11. Collect the time series data of the multi-energy supply system in the port and construct a synchronous time series tensor based on the time series data. Specifically, timestamp alignment and serialization processing can be performed on the collected data; the aligned data is constructed into a four-dimensional tensor; time series interpolation and resampling are performed on the four-dimensional tensor to obtain a synchronous time series tensor.

[0034] S12. Calculate the data quality evaluation matrix based on the synchronous time series tensor. Specifically, based on the calculation integrity index, consistency index, and timeliness index of the synchronous time series tensor, a comprehensive data quality score is obtained, and a data quality evaluation matrix is generated according to the comprehensive data quality score.

[0035] S13. Use the data quality evaluation matrix to perform anomaly detection and correction on the synchronous time series tensor to obtain a cleaned data tensor.

[0036] Specifically, the steps of anomaly detection and correction may specifically include steps S131 to S135:

[0037] S131. Obtain the synchronous time series tensor and the data quality evaluation matrix, calculate the first-order difference and second-order difference values of the time series, and generate a time series feature vector; use the time series feature vector to perform a sliding window statistical operation to generate a time domain anomaly score.

[0038] Specifically, the steps of generating the time domain anomaly score specifically include steps S1311 to S1313:

[0039] S1311. Obtain the synchronous time series tensor, calculate the first-order difference value and second-order difference value of adjacent time points, and generate a difference feature matrix.

[0040] S1312. Construct a sliding window using the differential feature matrix, calculate the mean, variance, and skewness within the window, and generate a statistical feature vector.

[0041] S1313. Calculate the Z-score based on the statistical feature vector to generate a time-domain anomaly score.

[0042] Among them, the calculation formula for the time-domain anomaly score TIME_S(x) is:

[0043] TIME_S(x)=DBSCAN_S(x)+TS_features(x)

[0044] Among them, DBSCAN_S(x) is the density clustering score, and TS_features(x) includes time-domain features such as trend, periodicity, and mutation. Specifically, the calculation formula for DBSCAN_S(x) is:

[0045] DBSCAN_S(x)=exp(-ρ(x) / ε)·(1-μ(x))

[0046] Among them, ρ(x) is the local density of sample x, ε is the density threshold parameter, and μ(x) is the clustering membership of sample x.

[0047] S132. Perform multi-scale wavelet decomposition on the synchronous time series tensor, calculate the energy distribution characteristics of the decomposition coefficients, and generate a frequency-domain anomaly score.

[0048] Specifically, the steps for generating the frequency-domain anomaly score can specifically include steps S1321 to S1323:

[0049] S1321. Obtain the synchronous time series tensor, perform six-layer decomposition operations using the db4 wavelet basis, obtain the wavelet coefficients at each scale, and generate a wavelet coefficient matrix.

[0050] S1322. Calculate the energy distribution of each frequency band in the wavelet coefficient matrix to generate an energy feature vector.

[0051] S1323. Use the energy feature vector to calculate the frequency-domain anomaly metric value and generate a frequency-domain anomaly score.

[0052] Among them, the calculation formula for the frequency-domain anomaly score FREQ_S(x) is:

[0053] FREQ_S(x)=W_S(x)+FFT_features(x)

[0054] Among them, W_S(x) is the wavelet coefficient anomaly score, and FFT_features(x) includes frequency-domain features such as spectral features and harmonic analysis.

[0055] Specifically, the calculation formula of W_S(x) is as follows:

[0056] W_S(x)=Σ(ψj·|dj(x)|)

[0057] where ψj is the weight of the j-th layer wavelet coefficient, and dj(x) is the j-th layer wavelet coefficient of the sample x.

[0058] S133. Calculate the local density distribution and distance distribution of sample points based on the synchronous time series tensor to generate the spatial domain anomaly score. Specifically, the DBSCAN algorithm can be used to calculate the density distribution of sample points to generate a density feature matrix; calculate the core point set and boundary point set based on the density feature matrix to generate a clustering result vector; calculate the Local Outlier Factor (LOF) value for each type of sample in the clustering result vector to generate an LOF feature vector; perform weighted fusion on the density feature matrix and the LOF feature vector to generate the spatial domain anomaly score. Among them, the calculation formula of the spatial domain anomaly score SPATIAL_S(x) is:

[0059] SPATIAL_S(x)=LOF_S(x)+SPATIAL_corr(x)

[0060] where LOF_S(x) is the local outlier factor score, and SPATIAL_corr(x) represents the spatial correlation feature. Specifically, the calculation formula of LOF_S(x) is:

[0061] LOF_S(x)=(Σk_dist(x,Nk(x))) / (k·avgDist(Nk(x)))

[0062] where k_dist(x,Nk(x)) is the distance from the sample x to its k nearest neighbors, and avgDist(Nk(x)) is the average distance between the k nearest neighbor samples.

[0063] S134. Input the time domain anomaly score, frequency domain anomaly score, and spatial domain anomaly score into the evidence theory fusion framework to calculate the comprehensive confidence and generate the anomaly comprehensive score. Specifically, the basic probability assignment value of each score can be calculated to generate an evidence weight matrix; perform the Dempster combination rule operation based on the evidence weight matrix to generate a fusion confidence vector; use the fusion confidence vector to calculate the final degree of anomaly to generate the anomaly comprehensive score.

[0064] Among them, in the fusion process, an anomaly detection fusion method can be adopted. Specifically, the anomaly detection result, that is, the anomaly score AS(x), is calculated based on the following formula:

[0065] AS(x) = α·TIME_S(x) + β·FREQ_S(x) + γ·SPATIAL_S(x)

[0066] Wherein, TIME_S(x) is the time-domain anomaly score, FREQ_S(x) is the frequency-domain anomaly score, SPATIAL_S(x) is the spatial-domain anomaly score, and α, β, and γ are the weights of the time-domain anomaly score, frequency-domain anomaly score, and spatial-domain anomaly score, respectively. And α + β + γ = 1.

[0067] By combining the dual anomaly detection mechanisms of DBSCAN density clustering and Local Outlier Factor (LOF), the accuracy and robustness of anomaly detection are significantly improved. Specifically, the DBSCAN algorithm first identifies core samples, boundary samples, and noise points based on density, providing a preliminary classification for different types of data points; then calculates the LOF score for each type of sample separately. This hierarchical calculation strategy avoids the limitations of the traditional LOF algorithm on data with uneven density. By calculating differential features in the time domain, performing wavelet decomposition in the frequency domain, and analyzing the local density distribution in the spatial domain, an all-round anomaly detection framework is constructed. Finally, the detection results in the three dimensions are fused through the Dempster-Shafer evidence theory, which not only considers the reliability of each detector but also improves the robustness of the system through the evidence conflict handling mechanism. This multi-dimensional anomaly detection method can effectively identify and process various anomaly patterns in the port multi-energy supply system, including instantaneous fluctuations, gradual anomalies, and structural anomalies.

[0068] S135. Use the comprehensive anomaly score to correct the anomaly data in the synchronous time series tensor to generate a cleaned data tensor. Specifically, the anomaly data points in the synchronous time series tensor can be identified based on the comprehensive anomaly score; a data reconstruction model is constructed using the locally linear embedding method; the anomaly data points are corrected using the data reconstruction model, and the cleaned data tensor is output.

[0069] S14. Adopt a sliding window method to extract the dynamic characteristics of the cleaned data tensor to obtain a standardized system characteristic matrix and a dynamic response ability index.

[0070] Wherein, the calculation formula of the feature importance score FI(f) is:

[0071] FI(f) = η·GainLB(f) + θ·StabLB(f)

[0072] Where: GainLB(f)=Σ(gi·splitf,i) / N; GainLB(f) is the gain value of feature f in the LightGBM model, gi is the gain value of the i-th split, splitf,i represents the number of times feature f is used in the i-th split, and N is the total number of splits; StabLB(f)=1-σ(gainf) / μ(gainf); StabLB(f) is the stability score of feature f, σ(gainf) is the standard deviation of the gain of feature f, and μ(gainf) is the average gain of feature f; η and θ are weight coefficients and η+θ=1; f is the input feature.

[0073] In a further embodiment, the temporal correlation score may also be considered when calculating the feature importance score FI(f). At this time, the calculation formula for the feature importance score FI(f) is:

[0074] FI(f)=η·GainLB(f)+θ·StabLB(f)+λ·TempCorr(f)

[0075] Where TempCorr(f) represents the temporal correlation score of the feature, and η, θ, and λ are weight coefficients and η+θ+λ=1.

[0076] According to one aspect of the present application, step S14 may specifically be:

[0077] S141. Obtain a preset time window parameter set; calculate the optimal window width based on an adaptive algorithm; construct multiple overlapping sliding windows, and output a multi-scale window sequence and a window parameter matrix.

[0078] S142. Obtain the cleaned data tensor and the multi-scale window sequence, calculate the statistical features within the window, extract the time-frequency features within the window, and construct a feature candidate set; perform feature correlation analysis on the feature candidate set to obtain an original feature matrix.

[0079] S143. Obtain the original feature matrix; construct a feature evaluation model based on LightGBM (Light Gradient Boosting Machine), and calculate the importance score of the original feature for the target variable (the result of the original control instruction of the system participating in the inertia response in the historical operation data); perform feature stability analysis to obtain a feature importance matrix.

[0080] S144. Obtain the original feature matrix and the feature importance matrix; construct a feature weighting model; calculate the dynamic response ability index, and perform response ability normalization processing to obtain a normalized system characteristic matrix and a dynamic response ability index.

[0081] By adopting a multi-scale window sequence processing scheme and dynamically calculating the optimal window width with an adaptive algorithm, the precise capture of features at different time scales is achieved. In particular, statistical features and time-frequency features are calculated simultaneously within multiple sliding windows, and the feature importance scores are calculated by combining the feature evaluation model of LightGBM. This not only enables the identification of key features that have the most influence on the system response ability but also ensures the reliability of the selected features through feature stability analysis. In this step, a dynamic response ability index is calculated through a feature weighting model, enabling the system to accurately evaluate the response characteristics of the port multi-energy supply system under different operating conditions and providing reliable feature support for subsequent optimal control.

[0082] In step S1 of this embodiment, through an innovative method of constructing a synchronous time series tensor, real-time power data, device operating status data, historical operating data, and feedback data are unified and integrated into a four-dimensional tensor space, achieving the spatio-temporal alignment of heterogeneous data sources; by introducing a data quality evaluation matrix to conduct a comprehensive quality assessment of the data, including three dimensions of integrity, consistency, and timeliness, data quality problems can be accurately quantified and located; a triple anomaly detection mechanism in the time domain, frequency domain, and spatial domain is adopted, which can not only identify abrupt anomalies but also discover slowly accumulating anomaly patterns, effectively reducing the false alarm rate; finally, through a dynamic feature extraction method of a sliding window, both the instantaneous characteristics of the system are captured and the historical evolution information is retained, providing a reliable feature basis for subsequent response ability evaluation. This multi-dimensional and multi-level data preprocessing scheme significantly improves the data quality, enabling subsequent analysis and decision-making to be based on high-quality data.

[0083] As Figure 3 shown, according to one aspect of the present application, step S2 is specifically as follows:

[0084] S21. Obtain a standardized system characteristic matrix, pre-stored power grid operation parameters, pre-stored load fluctuation data, and a dynamic response ability index, and construct a time-varying network-load interaction graph. Specifically, a time-varying weight function can be constructed by obtaining a standardized system characteristic matrix, a dynamic response ability index, pre-stored power grid operation parameters, and load fluctuation data, the edge weights between network nodes are calculated based on the time-varying weight function, and a time-varying network-load interaction graph is constructed using the edge weights.

[0085] Among them, the core features of the time-varying network-load interaction graph include: (1) Node types and attributes: Source nodes: represent energy output points such as power generation units and energy storage devices; Load nodes: represent energy consumption points such as load users and electrical equipment; Each node contains dynamic characteristic parameters (such as power, voltage, etc.); (2) Edge attributes: Weight: represents the interaction intensity between nodes; Direction: represents the energy flow direction; Time-varying characteristic: The weight changes dynamically with time; (3) Expression of time-varying characteristics: Edge weight function: w(t)=f(P(t),V(t),I(t)); where P(t), V(t), and I(t) are power, voltage, and current at time t respectively; Node state function: s(t)=g(load characteristics, response ability); (4) Calculation of interaction intensity: Instantaneous interaction intensity: IS(t)=w(t)×s(t); Cumulative interaction intensity: IC=∫IS(t)dt.

[0086] Through the time-varying network-load interaction graph, the following key information can be analyzed: The power transmission path and intensity between the power source and the load; The dynamic coupling relationship between the nodes of the system; The temporal characteristics and patterns of energy flow; The system response ability and stability performance, etc.

[0087] S22. Construct a spatio-temporal attention model based on the time-varying network-load interaction graph and the interaction intensity matrix; Use the spatio-temporal attention model to extract features from the time-varying network-load interaction graph, and generate a spatio-temporal feature matrix and an attention weight matrix.

[0088] According to one aspect of the present application, step S22 can specifically be steps S221~S224:

[0089] S221. Obtain the time-varying network-load interaction graph; Construct a multi-head attention calculation unit; Calculate the query matrix, key matrix, and value matrix; Perform a scaled dot product operation to obtain the original attention scores; Apply the Softmax function to obtain the normalized attention weights, and obtain the initial attention matrix.

[0090] S222. Obtain the time-varying network-load interaction graph and the initial attention matrix; Construct a graph convolutional layer, and calculate the first-order neighborhood features of the nodes in the time-varying network-load interaction graph; Aggregate the neighborhood information to obtain the node embedding vectors; Apply an activation function for processing, and output the graph feature matrix.

[0091] S223. Construct a bidirectional gated recurrent unit, calculate the states of the input gate, forget gate, and output gate based on the graph feature matrix, update the hidden layer state to obtain the temporal feature sequence, and perform sequence normalization processing to obtain the normalized temporal features.

[0092] S224. Construct a feature fusion network, calculate the weight coefficients of the static features and dynamic features based on the graph feature matrix and the normalized temporal features according to the feature fusion network, and perform a weighted fusion operation to obtain the spatio-temporal feature matrix and the attention weight matrix.

[0093] Among them, the spatio-temporal attention model can specifically be as follows:

[0094] The attention score A(Q, K, V) = softmax(Q·K^T / √dk)·V·exp(-λt·Δt); where: Q = WqX is the query matrix, K = WkX is the key matrix, V = WvX is the value matrix, X is the input feature, Wq, Wk, and Wv are the corresponding weight matrices; dk is the feature dimension; λt is the time decay coefficient; Δt is the time interval matrix; exp(-λt·Δt) is the time decay factor.

[0095] Preferably, in step S22, a time decay factor can also be introduced on the basis of the multi-head attention calculation unit to solve the long-term dependence problem in long sequence modeling. Specifically, the steps of generating the spatio-temporal feature matrix and the attention weight matrix (step S22) specifically include: constructing a multi-head attention calculation unit containing a time decay factor based on the time-varying network-load interaction graph and the interaction intensity matrix to generate a decay attention matrix; performing graph convolution operations using the decay attention matrix to extract the local features of the nodes and generate a graph feature matrix; inputting the graph feature matrix into a bidirectional gated recurrent unit to extract the temporal dependence features and generate a temporal feature vector; performing feature fusion operations on the graph feature matrix and the temporal feature vector to generate the spatio-temporal feature matrix and the attention weight matrix. Specifically, it can include the following steps S221’~S224’:

[0096] S221’. Based on the time-varying network-load interaction graph, calculate the time decay factor in the form of exponential decay to generate a time decay vector; construct a multi-head attention calculation unit, calculate the query matrix, the key matrix, and the value matrix to generate an attention basis matrix; perform an element-wise product operation on the time decay vector and the attention basis matrix to generate a weighted attention score; perform Softmax normalization operation on the weighted attention score to generate a decay attention matrix.

[0097] S222’. Based on the time-varying network-load interaction graph and the decay attention matrix, calculate the first-order neighborhood node set of each node in the graph to generate a neighborhood index matrix; perform spatial attention convolution operations using the neighborhood index matrix to extract the local topological features of the nodes and generate a local feature matrix; input the local feature matrix into a residual connection module to fuse the original features and the convolution features to generate a graph feature matrix.

[0098] S223’. Obtain the graph feature matrix, construct a bidirectional recurrent neural network with a gating mechanism, calculate the forward propagation sequence and the backward propagation sequence to generate a bidirectional feature sequence; use the bidirectional feature sequence to construct a temporal memory module, calculate the long-term dependence features and the short-term dependence features to generate a memory feature matrix; perform sequence normalization processing on the memory feature matrix to generate a temporal feature vector.

[0099] S224. Obtain the graph feature matrix and the temporal feature vector, construct a feature fusion network, calculate the importance weights of the static features and the dynamic features, and generate a feature weight vector; use the feature weight vector to perform a weighted fusion operation to generate a spatio-temporal feature matrix; calculate multi-level attention scores based on the spatio-temporal feature matrix to generate an attention weight matrix.

[0100] This method combining graph structure and sequence processing enables the system to simultaneously capture the spatial structure features and the time evolution features of the grid-load interaction, and is particularly suitable for dealing with the complex grid-load interaction patterns in the port multi-energy supply system, improving the system's ability to model long-term dependence relationships.

[0101] S23. Obtain the spatio-temporal feature matrix and the attention weight matrix, and construct a dynamic density clustering model; use the dynamic density clustering model to perform clustering analysis on the spatio-temporal feature matrix, track the time evolution features of the clustering results, and obtain the grid-load interaction feature vector, the key pattern matrix, and the pattern evolution matrix.

[0102] S24. Based on the grid-load interaction feature vector, the key pattern matrix, and the pattern evolution matrix, construct a three-layer feature fusion network to perform multi-level fusion to obtain a fusion feature vector and a feature mapping matrix.

[0103] According to one aspect of the present application, the steps of multi-level fusion specifically include steps S241 to S244:

[0104] S241. Obtain the grid-load interaction feature vector, construct an autoencoder network, perform feature dimensionality reduction to obtain compressed features, reconstruct the features to obtain reconstructed features, calculate the reconstruction error, and output the underlying fusion features;

[0105] S242. Obtain the key pattern matrix and the underlying fusion features, construct a dynamic convolution kernel, perform multi-scale convolution operations, perform pooling processing to obtain an intermediate feature map, extract significant features, and output the middle-level fusion features;

[0106] S243. Obtain the pattern evolution matrix and the middle-level fusion features, construct an evolution feature extractor, calculate the state transition probability of the time series, extract the evolution pattern features, fuse the static and dynamic features, and output the high-level fusion features;

[0107] S244. Obtain the underlying fusion features, the middle-level fusion features, and the high-level fusion features, construct a multi-layer optimization objective function, calculate the loss function values of the features at each layer, perform gradient descent optimization, update the fusion weights, and output the fusion feature vector and the feature mapping matrix.

[0108] Among them, the three-layer feature fusion method can specifically be: the fused feature F = WH·(WM·(WL·FL + bL) + bM) + bH; where: FL = AE(X) is the low-level feature extracted by the autoencoder, WL is the low-level weight matrix, bL is the low-level bias; WM is the middle-level weight matrix, bM is the middle-level bias; WH is the high-level weight matrix, bH is the high-level bias; the autoencoder function AE(X) = Dec(Enc(X)), where X is the input feature, Enc(X) represents the encoding output by the encoder function, and Dec(Enc(X)) represents the reconstructed output of the decoder function for the encoding.

[0109] By constructing a three-layer feature fusion network, effective fusion of different types of features is achieved at the low, middle, and high levels respectively. At the low level, an autoencoder network is used for feature dimensionality reduction and reconstruction, and the quality of feature extraction is optimized through the feedback of the reconstruction error; at the middle level, a dynamic convolution kernel is used to perform multi-scale convolution operations to extract mesoscopic scale pattern features; at the high level, a state transition feature of the time series is captured through an evolutionary feature extractor. This hierarchical feature fusion strategy not only realizes the organic combination of static and dynamic features, but also ensures the accuracy and stability of the fusion result through the gradient descent optimization of the multi-layer optimization objective function.

[0110] In step S2 of this embodiment, through an innovative method of constructing a time-varying network-load interaction graph, dynamic capture of network-load interaction features is achieved. Specifically, the edge weights between network nodes are calculated through a time-varying weight function to construct an interaction intensity matrix reflecting the dynamic characteristics of the system; a spatio-temporal attention model is used to extract features from the time-varying network-load interaction graph, and a spatio-temporal feature matrix is obtained through temporal dependence modeling; a dynamic density clustering model is used to perform clustering analysis on the spatio-temporal feature matrix to achieve accurate identification and tracking of network-load interaction patterns. Especially in the feature fusion network, through the step-by-step fusion of the low, middle, and high levels, multi-level integration of original features, pattern features, and evolutionary features is achieved, which not only improves the richness of feature expression, but also ensures the accuracy of the fusion result through optimizing the objective function, providing reliable feature support for the response ability assessment of the port multi-energy supply system.

[0111] As Figure 4 shown, according to one aspect of the present application, step S3 is specifically:

[0112] S31. Obtain dynamic response ability indicators, network-load interaction feature vectors, and fused feature vectors, construct a non-linear response model, calculate response characteristic parameters using the non-linear response model, and construct a response characteristic tensor based on the response characteristic parameters.

[0113] Among them, the non-linear response model can specifically be: the response output R(xr) = K·xr·(1 - exp(-tr / τ))·φ(xr); where: φ(xr) = 0, When |xr| ≤ δ; φ(xr) = sign(xr)·(|xr| - δ), when δ < |xr| ≤ Δ; φ(xr) = sign(xr)·(Δ - δ), When |xr| > Δ; K is the gain coefficient, τ is the time constant, tr is the response time, δ is the dead zone threshold, Δ is the saturation threshold; xr is the input deviation.

[0114] According to one aspect of the present application, step S31 specifically includes the following steps S311 to S314:

[0115] S311. Obtain the dynamic response ability index and the network-load interaction feature vector; construct a recurrent neural network structure; set the weights of the input layer neurons, and calculate the hidden layer state transition matrix based on the dynamic response ability index and the network-load interaction feature vector according to the recurrent neural network structure, generate the initial model parameters of the non-linear response model, and construct the non-linear response basis function.

[0116] S312. Calculate the response amplitude attenuation coefficient based on the fusion feature vector and the initial model parameters; perform phase shift calculation to generate the frequency modulation parameter, and fuse to obtain the initial characteristic parameter set and the response parameter matrix.

[0117] S313. Obtain the non-linear response basis function and the response parameter matrix; construct a three-dimensional tensor framework, calculate the time dimension feature mapping of the response parameter matrix, perform frequency dimension transformation, generate the power dimension mapping, and output the response characteristic tensor.

[0118] S314. Obtain the initial characteristic parameter set and the response characteristic tensor; construct a parameter optimization objective function; calculate the gradient update direction based on the initial characteristic parameter set and the response characteristic tensor according to the parameter optimization objective function, perform Adam optimization iteration, update the model parameters, and obtain the non-linear parameter set.

[0119] By introducing a dead zone characteristic description function into the non-linear response model, the start-stop characteristics and hysteresis characteristics of the system are accurately characterized. Specifically, the hidden layer state transition matrix is calculated through a recurrent neural network structure, and the long-term dependence relationship is processed by combining bidirectional gated units, and then the boundary characteristics of the dead zone interval are accurately described by a piecewise function model. This step calculates the forward dead zone threshold and the reverse dead zone threshold, establishes a complete dead zone characteristic matrix, and realizes the unified mapping of the time dimension, frequency dimension, and power dimension through a three-dimensional characteristic mapping framework. Finally, through the iterative optimization of the parameter optimization objective function, the modeling accuracy of the system for non-linear characteristics is significantly improved.

[0120] In another aspect of the present application, a dead zone characteristic description function can be introduced into the non-linear response model to accurately characterize the start-stop characteristics and hysteresis characteristics of the system. The hidden layer state transition matrix is calculated through a recurrent neural network structure, and the long-term dependence relationship is processed by combining bidirectional gated units. Then, the boundary characteristics of the dead zone interval are accurately described by a piecewise function model. At this time, the steps of constructing the response characteristic tensor (step S31) may specifically include: obtaining the dynamic response ability index and the network-load interaction feature vector, calculating the hidden layer state transition matrix and the initial neuron weights based on the recurrent neural network structure, and generating a basic model parameter set; using the basic model parameter set and the pre-stored system operation historical data to construct a piecewise function model, calculating the forward dead zone threshold and the reverse dead zone threshold, and generating a dead zone characteristic matrix; calculating the response amplitude attenuation coefficient and the phase shift parameter based on the basic model parameter set and the dead zone characteristic matrix according to the non-linear response function, and generating an initial characteristic parameter set; constructing a three-dimensional characteristic mapping framework based on the initial characteristic parameter set, calculating the time dimension characteristic mapping value, the frequency dimension transformation coefficient, and the power dimension mapping parameter, and generating a response characteristic tensor. Specifically, step S31 may include the following steps S311’ to S314’:

[0121] S311’. Obtain the dynamic response ability index and the network-load interaction feature vector, construct a recurrent neural network with bidirectional gated units, calculate the weight matrix from the input layer to the hidden layer and the weight matrix from the hidden layer to the output layer, and generate a neural network weight set; use the neural network weight set to calculate the hidden layer state transition probability matrix, and generate a state transition feature vector; input the state transition feature vector into the parameter initialization module, calculate the initial neuron threshold and the activation function parameters, and generate a basic model parameter set.

[0122] S312’. Obtain the basic model parameter set and the pre-stored system operation historical data, calculate the forward start threshold and the reverse start threshold of the system response, and generate a start characteristic vector; use the start characteristic vector to construct a piecewise function model, calculate the upper boundary value and the lower boundary value of the dead zone interval, and generate a boundary parameter matrix; calculate the hysteresis characteristic parameters of the dead zone interval based on the boundary parameter matrix, and generate a dead zone characteristic matrix.

[0123] S313’. Based on the basic model parameter set and the dead zone characteristic matrix, calculate the amplitude attenuation characteristic and the phase shift characteristic of the system response, and generate a dynamic characteristic vector; use the dynamic characteristic vector to construct a non-linear response function, calculate the gain coefficient and the time constant of the system, and generate an initial characteristic parameter set.

[0124] S314'. Obtain the initial set of characteristic parameters, construct a three-dimensional characteristic mapping framework, calculate the mapping coefficients of the initial set of characteristic parameters in the time dimension, frequency dimension, and power dimension respectively to generate a dimension mapping matrix; use the dimension mapping matrix and the dead zone characteristic matrix to construct a characteristic tensor, calculate the components of each dimension of the tensor to generate a response characteristic tensor; construct a parameter optimization objective function based on the response characteristic tensor, perform parameter optimization calculation based on the Adam algorithm to generate a non-linear parameter set and a response model parameter matrix.

[0125] In this embodiment, a bidirectional gated unit is introduced to enhance the ability to capture long-term temporal dependencies; dead zone characteristic modeling is added to better conform to the characteristics of the actual system; a piecewise function model is adopted to improve the expression accuracy of non-linear characteristics. The response characteristics are characterized more accurately, including: describing the system startup threshold through the dead zone characteristic matrix; considering the hysteresis characteristic to be closer to the actual system behavior. The piecewise function model simplifies the computational complexity; the Adam algorithm accelerates parameter convergence. In summary, in step S31 of this embodiment, the key characteristics of the actual system are considered; the parameter optimization is more comprehensive, providing better adaptability; and the engineering practicability is stronger.

[0126] S32. Based on the response characteristic tensor and the pre-stored system operation constraints, construct a multi-dimensional evaluation function, use the multi-dimensional evaluation function to calculate sub-item indicators, perform weighted calculation on the sub-item indicators, and output a comprehensive evaluation index matrix.

[0127] S33. Use the comprehensive evaluation index matrix and the pre-stored historical response data to construct a time series prediction model, perform time series prediction based on the time series prediction model, and output an inertia response ability matrix.

[0128] In step S3 of this embodiment, by constructing a non-linear response model and a comprehensive evaluation index system, the accurate evaluation of the system response ability is realized. First, use the non-linear response model to calculate the response characteristic parameters, and comprehensively describe the dynamic response characteristics of the system through the response characteristic tensor; then construct a comprehensive evaluation index matrix based on the multi-dimensional evaluation function to realize the quantitative evaluation of the system response ability; finally, construct a time series prediction model, which can not only accurately predict the future response state of the system, but also capture key time series features through the attention mechanism. This method combining non-linear modeling and time series prediction is particularly suitable for dealing with the complex response characteristics in the port multi-energy supply system and provides accurate evaluation results of the inertia response ability.

[0129] In another embodiment of this application, in step S31, a non-linear parameter set can also be constructed while constructing the response characteristic tensor based on the response characteristic parameters. At this time, steps S32 and S33 can be:

[0130] S32': Based on the response characteristic tensor, the set of non-linear parameters, and the pre-stored system operation constraints, construct a multi-dimensional evaluation function including dead zone characteristics, calculate the time characteristic index, amplitude characteristic index, and frequency characteristic index of the system response, and generate a sub-index matrix; perform weighted fusion operation using the sub-index matrix and the set of non-linear parameters to generate a comprehensive evaluation index matrix.

[0131] S33': Obtain the comprehensive evaluation index matrix, the set of non-linear parameters, and the pre-stored historical response data, construct a temporal attention network, use the set of non-linear parameters as the modulation parameters of the network to generate prediction model parameters; construct a temporal prediction model using the prediction model parameters, perform multi-step prediction operation, and generate an inertia response ability matrix.

[0132] As Figure 5 shown, according to one aspect of the present application, step S4 may specifically be:

[0133] S41: Based on the inertia response ability matrix and the key mode matrix, construct a wavelet transform model, perform decomposition and feature extraction of different regimes through wavelet transform to obtain a multi-scale feature matrix. Specifically, it may include obtaining the inertia response ability matrix and the key mode matrix, constructing an improved wavelet transform model; using the wavelet transform model to decompose the inertia response ability matrix and the key mode matrix at different time scales; performing feature extraction on the decomposition results to obtain a multi-scale feature matrix.

[0134] Among them, the improved wavelet transform model is specifically:

[0135] Wψ(a,b)=(1 / √a)·Σx(t)·ψ*((tw-b) / a)·w(tw)

[0136] Among them, Wψ(a,b) is the wavelet coefficient, ψ((tw-b) / a) is the wavelet basis function, a is the scale parameter, b is the translation parameter; w(tw)=exp(-ξ·|tw-tc|) is the adaptive weight function, tw represents the current moment, tc is the central moment, ξ is the attenuation coefficient; x(t) is the input signal.

[0137] S42: Construct an objective function, and obtain an optimized objective function set based on the multi-scale feature matrix and the pre-stored system constraint conditions according to the objective function. Specifically, it may include obtaining the multi-scale feature matrix and the pre-stored system constraint conditions; constructing a multi-level objective function; performing weight assignment on the multi-level objective function; combining the weighted objective functions; obtaining an optimized objective function set based on the multi-scale feature matrix and the pre-stored system constraint conditions according to the multi-level objective function.

[0138] S43. Construct a distributed optimization algorithm; use the distributed optimization algorithm to solve the multi-objective optimization problem to obtain an optimized control sequence and a response strategy matrix. Specifically, it may include obtaining a set of optimization objective functions and pre-stored system operation boundaries; constructing a distributed optimization algorithm; using the distributed optimization algorithm to solve the multi-objective optimization problem between the set of optimization objective functions and the pre-stored system operation boundaries to obtain an optimized control sequence and a response strategy matrix.

[0139] According to one aspect of the present application, step S43 is specifically as follows:

[0140] S431. Obtain a set of optimization objective functions and pre-stored system operation boundaries; construct a Lagrange multiplier vector; calculate the gradient of the constraint condition, initialize the primal variable and the dual variable, generate an initial solution vector, and output an optimization problem parameter set.

[0141] S432. Obtain the initial solution vector and the optimization problem parameter set; construct an ADMM (Alternating Direction Method of Multipliers) algorithm framework; calculate the augmented Lagrangian function value, perform primal variable update, and output an iterative solution sequence.

[0142] S433. Based on the iterative solution sequence, calculate the change amount of adjacent iterative solutions; perform primal residual calculation to generate a dual residual; judge the convergence condition; and output a convergence status flag when the convergence condition is met.

[0143] S434. Based on the iterative solution sequence and the convergence status flag, extract the optimal primal variable value, calculate the optimal dual variable value, generate a control decision sequence, and output an optimized control sequence and a response strategy matrix.

[0144] The distributed optimization algorithm based on ADMM is adopted to solve large-scale constrained optimization problems. By constructing a Lagrange multiplier vector and calculating the gradient of the constraint condition, the original problem is decomposed into several sub-problems that are easy to solve; by iteratively updating the primal variable and the dual variable, the rapid convergence of the optimization objective is achieved. Especially when calculating the primal residual and the dual residual, a dynamic adjustment strategy is adopted to ensure the stability of the algorithm. Finally, the generated optimized control sequence and response strategy matrix can effectively balance the control effect and the calculation efficiency, and meet the real-time control requirements of the port multi-energy supply system.

[0145] In step S4 of this embodiment, the multi-scale analysis of the system response characteristics is realized through an improved wavelet transform method, and the optimal control strategy is obtained through multi-objective optimization. Specifically, the wavelet transform is performed on the inertia response ability matrix to obtain the characteristic representation on different time scales; by constructing a multi-level objective function and performing weight allocation, a complete set of optimization objective functions is formed; the distributed optimization algorithm is used to solve the multi-objective optimization problem, and the optimal control sequence is obtained on the premise of satisfying the system constraints. This method based on multi-scale analysis and multi-objective optimization can not only adapt to the response requirements of the system on different time scales, but also improve the solution efficiency through the distributed processing of the optimization algorithm, ensuring the real-time and optimality of the control strategy.

[0146] In another embodiment of the present application, steps S41 and S42 may also be:

[0147] S41’. Obtain the inertia response ability matrix, the key mode matrix and the feature mapping matrix, construct a feature transformation function based on the feature mapping matrix, calculate the multi-scale representation of the signal, and generate a transformation coefficient matrix; perform wavelet transform operations using the transformation coefficient matrix to generate a multi-scale feature matrix.

[0148] According to one aspect of the present application, step S41’ may specifically include the following steps S411 to S414:

[0149] S411. Based on the inertia response ability matrix, calculate the signal energy distribution; perform spectrum analysis; construct an optimal wavelet basis selection criterion, and output optimized wavelet parameters.

[0150] S412. Obtain the key mode matrix and the optimized wavelet parameters, construct a multi-level wavelet decomposition framework; perform high-frequency and low-frequency separation, calculate the coefficients of each scale, generate a decomposition coefficient matrix, and output multi-scale components.

[0151] S413. Obtain the multi-scale components, calculate the statistical characteristics of each scale; perform energy aggregation analysis, extract scale correlation characteristics, construct a feature description vector, and output a scale feature matrix.

[0152] S414. Based on the decomposition coefficient matrix and the scale feature matrix, construct a signal reconstruction framework; calculate the reconstruction weight coefficients, perform inverse wavelet transform, fuse the multi-scale features, and output a multi-scale feature matrix.

[0153] By constructing an adaptive wavelet transform framework, the multi-scale analysis of the inertia response ability is realized. Specifically, the optimal wavelet basis function is selected by calculating the signal energy distribution, and a multi-layer wavelet decomposition framework is constructed to achieve the precise separation of high-frequency and low-frequency components. At the same time, through the calculation of statistical features and energy aggregation degree analysis at each scale, the feature descriptions of different frequency bands are obtained. This method calculates the reconstruction weight coefficients through the signal reconstruction framework, extracts multi-scale features while maintaining the integrity of the signal, and is particularly suitable for processing the power fluctuation characteristics of different time scales in the port multi-energy supply system, providing comprehensive feature support for subsequent optimal control.

[0154] S42': Based on the multi-scale feature matrix, the feature mapping matrix, and the pre-stored system constraint conditions, construct a feature mapping function, calculate the importance weights of features at different scales, and generate a feature weight vector; construct a multi-level objective function based on the feature weight vector to generate a set of optimization objective functions.

[0155] In this embodiment, by introducing the feature mapping matrix as the core component, the adaptive mapping from the feature space to the target space is realized, which has significant advantages compared with the first solution: it simplifies the feature extraction process, avoids the cumbersome wavelet basis selection and multi-layer decomposition process, directly uses the feature mapping matrix to construct the transformation function, which not only ensures the accuracy of feature extraction but also improves the calculation efficiency; in terms of constructing the objective function, the importance weights of features at different scales are automatically calculated through the feature mapping function, making the objective function better reflect the actual characteristics of the system and avoiding the subjectivity that may be brought by artificially specifying weights. This design not only reduces the algorithm complexity but also enhances the adaptive ability and generalization performance of the model, and is particularly suitable for processing the rapidly changing dynamic characteristics in the power system, and can more efficiently support subsequent optimal control decisions. Under a certain working condition, through the introduction of the feature mapping matrix, obvious improvements have been achieved in three aspects: the calculation time is reduced by more than 40% on average, the accuracy of feature expression is increased by about 25%, and the adaptation speed to system state changes is increased by about 35%.

[0156] As Figure 6 shown, according to one aspect of the present application, step S5 is specifically as follows:

[0157] S51: Obtain the optimal control sequence and the response strategy matrix, and construct an adaptive control law; based on the optimal control sequence and the response strategy matrix, obtain the control instruction according to the adaptive control law, and generate a control instruction matrix.

[0158] S52: Obtain the control instruction matrix and the pre-stored system state vector, and construct a Lyapunov function; use the Lyapunov function to analyze the system stability according to the control instruction matrix and the system state vector, and obtain the stability evaluation index.

[0159] S53. Obtain the control instruction matrix, stability evaluation index, and pre-stored real-time system state, and construct a control instruction correction function; input the control instruction matrix, stability evaluation index, and real-time system state into the control instruction correction function to perform real-time correction on the control instruction, and obtain a corrected control instruction sequence and execution feedback data.

[0160] Among them, the instruction correction can specifically be:

[0161] The corrected instruction u*(t)=u(t)+Kp·e(t)+Ki·∫e(t)dt+Kd·de(t) / dt+δu(t)

[0162] Among them: u(t) is the control instruction matrix at time t, and δu(t)=-μ·sign(s(t))·|s(t)|^αu; s(t)=e(t)+ λu·∫e(t)dt is the sliding mode surface; e(t) is the tracking error; Kp, Ki, Kd are PID parameters; μ is the approaching rate; αu is the convergence exponent; λu is the sliding mode surface parameter.

[0163] According to one aspect of the present application, step S53 may specifically include the following steps S531 to S534:

[0164] S531. Based on the control instruction matrix, stability evaluation index, and pre-stored real-time system state, calculate the system state deviation; perform state prediction and generate a state evaluation vector.

[0165] S532. Obtain the state evaluation vector; construct an adaptive correction function; calculate a correction coefficient matrix based on the state evaluation vector according to the adaptive correction function; perform response characteristic analysis and generate correction function parameters.

[0166] S533. Based on the control instruction matrix and correction function parameters, calculate the corrected control quantity; perform constraint check and generate a corrected control instruction sequence.

[0167] S534. Based on the corrected control instruction sequence and state evaluation vector, calculate the execution effect evaluation index, generate a feedback data packet; perform data format conversion on the execution feedback data packet and output the execution feedback data.

[0168] By constructing an adaptive correction function mechanism, real-time optimization and adjustment of control instructions are achieved. This step first calculates the system state deviation based on the state evaluation vector, evaluates the control effect through a state prediction model, and then uses the correction function parameters to perform real-time adjustment on the control instruction to ensure the execution effect of the control instruction. This closed-loop correction mechanism not only improves the control accuracy but also ensures the reliability and stability of the control system through the real-time calculation of the execution effect evaluation index.

[0169] In step S5, by constructing an adaptive control law and a real-time correction mechanism, the reliable execution of control instructions is achieved. First, a control instruction matrix is generated based on an optimized control sequence and a response strategy matrix; then, the stability of the system is analyzed through a Lyapunov function to ensure the stability of the control strategy; finally, the control instructions are adjusted in real time through a control instruction correction function, realizing a fast response to system state changes. This control method combining stability analysis and real-time correction not only ensures the stability of the control system but also improves the control accuracy through a feedback mechanism, and is especially suitable for scenarios such as port multi-energy systems that require high-precision control. By performing real-time acquisition and analysis of feedback data, a complete control closed-loop is formed, significantly enhancing the control reliability and robustness of the system.

[0170] In some embodiments of the present application, the data processing process is specifically as follows:

[0171] In step S1, dynamic characteristic data of the port multi-energy system is acquired and data preprocessing is performed, which specifically includes the following steps:

[0172] S11. Data acquisition and synchronization

[0173] Real-time power data PO(t) is acquired; equipment operation status data SO(t); historical operation data HO(t); system feedback data D_feedback(t - 1); when performing data pre-alignment, a timestamp alignment algorithm is applied; an enhanced four-dimensional tensor TO(t)=[PO(t), SO(t), HO(t), D_feedback(t - 1)] is constructed; time series interpolation and resampling are performed on the four-dimensional tensor to obtain a synchronized time series tensor T(t).

[0174] S12. Data quality evaluation and enhancement

[0175] The synchronized time series tensor T(t) is acquired; the quality index of T(t) is calculated: the comprehensive data quality score is calculated based on the quality index: , a data quality evaluation matrix DQM and a data quality enhancement recommendation matrix E are generated. Among them, the quality index may include an integrity index, a consistency index, and a timeliness index, and the calculation formulas for the three are respectively:

[0176] Integrity index: C(i)=Σ(valid data points) / total data points

[0177] Consistency index: U(i)=1 - Σ(data deviation) / reference value

[0178] Timeliness index: T(i)=exp(-λ·Δt)

[0179] S13. Multidimensional anomaly detection and processing

[0180] Obtain the synchronized time series tensor T(t) and the data quality evaluation matrix DQM; construct a multi-dimensional anomaly detection model, perform anomaly detection on the synchronized time series tensor in the time domain, perform wavelet transform anomaly detection on the synchronized time series tensor in the frequency domain, and perform local outlier factor detection on the synchronized time series tensor in the spatial domain; fuse the above detection results to obtain the anomaly detection result; correct the synchronized time series tensor based on the anomaly detection result to obtain the cleaned data tensor T'(t).

[0181] S14. Dynamic characteristic extraction and evaluation

[0182] Obtain the cleaned data tensor T'(t); construct multiple time windows of different scales, and use the time windows to extract features from the cleaned data tensor: ; calculate the importance score of the features: I(f)=XGBoost feature importance score, and calculate the dynamic response ability based on the feature importance score: D = Σ(I(f)·W(t,δ)), obtain the standardized system characteristic matrix M and the dynamic response ability index D;

[0183] In step S2, perform network-load interaction pattern recognition and feature extraction, which specifically includes the following steps:

[0184] S21. Enhanced network-load interaction pattern analysis

[0185] Obtain the standardized system characteristic matrix M, the power grid operation parameters E(t), the load fluctuation data L(t), and the dynamic response ability index D, construct a time-varying weight function w(t)=β0·exp(-γ0t), where β0 is the initial weight coefficient and γ0 is the decay coefficient; calculate the edge weight between network nodes E(i,j,t)=w(t)·sim(i,j) based on the time-varying weight function, where sim(i,j) is the cosine similarity of the load curves between nodes i and j; construct a time-varying network-load interaction graph G(V,E,t)=F[M,E(t),L(t),D] using the edge weights; calculate the interaction intensity matrix S based on the time-varying network-load interaction graph G(V,E,t).

[0186] S22. Deep spatio-temporal feature extraction

[0187] Obtain the time-varying network-load interaction graph G(V,E,t) and the interaction intensity matrix S; construct a spatio-temporal attention model: A(Q,K,V)=softmax(Q·K^T / √dk)·V·exp(-λt·Δt); and extract features from the time-varying network-load interaction graph G(V,E,t): F(t)=GNN(G(V,E,t),A(t)); perform temporal dependence modeling on the extracted features ST=LSTM(F(t)) to obtain the spatio-temporal feature matrix ST and the attention weight matrix A.

[0188] Attention calculation process: Similarity calculation: score = Q × K^T; Scaling: scaled_score = score / sqrt(dk); Softmax normalization: weight = softmax(scaled_score); Weighted summation: output = weight × V.

[0189] Among them, Q is the query matrix, representing the target information that needs to be focused on currently; in the network-load interaction graph, it represents the features of the node or time point that one wants to understand; the dimension is [N × dk], where N is the number of nodes and dk is the feature dimension; K represents the information key values that can be queried; in the network-load interaction graph, it represents the feature information of all nodes; the dimension is [N × dk], which needs to match the dimension of Q; V represents the actual information content that is extracted; in the network-load interaction graph, it contains the actual feature values of the nodes; the dimension is [N × dv], where dv is the dimension of the values.

[0190] S23. Adaptive pattern recognition

[0191] Obtain the spatio-temporal feature matrix ST and the attention weight matrix A; Based on the spatio-temporal feature matrix and the attention weight matrix, construct a dynamic density clustering model. Specifically, the DBSCAN algorithm + time window can be used, and an adaptive density threshold is adopted to construct the dynamic density clustering model; Use the dynamic density clustering model to perform clustering analysis on the spatio-temporal feature matrix, and track the time evolution feature of the clustering result Ev(t) = TrackClusterEvolution(C(t), C(t - 1)), where C(t) represents the time series obtained after clustering analysis, to obtain the network-load interaction feature vector I, the key pattern matrix K, and the pattern evolution matrix E.

[0192] S24. Multi-level feature fusion

[0193] Obtain the network-load interaction feature vector I, the key pattern matrix K, and the pattern evolution matrix E; Based on the network-load interaction feature vector, the key pattern matrix, and the pattern evolution matrix, construct a three-layer feature fusion network; At the bottom layer of the feature fusion network, fuse the original features; At the middle layer of the feature fusion network, fuse the pattern features; At the top layer of the feature fusion network, fuse the evolution features; Based on the optimization objective function, optimize the fusion result to obtain the fusion feature vector R and the feature mapping matrix T.

[0194] Among them, the optimization objective function is: ;

[0195] Among them, L_feature (feature layer loss) is used to measure the accuracy of the original feature fusion and calculate the reconstruction error of the underlying network-load interaction feature I; the expression is usually: ||I - I'||², where I' is the reconstructed feature; L_pattern (pattern layer loss) is used to measure the preservation degree of the key pattern K and measure the structural similarity of the pattern features; it can be expressed as: ||K - f(R)||², where f(·) is the pattern mapping function; L_evolution (evolution layer loss) is used to measure the expression degree of the system evolution characteristic E; characterize the time-series dynamic change law; in the form of: ||E - g(R, t)||², where g(·) is the evolution mapping function; , , respectively represent the weight coefficients of L_feature, L_pattern, and L_evolution.

[0196] In step S3, the automatic inertia response ability is evaluated and predicted, which specifically includes the following steps:

[0197] S31. Nonlinear inertia response characteristic analysis

[0198] Obtain the dynamic response ability index D, the network-load interaction feature vector I, and the fusion feature vector R, and construct a nonlinear response model: R(tr, fr, pr) = F[D, I, φ(tr)], where φ(tr) = αr·exp(-βr·tr)·sin(ωr·tr + θr); use the nonlinear response model to calculate the response characteristic parameters and perform parameter adaptive optimization: [αr, βr, ωr, θr] = ArgMin(L_response); construct the response characteristic tensor R(tr, fr, pr) and the nonlinear parameter set Θ based on the response characteristic parameters. Among them, R(tr, fr, pr) is the response characteristic tensor, fr is the response frequency, pr is the parameter in the nonlinear response model, and φ(tr) is the time-varying attenuation function of the nonlinear response characteristic: αr is the response amplitude coefficient; βr is the attenuation rate coefficient; ωr is the oscillation angular frequency; θr is the initial phase angle; exp(-βr·tr) represents the response attenuation characteristic; sin(ωr·tr + θr) represents the oscillation characteristic.

[0199] S32. Construction of multi-dimensional evaluation indicators

[0200] Obtain the response characteristic tensor R(tr, fr, pr) and the system operation constraint C, and design a multi-dimensional evaluation function E(x) = Σ(wi·xi), where wi is the weight coefficient and xi is the sub-item index; use the multi-dimensional evaluation function to calculate the sub-item index xi; perform weighted calculation on the sub-item index xi; the output final result is the comprehensive evaluation index matrix A.

[0201] S33. Construction of Time Series Prediction Model

[0202] Obtain the comprehensive evaluation index matrix A and historical response data HR; construct the time series attention network M(t), and use the time series attention network to construct the prediction model M(t + τ)=F[M(t), A, HR]; obtain the inertia response ability matrix B;

[0203] In step S4, perform multi-time scale collaborative optimization decision-making, which specifically includes the following steps:

[0204] S41. Time Scale Decomposition

[0205] Obtain the inertia response ability matrix B and the key mode matrix Key, construct an improved wavelet transform model, and use the wavelet transform model to decompose the inertia response ability matrix B and the key mode matrix Key into different time scales [T1, T2,..., Tn]; extract features from the decomposition results to obtain the multi-scale feature matrix MT;

[0206] S42. Hierarchical Optimization Objective Construction

[0207] Construct a multi-level optimization objective function O(y)=Σ(αi·fi(y)), where αi is the weight of each level and y is the parameter to be optimized; perform weight allocation on the multi-level objective function; combine the weighted objective functions and output the optimization objective function set F.

[0208] S43. Collaborative Optimization Solution

[0209] Obtain the optimization objective function set F and the system operation boundary BB; design a distributed optimization algorithm to solve the multi-objective optimization problem and obtain the optimized control sequence U and the response strategy matrix SS;

[0210] In step S5, generate and execute the adaptive control strategy, which specifically includes the following steps:

[0211] S51. Control Strategy Generation

[0212] Obtain the optimized control sequence U and the response strategy matrix SS, construct the adaptive control law CTL(t)=G[U(t), SS(t), e(t)], where e(t) is the tracking error; obtain the control instruction based on the adaptive control law and obtain the control instruction matrix CM.

[0213] S52. Stability Analysis

[0214] Obtain the control instruction matrix CM and the system state vector X, construct the Lyapunov function V(x), and use the Lyapunov function to analyze the system stability to obtain the stability evaluation index.

[0215] S53. Real-time Correction and Execution

[0216] Obtain the control instruction matrix CM, the stability evaluation index Z, and the real-time system state Y; design a correction function ΔC(t)=H[CM(t),Z(t),Y(t)] based on real-time feedback; use the control instruction correction function to correct the control instructions in real time to obtain a corrected control instruction sequence C' and execution feedback data D.

[0217] In this embodiment, first, various types of real-time data are collected, including power data, equipment operating status, etc. Triple-insurance data quality control is adopted: first, the data quality is evaluated, then anomalies are detected from three dimensions of time, frequency, and space, and finally dynamic features are extracted using a sliding window. Then, an intelligent network-load interaction analysis model is constructed. Using the "time-varying network-load interaction diagram", the changing characteristics of the system can be dynamically captured. This is like taking a "dynamic X-ray" of the system, which can not only show the current state but also analyze the changing trend. Especially when analyzing the network-load interaction characteristics, a three-level feature fusion network from the bottom layer to the high layer is adopted to ensure that no important information is missed. Then, based on the previous analysis results, through multi-scale feature analysis and multi-objective optimization, the optimal control strategy is calculated. The innovation lies in the use of an improved wavelet transform and a distributed optimization algorithm, which not only ensures the quality of the decision-making but also improves the calculation efficiency. Finally, instead of simply executing the control instructions, a real-time correction mechanism is added. Through Lyapunov stability analysis and adaptive correction, the reliable execution of the control instructions is ensured. This is like installing an "automatic error corrector" in the control system, which can adjust the control strategy in a timely manner according to the actual situation.

[0218] The technical advantages of the entire solution are mainly reflected in: first, achieving all-round data quality control; second, innovatively introducing the time-varying network-load interaction diagram and the three-layer feature fusion network; third, considering the dead zone characteristics of the system and improving the control accuracy; fourth, adopting distributed optimization to improve the calculation efficiency. These innovation points jointly solve the key technical problems in the participation of the port multi-energy supply system in power grid regulation, enabling the system to perform inertia response control more intelligently and efficiently.

[0219] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A control method for a multi-energy supply system in a port based on automatic inertia response, characterized in that, It includes the following steps: S1. Collect the time series data of the port multi-energy system, construct a synchronous time series tensor based on the time series data, and perform dynamic characteristic extraction according to the synchronous time series tensor to obtain a standardized system characteristic matrix and a dynamic response ability index. The time series data includes system real-time power data, port equipment operation status data, historical operation data, and feedback data; S2. Construct a time-varying network-load interaction graph based on the standardized system characteristic matrix, dynamic response ability index, pre-stored power grid operation parameters, and load fluctuation data, extract features from the time-varying network-load interaction graph to generate a spatio-temporal feature matrix, use the spatio-temporal feature matrix for pattern recognition to obtain a network-load interaction feature vector and a key pattern matrix, and obtain a fused feature vector and a feature mapping matrix through multi-level fusion; S3. Construct a response characteristic tensor based on the dynamic response ability index, network-load interaction feature vector, and fused feature vector, construct a time series prediction model based on the response characteristic tensor, pre-stored system operation constraints, and historical response data and perform prediction to obtain an inertia response ability matrix; S4. Based on the inertia response ability matrix and the key pattern matrix, obtain a multi-scale feature matrix through wavelet transform and optimize it to obtain an optimized control sequence and a response strategy matrix; S5. Generate a control instruction matrix for the system based on the optimized control sequence and the response strategy matrix, and perform stability analysis and real-time correction on the control instruction matrix to obtain a corrected control instruction sequence and execution feedback data.

2. The control method of the port multi-energy system based on automatic inertia response according to claim 1, characterized in that, Step S1 includes: Collect the time series data of the port multi-energy system and construct a synchronous time series tensor based on the time series data; Calculate a data quality evaluation matrix based on the synchronous time series tensor; Use the data quality evaluation matrix to perform anomaly detection and correction on the synchronous time series tensor to obtain a cleaned data tensor; Adopt a sliding window method to perform dynamic characteristic extraction on the cleaned data tensor to obtain a standardized system characteristic matrix and a dynamic response ability index.

3. The control method of the port multi-energy system based on automatic inertia response according to claim 2, characterized in that, The steps of anomaly detection and correction specifically include: Obtain the synchronous time series tensor and the data quality evaluation matrix, calculate the first-order difference and second-order difference values of the time series, and generate a time series feature vector; use the time series feature vector to perform sliding window statistical operations to generate a time domain anomaly score; Perform multi-scale wavelet decomposition on the synchronous time series tensor, calculate the energy distribution characteristics of the decomposition coefficients, and generate a frequency domain anomaly score; Calculate the local density distribution and distance distribution of sample points based on the synchronous time series tensor to generate a spatial domain anomaly score; Input the time domain anomaly score, frequency domain anomaly score, and spatial domain anomaly score into an evidence theory fusion framework, calculate the comprehensive confidence degree, and generate an anomaly comprehensive score; Use the anomaly comprehensive score to correct the abnormal data in the synchronous time series tensor to generate a cleaned data tensor.

4. The control method of the port multi-energy system based on automatic inertia response according to claim 1, characterized in that Step S2 specifically includes: Based on the standardized system characteristic matrix, dynamic response ability index, power grid operation parameters, and load fluctuation data, construct a time-varying weight function, calculate the edge weights between network nodes based on the time-varying weight function, construct a time-varying network-load interaction graph using the edge weights, and calculate an interaction intensity matrix based on the time-varying network-load interaction graph; Construct a spatio-temporal attention model based on the time-varying network-load interaction graph and the interaction intensity matrix; use the spatio-temporal attention model to extract features from the time-varying network-load interaction graph, generating a spatio-temporal feature matrix and an attention weight matrix; Obtain the spatio-temporal feature matrix and the attention weight matrix, and construct a dynamic density clustering model; use the dynamic density clustering model to perform clustering analysis on the spatio-temporal feature matrix, track the time-evolution characteristics of the clustering results, and obtain the network-load interaction feature vector, the key pattern matrix, and the pattern evolution matrix; Based on the network-load interaction feature vector, the key pattern matrix, and the pattern evolution matrix, construct a three-layer feature fusion network for multi-level fusion to obtain a fusion feature vector and a feature mapping matrix.

5. The control method of the multi-energy supply system for ports based on automatic inertia response according to claim 4, characterized in that, The steps of generating the spatio-temporal feature matrix and the attention weight matrix specifically include: Construct a multi-head attention calculation unit with a time decay factor based on the time-varying network-load interaction graph and the interaction intensity matrix to generate a decay attention matrix; Use the decay attention matrix to perform graph convolution operations to extract the local features of the nodes, generating a graph feature matrix; Input the graph feature matrix into a bidirectional gated recurrent unit to extract the temporal dependence features, generating a temporal feature vector; Perform feature fusion operations on the graph feature matrix and the temporal feature vector to generate the spatio-temporal feature matrix and the attention weight matrix.

6. The control method of the port multi-energy system based on automatic inertia response according to claim 4, characterized in that, The steps of multi-level fusion specifically include: Obtain the network-load interaction feature vector, construct an autoencoder network, perform feature dimensionality reduction to obtain compressed features, reconstruct features to obtain reconstructed features, calculate the reconstruction error, and output the underlying fusion features; Obtain the key pattern matrix and the underlying fusion features, construct a dynamic convolution kernel, perform multi-scale convolution operations, perform pooling processing to obtain an intermediate feature map, extract significant features, and output the middle-level fusion features; Obtain the pattern evolution matrix and the middle-level fusion features, construct an evolution feature extractor, calculate the state transition probability of the time series, extract the evolution pattern features, fuse static and dynamic features, and output the high-level fusion features; Obtain the underlying fusion features, the middle-level fusion features, and the high-level fusion features, construct a multi-layer optimization objective function, calculate the loss function values of the features at each layer, perform gradient descent optimization, update the fusion weights, and output the fusion feature vector and the feature mapping matrix.

7. The control method of the port multi-energy system based on automatic inertia response according to claim 1, characterized in that Step S3 includes: Obtain the dynamic response ability index, the network-load interaction feature vector, and the fusion feature vector, construct a non-linear response model, use the non-linear response model to calculate the response characteristic parameters, and construct a response characteristic tensor based on the response characteristic parameters; Based on the response characteristic tensor and the pre-stored system operation constraints, construct a multi-dimensional evaluation function, use the multi-dimensional evaluation function to calculate the sub-item indicators, perform weighted calculation on the sub-item indicators, and output the comprehensive evaluation index matrix; Use the comprehensive evaluation index matrix and the pre-stored historical response data to construct a time series prediction model, and perform time series prediction based on the time series prediction model to output the inertia response ability matrix.

8. The control method of the port multi-energy supply system based on automatic inertia response according to claim 7, characterized in that The steps of constructing the response characteristic tensor specifically include: Obtain the dynamic response ability index and the network-load interaction feature vector, calculate the hidden layer state transition matrix and the initial neuron weights based on the recursive neural network structure, and generate a basic model parameter set; Construct a piecewise function model using the basic model parameter set and the pre-stored system operation historical data, calculate the forward dead zone threshold and the reverse dead zone threshold, and generate a dead zone characteristic matrix; Based on the basic model parameter set and the dead zone characteristic matrix, calculate the response amplitude attenuation coefficient and the phase shift parameter according to the non-linear response function, and generate an initial characteristic parameter set; Based on the initial characteristic parameter set, construct a three-dimensional characteristic mapping framework, calculate the time dimension feature mapping value, the frequency dimension transformation coefficient, and the power dimension mapping parameter, and generate a response characteristic tensor.

9. The control method of the multi-energy supply system for ports based on automatic inertia response according to claim 1, characterized in that Step S4 is specifically as follows: Construct a wavelet transform model, use the wavelet transform model to decompose the inertia response ability matrix and the key mode matrix at different time scales, extract features from the decomposition results, and obtain a multi-scale feature matrix; Construct a multi-level objective function, allocate weights to the multi-level objective function based on the multi-scale feature matrix and the pre-stored system constraints, and obtain an optimized objective function set according to the multi-level objective function; Use a distributed optimization algorithm to solve the multi-objective optimization problem between the optimized objective function set and the pre-stored system operation boundary, and obtain an optimized control sequence and a response strategy matrix.

10. The control method of the port multi-energy system based on automatic inertia response according to claim 1, characterized in that, Step S5 is specifically as follows: Obtain the optimized control sequence and the response strategy matrix, construct an adaptive control law, and obtain control instructions based on the optimized control sequence and the response strategy matrix according to the adaptive control law, and generate a control instruction matrix of the system; Obtain the control instruction matrix and the pre-stored system state vector, construct a Lyapunov function, and analyze the system stability using the Lyapunov function according to the control instruction matrix and the system state vector to obtain a stability evaluation index; Obtain the control instruction matrix, the stability evaluation index, and the pre-stored real-time system state, construct a control instruction correction function, input the control instruction matrix, the stability evaluation index, and the real-time system state into the control instruction correction function to correct the control instructions in real time, and obtain a corrected control instruction sequence and execution feedback data.

Images