Multi-level Interactive Control Method for Flexible Loads in Ports Oriented to Ship-Port-Network Collaboration

Abstract

The present invention relates to the field of port intelligent technologies, and discloses a multi-level interactive control method for flexible loads in ports facing ship-port-network collaboration, including: obtaining a basic data set, and processing it based on a graph attention mechanism and a cross-layer attention mechanism to obtain a multi-layer graph feature tensor; performing probability distribution calculation and tensor decomposition processing on the multi-layer graph feature tensor and historical time-series data to obtain a time-series feature matrix; using a dynamic programming algorithm to determine time window division points, and processing based on a cross-scale attention mechanism to obtain a multi-scale dynamic feature set; combining physical constraint data and logical constraint data to construct a coupled feature matrix; based on a multi-time-scale objective function and system constraint data, using an improved NSGA-III algorithm to obtain a collaborative control strategy set; using a reinforcement learning method to iteratively optimize the control strategy to generate an optimal control strategy. Thus, multi-level collaborative control of port loads is achieved, and the dynamic response ability and control performance of the system are improved.

Description

Technical Field

[0001] The present invention belongs to the field of port intelligentization, and particularly relates to a multi-level interactive control method for flexible loads in ports facing ship-port-network collaboration. Background Art

[0002] With the deep promotion of the transformation of ports towards intelligentization and greening, the electrical loads in ports exhibit significant flexible characteristics. How to make full use of the flexible characteristics of port loads to achieve the collaborative optimization of ship electricity demand and port power systems is of great significance for improving port energy utilization efficiency and ensuring the stable operation of the power system.

[0003] Currently, the research in the field of port load flexible control mainly focuses on three aspects: First, the demand response method based on load aggregation, which realizes the centralized control of load groups by classifying and aggregating port equipment loads; second, the hierarchical control strategy based on model prediction, which decomposes and coordinates control objectives at different time scales using a hierarchical architecture; third, the multi-objective optimization method based on heuristic algorithms, which optimizes the economy and stability of the system by constructing a multi-objective function. These methods have improved the flexible control ability of port loads to a certain extent, but there are still some technical limitations.

[0004] By deeply analyzing the existing technical solutions, it is found that the following problems mainly exist: First, most of the existing load characteristic extraction methods adopt an independent modeling method, ignoring the coupling relationship between ship electricity consumption, port equipment, and power grid operation, and it is difficult to accurately capture the dynamic characteristics of the system; second, the traditional time window division method usually uses a fixed window size, which cannot adapt to the differences in load changes at different time scales, resulting in insufficient accuracy of feature extraction; third, the existing control strategy optimization methods have convergence problems when dealing with high-dimensional constraint conditions, and lack an effective strategy evaluation and optimization mechanism, making it difficult to ensure the continuous improvement of control effects; finally, when dealing with sudden changes in ship power demand, the existing solutions often have problems such as response lag and control oscillation due to the lack of inter-layer dynamic coupling modeling and multi-time scale collaborative control mechanisms. These technical problems seriously restrict the effect of port load flexible control. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a multi-level interactive control method for flexible loads in ports facing ship-port-network collaboration.

[0006] The technical solution adopted by the present invention is as follows:

[0007] An embodiment of the present invention proposes a multi-level interactive control method for port flexible loads oriented to ship-port-network collaboration, including the following steps: Step S1, obtain a pre-stored basic data set, the basic data set includes a ship data set, a port equipment data set, and a power grid topology data set, construct adjacency matrices for the ship layer, the port equipment layer, and the power grid layer based on the basic data set, and use the graph attention mechanism to process to obtain intra-layer feature weights and inter-layer feature weights, and generate a multi-layer graph feature tensor; Step S2, obtain and calculate the node feature probability distribution of the multi-layer graph feature tensor and the time series change probability distribution of the pre-stored historical time series data based on the multi-layer graph feature tensor, construct a multi-dimensional probability tensor, and use the tensor decomposition method to process to obtain a time series feature matrix; Step S3, obtain the time series feature matrix, use the dynamic programming algorithm to determine the time window division points, generate a multi-scale time window sequence, and process based on the cross-scale attention mechanism to obtain a multi-scale dynamic feature set; Step S4, obtain the pre-stored physical constraint data and logical constraint data, combine with the multi-scale dynamic feature set, construct an inter-layer static coupling matrix and a dynamic coupling matrix, and then generate a coupling feature matrix; Step S5, obtain the coupling feature matrix and the pre-stored system constraint data, construct a multi-time scale objective function and a constraint condition set, and use the multi-objective optimization algorithm to process to obtain a collaborative control strategy set; Step S6, obtain the collaborative control strategy set and the pre-stored evaluation index data, calculate the performance index value and perform strategy optimization, and generate an optimal control strategy.

[0008] Advantages of the present invention:

[0009] Through multi-layer graph feature extraction and probability time series modeling, the present invention improves the ability to capture system dynamic features; through adaptive time window division and multi-scale feature fusion, it enhances the response ability to load changes at different time scales; through the combination of multi-objective optimization and reinforcement learning, it realizes the continuous optimization and performance improvement of control strategies. The entire solution shows good adaptability and control effects in practical applications, especially in dealing with complex scenarios such as sudden changes in ship power demand, demonstrating significant performance advantages. Description of the drawings

[0010] Figure 1 is a flowchart of a multi-level interactive control method for port flexible loads oriented to ship-port-network collaboration according to an embodiment of the present invention.

[0011] Figure 2 is a flowchart of Step S1 of an embodiment of the present invention.

[0012] Figure 3 is a flowchart of Step S2 of an embodiment of the present invention.

[0013] Figure 4 is a flowchart of Step S3 of an embodiment of the present invention.

[0014] Figure 5 It is a flowchart of step S4 in an embodiment of the present invention.

[0015] Figure 6 It is a flowchart of step S5 in an embodiment of the present invention.

[0016] Figure 7 It is a flowchart of step S6 in an embodiment of the present invention. Detailed implementation manners

[0017] 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.

[0018] As Figure 1 shown, a multi-level interactive control method for flexible loads in ports oriented to ship-port-network collaboration is provided, including the following steps:

[0019] Step S1: Obtain a pre-stored basic data set, where the basic data set includes a ship data set, a port equipment data set, and a power grid topology data set. Based on the basic data set, construct adjacency matrices for the ship layer, the port equipment layer, and the power grid layer respectively, and use the graph attention mechanism to process to obtain intra-layer feature weights and inter-layer feature weights, and generate a multi-layer graph feature tensor;

[0020] Step S2: Obtain and calculate the node feature probability distribution of the multi-layer graph feature tensor and the time series change probability distribution of the pre-stored historical time series data based on the multi-layer graph feature tensor, construct a multi-dimensional probability tensor, and use the tensor decomposition method to process to obtain a time series feature matrix;

[0021] Step S3: Obtain the time series feature matrix, use the dynamic programming algorithm to determine the time window division points, generate a multi-scale time window sequence, and use the cross-scale attention mechanism to process to obtain a multi-scale dynamic feature set;

[0022] Step S4: Obtain the pre-stored physical constraint data and logical constraint data, combine with the multi-scale dynamic feature set, construct an inter-layer static coupling matrix and a dynamic coupling matrix, and further generate a coupling feature matrix;

[0023] Step S5: Obtain the coupling feature matrix and the pre-stored system constraint data, construct a multi-time scale objective function and a constraint condition set, and use the multi-objective optimization algorithm to process to obtain a collaborative control strategy set;

[0024] Step S6: Obtain the collaborative control policy set and pre-stored evaluation index data, calculate the performance index value and perform policy optimization to generate the optimal control policy.

[0025] In another embodiment of the present application, the basic data set may further include: power grid topology, ship callable power, port charging pile equipment power, and power grid demand response power. By constructing a flexible load control system for ship-port-grid three-layer collaboration, intelligent and precise regulation of the port power system is achieved. The overall solution adopts a technical route of "multi-layer graph feature extraction - probabilistic time series modeling - multi-scale window division - coupling constraint construction - multi-objective optimization solution - reinforcement learning optimization" to form a complete closed-loop control system. At the feature extraction level, the dynamic association of ship-port-grid three-layer data is realized through the graph attention mechanism; at the time series modeling level, the dynamic characteristics of the system are captured through probability distribution estimation and tensor decomposition; at the control optimization level, the continuous optimization of the control policy is realized through multi-objective optimization and reinforcement learning. This solution not only solves the problems of inflexible load response and poor inter-layer collaboration in traditional port power systems, but also significantly improves the overall performance of the system through multi-time scale collaborative control. Especially when dealing with complex power consumption scenarios in large ports, the solution can adaptively adjust the control policy to ensure the unified coordination of power quality, power balance, and economy.

[0026] According to one aspect of the present application, as Figure 2 shown, step S1 is specifically as follows:

[0027] Step S11: Obtain the pre-stored ship data set, extract the ship identification, power demand, and time window information therein; obtain the pre-stored port equipment data set, extract the equipment identification, capacity, and operating status information therein; obtain the pre-stored power grid topology data set, extract the node information and edge information therein, and generate an initial feature data set according to the ship identification, power demand, and time window information, equipment identification, capacity, and operating status information, and node information and edge information.

[0028] Step S12: Based on the ship power demand data in the initial feature data set, calculate the ship layer adjacency matrix; based on the port equipment operating status information in the initial feature data set, calculate the port equipment layer adjacency matrix; based on the power grid nodes and edge information in the initial feature data set, calculate the power grid layer adjacency matrix; generate a ship-port association matrix according to the connection relationship between the ship and the port equipment, and generate a port-grid association matrix according to the connection relationship between the port equipment and the power grid nodes.

[0029] Step S13: Process the ship layer adjacency matrix, port equipment layer adjacency matrix, and power grid layer adjacency matrix respectively using the graph attention mechanism to obtain intra-layer feature weight data; process the ship-port association matrix and port-grid association matrix using the cross-layer attention mechanism to obtain inter-layer feature weight data; combine the intra-layer feature weight data and inter-layer feature weight data for combined processing to generate a multi-layer graph feature tensor.

[0030] Among them, the multi-layer graph attention weight calculation formula:

[0031] Among them, the multi-layer graph attention weight calculation formula ; Where: A(i,j) is the attention weight of node i to node j; h_i and h_j are the feature vectors of nodes i and j; W is the trainable weight matrix; b is the bias vector; σ is the softmax activation function; LeakyReLU is the rectified linear unit with leakage; cos(h_i, h_j) is the cosine similarity of the feature vectors; p_i and p_j are the position encodings of the nodes; α, β, γ are the adaptive weight coefficients, optimized by gradient descent; || is the vector concatenation operation; is the square of the Euclidean distance. By introducing the position encoding and cosine similarity, the modeling ability of the topological relationship and feature similarity between nodes is enhanced.

[0032] In another embodiment of the present application, the ship data set may include: ship basic information: ship ID, tonnage level, maximum power demand; shore power demand: charging power curve, estimated berthing time, allowed charging time window; energy storage capacity: on-board energy storage capacity (if any); load characteristics: critical load ratio, interruptible load ratio; historical power consumption pattern data; port equipment data set: charging pile information: rated power, operating efficiency, dynamic loss; converter parameters: power capacity, control dead zone, dynamic response characteristics; energy storage system: capacity, charge and discharge efficiency, state of charge (SOC) range; load classification: rigid load, flexible load, transferable load; equipment operating status: online rate, failure rate, maintenance cycle; power grid topology data set; network structure: bus connection relationship, line impedance parameters; electrical parameters: voltage level, power factor, harmonic content; protection setting values: overcurrent, overvoltage setting values; operating limits: line capacity, transformer capacity; power quality indicators: voltage deviation, frequency fluctuation.

[0033] By introducing the graph attention mechanism and cross-layer attention mechanism, multi-level fusion processing is performed on ship, port equipment, and power grid data, realizing the adaptive association of the three-layer data of ship-port-grid. Specifically, first, adjacency matrices are constructed for ship power demand data, port equipment operation status information, and power grid topology data respectively to capture the topological relationships within each layer; then, the graph attention mechanism is used to adaptively learn the degree of mutual influence between nodes within the layer, and at the same time, the cross-layer attention mechanism is used to establish dynamic association relationships between different levels. Finally, the generated multi-layer graph feature tensor not only retains the original topological structure information of each layer but also reflects the coupling relationship between layers. This hierarchical data processing method enables the system to automatically identify and adjust the weights of key nodes within and between layers, thus providing a comprehensive and accurate feature representation for subsequent collaborative control and greatly improving the perception ability of the overall state of the ship-port-grid system. In addition, through the adaptive weight learning of the graph attention mechanism, the system can dynamically adapt to changes in port load and fluctuations in ship power consumption demand, ensuring the real-time and accuracy of feature extraction.

[0034] According to one aspect of the present application, as Figure 3 shown, step S2 is specifically as follows:

[0035] Step S21: Obtain the node feature data in the multi-layer graph feature tensor, and calculate the node feature probability distribution data by using the probability distribution estimation method; obtain the pre-stored historical time series data, and calculate the time series change probability distribution data by using the probability distribution estimation method; combine and process the node feature probability distribution data and the time series change probability distribution data to generate a multi-dimensional probability tensor.

[0036] Among them, the adaptive kernel density probability estimation formula:

[0037] ; Among them: P(x) is the probability density of feature x; n is the number of samples; K_h is the Gaussian kernel function with an adaptive bandwidth, ; u is the input vector, h is the adaptive bandwidth parameter, h = h_0·(1 + η·var(x_i)); w_i is the sample weight coefficient; s_i is the sample sparsity index; λ is the sparse penalty factor; x_i is the historical sample point; var(x_i) is the local variance; h_0 is the reference bandwidth; η is the bandwidth adjustment coefficient.

[0038] Step S22: Process the multi-dimensional probability tensor by using the Tucker decomposition method to obtain the core tensor and the factor matrix group; extract the main feature pattern data based on the core tensor, and combine and process it with the factor matrix group to generate a time series feature matrix.

[0039] Among them, the tensor decomposition model ; among them: X is the original tensor; G is the core tensor; , , is the factor moment matrix, is the tensor product operation of the i-th mode; is the orthogonal constraint term; λ is the regularization coefficient; α, β, γ are the modal weight coefficients; I is the identity matrix.

[0040] Based on the multi-layer graph feature tensor and historical time-series data, a dual processing mechanism of probability distribution estimation and tensor decomposition is adopted to achieve efficient extraction and compression of the system's dynamic features. Specifically, first, the probability distribution estimation method is used to calculate the probability distributions of node features and temporal changes respectively, accurately capturing the statistical characteristics of the system state; then, through the Tucker tensor decomposition method, the high-dimensional probability distribution information is decomposed into a core tensor and a set of factor matrices, significantly reducing the data dimension while retaining the key feature patterns. This probabilistic feature extraction method can not only effectively handle the randomness and uncertainty in the ship-port network system but also achieve data dimensionality reduction and feature compression through tensor decomposition, significantly reducing the computational complexity of subsequent processing. At the same time, due to the adoption of probabilistic modeling, the system has strong robustness to noise and abnormal data and can provide a stable and reliable feature representation. Especially when facing the complex and changeable load conditions in the port, this method effectively captures the statistical laws of load changes through probabilistic modeling, providing a reliable data basis for the optimization of subsequent control strategies.

[0041] According to one aspect of the present application, as Figure 4 shown, step S3 is specifically as follows:

[0042] Step S31: Obtain the time-series feature matrix, calculate the temporal correlation between features to obtain the correlation matrix; process the correlation matrix using the dynamic programming algorithm to obtain the optimal window division point set; generate a multi-scale time window sequence according to the optimal window division point set.

[0043] Among them, the window division cost function ; where: C(i,j) is the division cost of the interval [i,j]; is the feature variance term; V(i,j) = |dF(j) / dt - dF(i) / dt| is the change rate difference term; S(i,j) = exp(-(j-i) / τ) is the window scale penalty term; F(i:j) is the feature sequence of the interval [i,j]; μ(i:j) is the interval mean; , , are the adaptive weight coefficients; τ is the scale adjustment parameter.

[0044] Step S32: Segment the time-series feature matrix based on the multi-scale time window sequence to obtain a set of local feature matrices; aggregate the set of local feature matrices using a cross-scale attention mechanism to generate a multi-scale dynamic feature set.

[0045] By introducing a dynamic programming algorithm and a cross-scale attention mechanism for multi-scale partitioning and feature extraction of time windows, the adaptive capture of features at different time scales is realized. Specifically, first, calculate the temporal correlation between features based on the time-series feature matrix to construct a correlation matrix, which accurately reflects the association pattern of features changing over time; then use the dynamic programming algorithm to optimally partition the correlation matrix and automatically determine the optimal window partition points, effectively avoiding the information loss problem caused by artificially fixing the window size. Through the cross-scale attention mechanism for adaptive fusion of features at different time scales, the system can simultaneously handle power quality fluctuations at the millisecond level, power balance regulation at the second level, and load scheduling requirements at the minute level, significantly improving the capture ability of multi-time scale dynamic features. Especially when facing sudden changes in ship electrical loads, etc., this method can quickly locate the key time points and accurately extract the response features at different time scales, providing temporal feature support for subsequent hierarchical control strategies.

[0046] According to one aspect of the present application, as Figure 5 shown, step S4 is specifically as follows:

[0047] Step S41: Obtain the pre-stored physical constraint data and construct a physical constraint matrix; obtain the pre-stored logical constraint data and construct a logical constraint matrix; combine and process the physical constraint matrix and the logical constraint matrix to generate a static coupling matrix.

[0048] Step S42: Obtain the multi-scale dynamic feature set, calculate the inter-layer time-varying coupling strength to obtain a dynamic coupling matrix; fuse and process the static coupling matrix and the dynamic coupling matrix to generate a coupling feature matrix.

[0049] Among them, the coupling strength calculation formula ; where: H(t) is the coupling strength at time t; is the frequency-domain feature term, N is the number of sampling points; is the time-series feature term; is the modal feature term; w_i is the frequency component weight; f_i is the feature frequency; φ_i is the phase; h_t and c_t are the LSTM hidden state and memory unit; a_t is the attention weight; s_t is the state vector; α(t), β(t), γ(t) are time-varying weight coefficients.

[0050] By constructing a coupled feature matrix by combining static physical constraints and dynamic logical constraints, a comprehensive modeling of the multi-dimensional constraints of the ship-port-grid system is realized. Specifically, first, constraint matrices are constructed based on physical constraint data and logical constraint data respectively, comprehensively covering static constraints such as equipment operation limits and network security boundaries; then, the time-varying coupling strength between layers is calculated using a multi-scale dynamic feature set, accurately depicting the dynamic constraint relationship during the system operation. This modeling method that combines static and dynamic constraints not only ensures that the control strategy meets basic physical and safety constraints but also can adapt to the dynamic changes in the system operation state. Through the coupled feature matrix obtained by fusion processing, the system can achieve the coordinated optimization of ship power consumption demand, port equipment operation, and grid security while ensuring that various constraints are met. Especially when dealing with the coordinated control problem of multiple time scales, this method can timely respond to the changes in constraint conditions through the dynamic update of the coupled feature matrix, ensuring the real-time effectiveness of the control strategy.

[0051] According to one aspect of the present application, as Figure 6 shown, step S5 is specifically as follows:

[0052] Step S51: Obtain pre-stored frequency modulation index data and construct a millisecond-level objective function; obtain pre-stored voltage regulation index data and construct a second-level objective function; obtain pre-stored peak shaving index data and construct a minute-level objective function; combine and process these three objective functions to generate a multi-scale objective function set.

[0053] Among them, the millisecond-level objective mainly targets the fast response of power electronic devices; processes frequency perturbations and fast power fluctuations; and realizes primary frequency modulation function. The second-level objective: focuses on voltage stability control; coordinates the power distribution of multiple charging piles; and realizes secondary frequency modulation and voltage regulation. The minute-level objective: optimizes the load curve and peak-valley difference; coordinates the ship charging time sequence arrangement; and realizes economic dispatch and demand response. A hierarchical and time-sharing coordinated control system is formed, each aiming at different control objectives, jointly ensuring the stability and economy of the system. The fast response ensures stability, the medium-speed response optimizes the operation state, and the slow-speed response realizes economic dispatch, cooperating with each other.

[0054] Step S52: Obtain the equipment operation limits in the pre-stored system constraint data and generate an equipment constraint set; obtain the network security parameters in the pre-stored system constraint data and generate a network constraint set; obtain the time-series correlation data in the coupled feature matrix and generate a time-series constraint set; combine and process these three constraint sets to generate a comprehensive constraint condition set.

[0055] Step S53: Construct a multi-objective optimization model based on the multi-scale objective function set and the comprehensive constraint condition set; solve this model using an improved NSGA-III algorithm (Non-dominated Sorting Genetic Algorithm) to generate a coordinated control strategy set.

[0056] Among them, the improved NSGA-III algorithm is: the fitness evaluation function ; where: is the control accuracy target; is the control smoothness target; is the constraint violation target; w_i is the control error weight; λ_i is the reference point weight; r_i is the position of the current solution in the objective space; is the ideal reference point; α_j, β_j are the control cost weights; γ_k, η_k are the constraint violation penalty coefficients; g_k(x) and h_k(x) are the inequality and equality constraint functions respectively; Δu_j is the change in the control quantity; y_i and are the actual output and the desired output respectively.

[0057] Through constructing a multi-time-scale objective function and introducing the improved NSGA-III algorithm for multi-objective optimization, the collaborative optimization of different-level control objectives of the ship-port-network system is realized. Specifically, it is shown as follows: First, based on the control requirements of different time scales, the frequency modulation objective function at the millisecond level, the voltage regulation objective function at the second level, and the peak shaving objective function at the minute level are constructed respectively, comprehensively considering the multi-level control requirements of the system; then, a comprehensive constraint condition set is generated through the system constraint data to ensure the feasibility of the control strategy. The improved NSGA-III algorithm is used for solution. Through mechanisms such as non-dominated sorting, adaptive crossover, and guided mutation, it can efficiently explore the high-dimensional decision space and obtain the Pareto optimal solution set. This multi-level optimization method can not only meet the control requirements of different time scales simultaneously, but also ensure the diversity and convergence of the solutions through the improved evolutionary algorithm, significantly improving the quality and reliability of the control strategy. Especially in the face of complex scenarios such as port load fluctuations, this method can quickly generate a collaborative control strategy that meets multiple constraints, effectively ensuring the stable operation of the system.

[0058] According to one aspect of the present application, as Figure 7 shown, step S6 is specifically as follows:

[0059] Step S61: Obtain the accuracy requirements in the collaborative control strategy set and the pre-stored evaluation index data, and calculate the control accuracy index value; obtain the time requirements in the collaborative control strategy set and the pre-stored evaluation index data, and calculate the response time index value; obtain the stability requirements in the collaborative control strategy set and the pre-stored evaluation index data, and calculate the stability index value; combine and process these three types of index values to generate a comprehensive evaluation index set.

[0060] Step S62: Construct an evaluation function based on the comprehensive evaluation index set; use the reinforcement learning method to perform iterative optimization processing on the collaborative control strategy set to generate the optimal control strategy.

[0061] The reinforcement learning method adopts the Deep Deterministic Policy Gradient algorithm, and the policy gradient update formula ; where: J(θ) is the policy objective function; Q(s,a) is the state-action value function; V(s) is the state value function; π_θ(a|s) is the parameterized policy function; is the policy entropy term; is the parameter regularization term; α is the entropy weight coefficient; β is the regularization weight coefficient; is the parameter change penalty coefficient; is the parameter norm penalty coefficient; c is the parameter threshold; is the previous round of policy parameters.

[0062] The improved NSGA-III algorithm introduces reference point weights and control smoothness objectives, and achieves a trade-off between control accuracy and system stability through multi-objective optimization. At the same time, an adaptive penalty mechanism is adopted to handle constraint violations, improving the solution performance of the algorithm under complex constraints. The improved Deep Deterministic Policy Gradient algorithm adds a policy entropy term and a parameter regularization term. The policy entropy term promotes the exploration of the policy and prevents premature convergence to the local optimum; the parameter regularization term restricts the change range of the policy parameters and improves the stability of the learning process. Through these improvements, the algorithm improves the effect of policy optimization while ensuring convergence.

[0063] By introducing the reinforcement learning method to iteratively optimize the control policy, the adaptive optimization and performance improvement of the control policy are realized. Specifically, first, based on the evaluation index data, the control accuracy, response time, and stability index values are calculated respectively, and a comprehensive performance evaluation system is constructed; then, the Deep Deterministic Policy Gradient algorithm is used to evaluate and optimize the control policy, and the performance of the policy is continuously improved through the experience replay mechanism and the soft update algorithm. This optimization method based on reinforcement learning has significant advantages: through continuous interaction with the environment and policy updates, the system can continuously accumulate experience and improve the control policy, with strong adaptability; the experience replay mechanism improves the sample utilization efficiency and accelerates the policy convergence; the soft update algorithm ensures the smoothness of the policy update and avoids drastic fluctuations in the policy. Especially when dealing with complex dynamic systems such as ship-port networks, this method can gradually improve the performance of the control policy through continuous learning and optimization, achieving a comprehensive improvement in control accuracy, response speed, and system stability.

[0064] According to one aspect of the present application, step S11 is specifically as follows:

[0065] Step S111: Obtain the pre-stored ship dataset, encode the ship identification using a data segmentation algorithm to obtain identification coding data, segment the power demand data using a sliding time window method to obtain segmented power data, map the time window information using a time series encoder to obtain time feature data, and fuse the identification coding data, segmented power data, and time feature data to obtain initial ship feature data.

[0066] Among them, the segmentation coding function ;

[0067] Among them: E(x) is the segmentation coding value; is the feature transformation function; is the difference degree between adjacent segments; w_i is the segmentation weight coefficient; W_i is the feature transformation matrix; b_i is the bias vector; λ is the difference degree penalty factor; s_i is the feature mean of the i-th segment; N is the number of segments; sigmoid is the S-shaped activation function.

[0068] The improved sliding time window method is specifically: the power segmentation function ; Among them: P(t) is the segmented power feature at time t: is the weighted mean term; is the weighted standard deviation term; d(t) = (p_t - p_(t-L)) / L is the trend term; p_i is the original power value; w_i is the time decay weight; L is the window length; α(t), β(t), γ(t) are adaptive weight coefficients.

[0069] Step S112: Obtain the pre-stored port equipment dataset, discretize the equipment capacity data using an adaptive quantization method to obtain discrete capacity data, encode the equipment operation status data to obtain status coding data, and fuse the discrete capacity data and status coding data to obtain initial port feature data.

[0070] The adaptive quantization method is: the capacity quantization function ; Among them: Q(x) is the quantized capacity value; is the Gaussian kernel function; is the entropy term; q_i is the quantization center; μ_i is the Gaussian kernel mean; Σ_i is the covariance matrix; p_i is the quantization probability; λ is the entropy weight coefficient; M is the quantization level.

[0071] Step S113: Obtain the pre-stored power grid topology dataset, traverse the node information using a depth-first search algorithm to obtain node sequence data, simplify the edge information using a minimum spanning tree algorithm to obtain simplified edge set data, and combine the node sequence data and simplified edge set data to obtain initial power grid feature data.

[0072] Among them, the depth-first search algorithm is specifically as follows: the node importance calculation function ; where: I(v) is the importance of node v; C(v) = |N(v)| / max_u|N(u)| is the degree centrality of the node; is the betweenness centrality; D(v) = 1 / ∑(u∈V)d(v,u) is the closeness centrality; N(v) is the neighbor set of node v; σ_st is the number of shortest paths from s to t; σ_st(v) is the number of shortest paths from s to t passing through node v; d(v,u) is the shortest distance from node v to u; α, β, γ are adaptive weight coefficients.

[0073] Among them, the minimum spanning tree algorithm is specifically as follows: the edge weight calculation function ; where: W(e) is the comprehensive weight of edge e; is the feature similarity term; is the traffic weight term; is the time series correlation term; f_i and f_j are the feature vectors of the nodes connected by the edge; p_i and p_j are the node power values; t_i and t_j are the timestamps; α, β, γ are adaptive weight coefficients; σ is the similarity scale parameter; τ is the time series decay factor.

[0074] Step S114: Input the initial data of ship features, port features, and power grid features into the autoencoder for dimensionality reduction processing to obtain the initial feature dataset.

[0075] By introducing a multi-data processing mechanism to preprocess the basic data, the effective integration of heterogeneous data is realized. Specifically manifested as follows: First, the ship identification is encoded by using the data segmentation algorithm to ensure the unique identification of ship information; then, the power demand data is segmented by using the sliding time window method to accurately capture the load change characteristics; finally, the time window information is mapped by using the time series encoder to establish a unified time series representation. For the port equipment data, the discretization processing of the capacity data is realized by using the adaptive quantization method to ensure the standardization of the data; for the power grid topology data, the depth-first search and minimum spanning tree algorithms are used for processing to extract the key network structure information. The dimensionality reduction of the processed data by the autoencoder not only retains the key features of the original data but also reduces the computational complexity of subsequent processing. This systematic data preprocessing method lays a reliable data foundation for subsequent feature extraction and control optimization.

[0076] According to one aspect of the present application, step S13 is specifically as follows:

[0077] Step S131: Obtain in-layer feature data, extract spatial features using a gated graph convolutional network to obtain spatial feature data, extract temporal features using a recurrent neural network to obtain temporal feature data, and fuse the spatial feature data and the temporal feature data to obtain in-layer enhanced feature data.

[0078] Among them, the improved algorithm of the gated graph convolutional network is specifically: the spatial feature extraction function ; where: F(X) is the extracted spatial feature; is the normalized adjacency matrix; g(X) = sigmoid(XW_g + b_g) is the gating function; W_f and W_g are learnable weight matrices; D is the degree matrix; λ is the self-loop weight; ⊙ is the Hadamard product; σ is the activation function; X is the input feature matrix; I is the identity matrix.

[0079] Step S132: Obtain inter-layer correlation data, calculate the inter-layer feature correlation using a bidirectional attention mechanism to obtain correlation feature data, and perform feature enhancement processing on the correlation feature data using a residual network to obtain inter-layer enhanced feature data.

[0080] Step S133: Obtain the in-layer enhanced feature data and the inter-layer enhanced feature data, perform combined processing using an adaptive feature fusion algorithm to obtain fusion feature data, and perform high-dimensional feature mapping processing using a tensor network to generate a multi-layer graph feature tensor.

[0081] In another embodiment of the present application, step S13 may also be:

[0082] Step S131: Obtain in-layer feature data, perform dimensionality mapping processing on the node data using a node feature embedding algorithm to obtain node embedding data, extract spatial features from the node embedding data using a gated graph convolutional network to obtain spatial feature data, and extract temporal features from the spatial feature data using a recurrent neural network to obtain in-layer enhanced feature data.

[0083] Step S132: Obtain inter-layer correlation data, perform dimensionality alignment processing on the correlation data using a feature embedding network to obtain aligned feature data, calculate the inter-layer correlation of the aligned feature data using a bidirectional attention mechanism to obtain correlation feature data, and perform feature enhancement processing on the correlation feature data using a residual network to obtain inter-layer enhanced feature data.

[0084] Step S133: Obtain the in-layer enhanced feature data and the inter-layer enhanced feature data, perform feature combination processing using an adaptive feature fusion algorithm to obtain fusion feature data, perform data verification processing on the fusion feature data using a feature consistency verification algorithm to obtain verification feature data, and perform high-dimensional feature mapping processing on the verification feature data using a tensor network to generate a multi-layer graph feature tensor.

[0085] By introducing feature embedding preprocessing and multi-level attention mechanisms, the deep fusion and feature enhancement of the three-layer data of the ship-port network are achieved. Specifically, first, a node feature embedding algorithm is used to perform dimensional mapping on the original node data, solving the problem of dimensional mismatch; then, a gated graph convolutional network is used to extract spatial features, and a recurrent neural network is used to capture temporal features, realizing the comprehensive extraction of intra-layer features. In terms of inter-layer feature processing, a feature embedding network is used for dimensional alignment, the inter-layer correlation is calculated through a bidirectional attention mechanism, and a residual network is used for feature enhancement, ensuring the effective fusion of features at different levels. Finally, through an adaptive feature fusion algorithm and a feature consistency verification algorithm, the reliability and accuracy of the fused features are guaranteed. This multi-level feature processing method not only solves the problems of dimensional matching and feature loss in traditional methods, but also prominently strengthens key features through the attention mechanism, significantly improving the quality of feature representation. Especially when dealing with dynamic scenarios such as sudden changes in ship load, this method can quickly capture key features and make adaptive adjustments, providing reliable feature support for subsequent control decisions.

[0086] According to one aspect of the present application, step S21 is specifically as follows:

[0087] Step S211: Obtain a multi-layer graph feature tensor, calculate the probability distribution of node feature data using the adaptive kernel density estimation method to obtain the initial node probability data, perform conditional probability inference on the initial node probability data using a Bayesian network to obtain the node conditional probability data, and perform optimization processing on the node conditional probability data using the information entropy criterion to obtain the node feature probability distribution data.

[0088] Among them, the adaptive kernel density estimation method is specifically: the probability density function ; Among them: is the kernel function; is the adaptive bandwidth; d(x,x_i) is the feature distance; w_i is the sample weight; τ is the temperature parameter; h_0 is the reference bandwidth; σ(x) is the local standard deviation; d is the feature dimension; α is the bandwidth adjustment coefficient.

[0089] Among them, the information entropy criterion is specifically: the node probability optimization function ; Among them: H(p) is the optimization target; p_i is the node probability distribution; q_i is the prior probability distribution; KL(p||q) is the KL divergence; w_i is the node importance weight; λ is the regularization coefficient; N is the number of nodes.

[0090] Step S212: Obtain pre-stored historical time-series data, perform alignment processing on the time-series data using the dynamic time warping algorithm to obtain aligned time-series data, calculate the state transition probability of the aligned time-series data using the hidden Markov model to obtain transition probability data, and extract the evolution feature of the transition probability data using the sequence prediction model to obtain time-series change probability distribution data.

[0091] Among them, the dynamic time warping algorithm is specifically: the alignment cost function , ; where: D(i,j) is the cumulative alignment cost; d(x_i,y_j) is the local distance metric; γ is the path weight coefficient; x_i and y_j are the i-th and j-th points of the sequences to be aligned respectively; min{} is the minimum value operation.

[0092] Step S213: Obtain node feature probability distribution data and time-series change probability distribution data, perform high-dimensional mapping processing on the probability distribution data using the tensor network to obtain mapped probability data, perform integration processing on the mapped probability data using the adaptive probability fusion algorithm to obtain fusion probability data, and reconstruct the fusion probability data based on the tensor construction criterion to generate a multi-dimensional probability tensor.

[0093] Through the integration of probability distribution estimation and data densification processing, the high-quality extraction and representation of the system's dynamic features are realized. Specifically, first, the adaptive kernel density estimation method is used to calculate the preliminary probability distribution of node features, then conditional probability inference is performed through the Bayesian network, and optimization is carried out using the information entropy criterion to obtain an accurate node feature probability distribution. For time-series data, the dynamic time warping algorithm is used for alignment processing, the state transition probability is calculated through the hidden Markov model, and the evolution feature is extracted using the sequence prediction model. Importantly, a data densification processing mechanism is introduced in the probability distribution calculation process, effectively solving the data sparsity problem. Through high-dimensional mapping by the tensor network and integration by the adaptive probability fusion algorithm, the finally generated multi-dimensional probability tensor not only maintains the data density but also accurately reflects the dynamic features of the system. This probabilistic feature extraction method can not only effectively handle the randomness and uncertainty in the system but also significantly improve the quality of feature representation through data densification, providing a reliable data basis for subsequent dynamic feature analysis.

[0094] In another embodiment of the present application, step S2 can also be:

[0095] Step S21: Obtain the node feature data in the multi-layer graph feature tensor, calculate the initial node probability data using the probability distribution estimation method, perform density enhancement processing on the initial node probability data using the data densification algorithm to obtain the node density data, and perform optimization processing on the node density data using the probability correction network to obtain the node feature probability distribution data; obtain the pre-stored historical time series data, calculate the initial time series probability data using the time series probability estimation method, perform density enhancement processing on the initial time series probability data using the data densification algorithm to obtain the time series density data, and perform optimization processing on the time series density data using the probability correction network to obtain the time series change probability distribution data; combine and process the node feature probability distribution data and the time series change probability distribution data to generate a multi-dimensional probability tensor.

[0096] S22: Decompose and perform feature extraction processing on the multi-dimensional probability tensor to generate a time series feature matrix.

[0097] According to one aspect of the present application, step S31 is specifically as follows:

[0098] Step S311: Obtain the time series feature matrix, perform feature encoding processing on the time series data using the local binary pattern algorithm to obtain the feature encoding data, perform correlation measurement processing on the feature encoding data using the mutual information criterion to obtain the correlation measurement data, and perform screening processing on the correlation measurement data using the adaptive threshold method to obtain the correlation matrix.

[0099] Among them, the local binary pattern algorithm is specifically: the feature encoding function ; where: s(x) is the sign function, which is 1 when x≥0 and otherwise 0; g_c is the central point feature value; g_p is the neighborhood point feature value; w_p = exp(-d_p / σ) is the adaptive weight; d_p is the distance from the neighborhood point to the central point; P is the number of neighborhood points; σ is the distance scale parameter.

[0100] Step S312: Obtain the correlation matrix, perform graph structure conversion processing on the correlation data using the interval graph construction algorithm to obtain the interval graph data, perform optimized traversal processing on the interval graph data using the shortest path algorithm to obtain the path feature data, and perform division point calculation on the path feature data using the dynamic programming algorithm to obtain the optimal window division point set.

[0101] Among them, the improved shortest path algorithm is specifically: the path cost function ; where: C(p) is the total cost of path p; d(v_i,v_(i+1)) is the distance between nodes; f_i is the node feature vector; s(v_i,v_(i+1)) is the state transition cost; w_t, w_f, w_s are the adaptive weight coefficients; L is the path length.

[0102] The interval graph construction algorithm is specifically as follows: the interval similarity function ; where: S(I_i, I_j) is the similarity between intervals I_i and I_j; T(I_i, I_j) = |I_i∩I_j| / |I_i∪I_j| is the time overlap degree; is the feature overlap degree; is the duration similarity; f_i and f_j are interval feature vectors; d_i and d_j are interval durations; α, β, and γ are adaptive weight coefficients.

[0103] Step S313: Obtain the optimal window division point set, perform hierarchical division processing on the division point data using a multi-scale decomposition algorithm to obtain hierarchical sequence data, perform window construction processing on the hierarchical sequence data using an adaptive window generation algorithm to obtain window construction data, and perform boundary optimization processing on the window construction data using a time series alignment algorithm to generate a multi-scale time window sequence.

[0104] Among them, the multi-scale decomposition algorithm is specifically as follows: the scale decomposition function ; where: M(x, s) is the feature representation at scale s; φ_k(x) is the basis function of the k-th scale; w_k is the scale weight; s_k is the reference scale; τ is the scale decay parameter; K is the number of decomposition layers.

[0105] The time series alignment algorithm is specifically as follows: the alignment cost function ; where: A(i, j) is the cumulative alignment cost; is the feature distance; is the rate difference term; is the time penalty term; f() is the feature mapping function; r_i and r_j are the rates of change; t_i and t_j are timestamps; λ and μ are weight coefficients. The adaptive window generation algorithm can be implemented using existing technologies and will not be elaborated here.

[0106] In another embodiment of the present application, step S313 can also be:

[0107] Step S313: Obtain the optimal window division point set, perform hierarchical division processing on the division point data using a multi-scale decomposition algorithm to obtain hierarchical sequence data, perform time scale correlation processing on the hierarchical sequence data using a control strategy mapping algorithm to obtain correlation sequence data, perform window construction processing on the correlation sequence data using an adaptive window generation algorithm to obtain window construction data, and perform boundary optimization processing on the window construction data using a time series alignment algorithm to generate a multi-scale time window sequence.

[0108] Through the integration of local feature encoding and dynamic programming optimization, the adaptive segmentation and multi-scale representation of time series data are realized. Specifically, first, the local binary pattern algorithm is used to encode the features of time series data, the correlation is measured by the mutual information criterion, and an accurate feature correlation matrix is established; then, the interval graph construction algorithm and the shortest path algorithm are used for optimized traversal, and the optimal division points are calculated by combining the dynamic programming algorithm to achieve the adaptive division of time windows. On this basis, hierarchical division is performed through the multi-scale decomposition algorithm, a window sequence is constructed by the adaptive window generation algorithm, and the boundaries are optimized by the time series alignment algorithm, and finally a multi-scale time window sequence is generated. This method of time window division based on feature encoding and dynamic optimization can not only accurately capture the feature changes at different time scales, but also improve the accuracy of time series feature extraction through the adaptive division mechanism. Especially when dealing with multi-time scale cooperative control, this method can automatically adjust the size and position of the time window according to the dynamic characteristics of the system, ensuring the accurate capture of key time series features.

[0109] According to one aspect of the present application, step S42 is specifically as follows:

[0110] Step S421: Obtain a multi-scale dynamic feature set, perform multi-scale decomposition processing on the feature data by using the wavelet decomposition algorithm to obtain decomposed feature data, perform frequency domain feature extraction on the decomposed feature data by using the time-varying Fourier transform to obtain frequency domain feature data, and perform dynamic feature enhancement processing on the frequency domain feature data by using the adaptive spectrum analysis algorithm to obtain initial coupling strength data.

[0111] Among them, the wavelet decomposition algorithm in step S421 is specifically: the multi-scale feature extraction function ; where: W(f, s, τ) is the wavelet coefficient; f(t) is the input signal; ψ(t) is the wavelet basis function; s is the scale parameter; τ is the translation parameter; w(t) is the adaptive weight function; is the complex conjugate operation.

[0112] The adaptive spectrum analysis algorithm is specifically: the spectrum feature function ; where: F(ω, t) is the time-frequency feature; a_k(t) is the instantaneous amplitude; ω_k(t) is the instantaneous frequency; is the frequency weight; j is the imaginary unit; N is the number of components; σ is the frequency band parameter.

[0113] Step S422: Obtain the initial coupling strength data, perform extraction processing on the time series coupling features by using a recursive neural network to obtain time series coupling data, perform weight calculation processing on the time series coupling data by using the attention mechanism to obtain weight feature data, and perform feature mapping processing on the weight feature data by using the deep residual network to obtain a dynamic coupling matrix.

[0114] Step S423: Obtain the static coupling matrix and the dynamic coupling matrix, perform eigenmode extraction processing on the coupling data by using a modal decomposition algorithm to obtain modal feature data, perform feature fusion processing on the modal feature data by using an adaptive weighted fusion algorithm to obtain fusion feature data, and perform high-dimensional feature reconstruction processing on the fusion feature data by using a tensor network to generate a coupling feature matrix.

[0115] In another embodiment of the present application,

[0116] Step S421 may also be: Obtain a multi-scale dynamic feature set, perform multi-scale decomposition processing on the feature data by using a wavelet decomposition algorithm to obtain decomposed feature data, perform frequency-domain feature extraction on the decomposed feature data by using a time-varying Fourier transform to obtain initial frequency-domain feature data, perform validity verification processing on the initial frequency-domain feature data by using a spectrum consistency test algorithm to obtain frequency-domain feature data, perform dynamic feature enhancement processing on the frequency-domain feature data by using a spectrum correction network to obtain frequency-domain enhanced data, and perform coupling feature extraction processing on the frequency-domain enhanced data by using an adaptive spectrum analysis algorithm to obtain initial coupling strength data.

[0117] In another embodiment of the present application, Step S423 may also be:

[0118] Step S423: Obtain the static coupling matrix and the dynamic coupling matrix, perform eigenmode extraction processing on the coupling data by using a modal decomposition algorithm to obtain initial modal feature data, perform mapping relationship verification processing on the initial modal feature data by using a feature correlation analysis algorithm to obtain modal feature data, perform feature fusion processing on the modal feature data by using an adaptive weighted fusion algorithm to obtain fusion feature data, perform data verification processing on the fusion feature data by using a consistency verification algorithm to obtain verified feature data, and perform high-dimensional feature reconstruction processing on the verified feature data by using a tensor network to generate a coupling feature matrix.

[0119] Among them, the modal decomposition algorithm is specifically: a modal extraction function ; where: IMF(t) is the intrinsic mode function; x(t) is the original signal; c_i(t) is the component function; is the adaptive weight; H() is the Hilbert transform; M is the number of modes; λ is the weight adjustment parameter.

[0120] Through the combination of frequency-domain analysis and dynamic feature mapping, the accurate modeling and dynamic characterization of the system coupling relationship are achieved. Specifically, it is manifested in the following aspects: First, the wavelet decomposition algorithm and time-varying Fourier transform are used to process multi-scale dynamic features, and the effectiveness of frequency-domain features is ensured through the spectrum consistency test algorithm; then the spectrum correction network is used for dynamic feature enhancement, and the coupling features are extracted through the adaptive spectrum analysis algorithm to obtain the initial coupling strength data. In terms of dynamic feature processing, a recurrent neural network is used to extract time-series coupling features, the weights are calculated through the attention mechanism, and the depth residual network is used for feature mapping to generate a dynamic coupling matrix. Finally, through the modal decomposition algorithm, adaptive weighted fusion, and tensor network reconstruction, the complete expression of the coupling features is realized. The advantage of this dynamic coupling modeling method lies in that it can consider both frequency-domain and time-domain features simultaneously, and ensure the accuracy of feature extraction through multiple verification mechanisms. Especially when dealing with scenarios such as sudden changes in the electrical load of ships, it can quickly capture the dynamic coupling changes of the system, providing a reliable basis for the real-time adjustment of control strategies.

[0121] According to one aspect of the present application, step S53 is specifically as follows:

[0122] Step S531: Obtain a multi-scale objective function set and a comprehensive constraint condition set, use the Lagrangian relaxation algorithm to perform conversion processing on the constraint conditions to obtain relaxed constraint data, use the penalty function method to perform optimization problem construction processing on the relaxed constraint data to obtain initial optimization problem data, and use the gradient projection method to perform feasible domain mapping processing on the initial optimization problem data to obtain optimization model data.

[0123] Among them, the Lagrangian relaxation algorithm is specifically: Relax the objective function ; Where: L(x,λ) is the relaxed Lagrangian function; f(x) is the original objective function; g_i(x) is the constraint function; λ_i is the Lagrange multiplier; ρ is the penalty factor; m is the number of constraints.

[0124] The penalty function method is specifically: The penalty cost function ; Where: P(x) is the objective function with penalty terms; f(x) is the original objective function; g_i(x) is the inequality constraint; h_i(x) is the equality constraint; μ is the penalty factor; α_i, β_i are the constraint weights; p, q are the penalty exponents.

[0125] The gradient projection method is specifically: The projection update function ; Where: P_C is the projection operator on the feasible domain C; is the gradient of the objective function; is the gradient of the constraint function; is the adaptive step size; μ_i is the constraint weight; β is the step size adjustment parameter.

[0126] Step S532: Obtain the optimized model data, perform hierarchical processing on the solution space using the non-dominated sorting algorithm to obtain hierarchical solution set data, perform feature recombination processing on the hierarchical solution set data using the adaptive crossover operator to obtain recombined feature data, and perform diversity enhancement processing on the recombined feature data using the guided mutation algorithm to obtain candidate strategy data.

[0127] Among them, the non-dominated sorting algorithm is specifically: the dominance degree calculation function ; where: D(x) is the dominance degree of solution x; is the dominance indication function, which is 1 when x dominates y and 0 otherwise; d(x,y) is the solution space distance; w_i is the objective weight; τ is the distance scale parameter; N is the population size.

[0128] The guided mutation algorithm is specifically: the mutation operation function ; where: M(x) is the mutated solution; N(0,Σ) is the Gaussian noise; is the current optimal solution; is the gradient direction; α, β, γ are the adaptive weight coefficients; Σ is the covariance matrix.

[0129] Step S533: Obtain the candidate strategy data, perform dynamic partitioning processing on the strategy space using the reference point adaptive update algorithm to obtain partitioned strategy data, perform neighborhood optimization processing on the partitioned strategy data using the local search algorithm to obtain optimized strategy data, and perform screening processing on the optimized strategy data using the elite retention strategy to generate a collaborative control strategy set.

[0130] Among them, the reference point adaptive update algorithm is specifically: the reference point update function ; where: r_t is the reference point at time t; is the ideal point; z_c is the current optimal solution; z_n is the nearest neighbor non-dominated solution; w_1, w_2, w_3 are the adaptive weights; λ is the learning rate.

[0131] The local search algorithm is specifically: the search direction function ; where: d(x) is the search direction; is the gradient of the objective function; is the optimal solution direction; is the neighborhood direction; is the local optimal solution; x_i is the neighborhood solution; w_i is the neighborhood weight; α, β, γ are the adaptive coefficients.

[0132] In another embodiment of the present application, step S53 may also be:

[0133] Step S531: Obtain a multi-scale objective function set and a comprehensive constraint condition set, use the Lagrangian relaxation algorithm to perform transformation processing on the constraint conditions to obtain relaxed constraint data, use the penalty function method to perform optimization problem construction processing on the relaxed constraint data to obtain initial optimization problem data, use the convergence analysis algorithm to perform stability evaluation processing on the initial optimization problem data to obtain stability evaluation data, and use the gradient projection method to perform feasible region mapping processing on the stability evaluation data to obtain optimization model data.

[0134] Step S532: Obtain the optimization model data, use the non-dominated sorting algorithm to perform hierarchical processing on the solution space to obtain hierarchical solution set data, use the convergence monitoring algorithm to perform stability verification processing on the hierarchical solution set data to obtain verified solution set data, use the adaptive crossover operator to perform feature recombination processing on the verified solution set data to obtain recombined feature data, and use the guided mutation algorithm to perform diversity enhancement processing on the recombined feature data to obtain candidate strategy data.

[0135] Step S533: Obtain the candidate strategy data, use the reference point adaptive update algorithm to perform dynamic partitioning processing on the strategy space to obtain partitioned strategy data, use the convergence compensation algorithm to perform stability enhancement processing on the partitioned strategy data to obtain stable strategy data, use the local search algorithm to perform neighborhood optimization processing on the stable strategy data to obtain optimized strategy data, and use the elitist retention strategy to perform screening processing on the optimized strategy data to generate a cooperative control strategy set.

[0136] By integrating the multi-objective optimization and convergence guarantee mechanism, the stable solution and performance optimization of the high-dimensional control strategy are realized. Specifically, first, the Lagrangian relaxation algorithm and the penalty function method are used to transform the constraint conditions, and the stability evaluation is carried out through the convergence analysis algorithm to establish an optimization model with convergence guarantee; then, the non-dominated sorting algorithm is used to perform hierarchical processing of the solution space, the stability verification is carried out through the convergence monitoring algorithm, and the diversity of the solution is enhanced by using the adaptive crossover operator and the guided mutation algorithm. In the strategy optimization stage, the reference point adaptive update algorithm is used for dynamic partitioning, the stability is enhanced through the convergence compensation algorithm, and the quality of the solution is continuously improved by combining the local search algorithm and the elitist retention strategy. This optimization method with convergence guarantee not only solves the convergence problem in high-dimensional multi-objective optimization, but also improves the solution efficiency through multiple optimization mechanisms. Especially when dealing with the optimization of control strategies under complex constraint conditions, it can stably obtain a high-quality Pareto optimal solution set.

[0137] According to one aspect of the present application, step S62 is specifically as follows:

[0138] Step S621: Obtain the comprehensive evaluation index set, perform quantization processing on the index data using the fuzzy evaluation algorithm to obtain the quantized index data, perform weight calculation processing on the quantized index data using the analytic hierarchy process to obtain the weight allocation data, and perform index fusion processing on the weight allocation data using the comprehensive evaluation model to obtain the evaluation function data.

[0139] Among them, the fuzzy evaluation algorithm is specifically: the fuzzy evaluation function ; where: E(x) is the comprehensive evaluation value; μ_i(x) is the membership function of the i-th index; d_i(x) is the index deviation; w_i is the index weight; σ_i is the fuzzy scale parameter; n is the number of indexes.

[0140] The analytic hierarchy process is specifically: the weight calculation function ; where: W(A) is the eigenvector; A is the judgment matrix; e_i is the i-th eigenvector component; is the geometric mean; is the consistency deviation; a_ij is the judgment matrix element; λ_max is the maximum eigenvalue; n is the matrix order.

[0141] Step S622: Obtain the evaluation function data and the cooperative control strategy set, perform state evaluation processing on the strategy data using the deep deterministic policy gradient algorithm to obtain the state value data, perform action evaluation processing on the state value data using the advantage function method to obtain the action value data, and perform value update processing on the action value data using the double temporal difference algorithm to obtain the strategy update data.

[0142] Among them, the strategy network optimization method is specifically: the parameter update function ; where: L(θ_t) is the policy loss function; is the regularization term; η is the learning rate; β is the regularization weight; α is the momentum coefficient; γ is the norm penalty coefficient; c is the parameter threshold; θ_t is the policy parameter at time t.

[0143] The advantage function method is specifically: the advantage evaluation function ; where: A(s,a) is the advantage value; Q(s,a) is the state-action value function; V(s) is the state value function; is the policy entropy; is the reference action item; π(a|s) is the policy function; is the expert demonstration action; α, β are the adaptive weights; σ is the similarity parameter.

[0144] The double temporal difference algorithm is specifically: the value function update function ; where: ΔQ(s,a) is the Q-value update amount; r is the immediate reward; is the state value function; γ is the discount factor; λ is the mixing coefficient; α is the learning rate; is the next state; is the maximum action value.

[0145] Step S623: Obtain the policy update data, perform sampling processing on historical data using the experience replay mechanism to obtain the sampled policy data, perform parameter optimization processing on the sampled policy data using the policy network to obtain the optimized policy data, and perform stability enhancement processing on the optimized policy data using the soft update algorithm to generate the optimal control policy.

[0146] The soft update algorithm is specifically: the parameter update function ; where: is the target network parameter; θ is the online network parameter; θ is the historical optimal parameter; θ_e is the integrated parameter; τ is the soft update coefficient; w_1, w_2, w_3 are the adaptive weight coefficients; t is the update time.

[0147] By integrating deep reinforcement learning and adaptive policy optimization, continuous improvement and performance enhancement of the control policy are achieved. Specifically, first, the fuzzy evaluation algorithm and the analytic hierarchy process are used to quantify the evaluation indicators and assign weights, constructing a comprehensive evaluation function; then, the state is evaluated through the deep deterministic policy gradient algorithm, the action value is evaluated using the advantage function method, and the policy value is updated through the double temporal difference algorithm. During the policy optimization process, the experience replay mechanism is used for historical data sampling, parameter optimization is performed through the policy network, and the soft update algorithm is used to enhance the stability of the policy. This policy optimization method based on deep reinforcement learning has significant advantages: through continuous interactive learning and policy updates, the system can continuously accumulate experience and improve the control policy; the use of the experience replay and soft update mechanisms effectively improves the learning efficiency and policy stability; especially when dealing with complex load scenarios in ports, this method can continuously improve the adaptability and performance of the control policy through real-time learning and optimization.

[0148] In summary, the present invention addresses the problem of ship-port-network collaborative control and proposes a multi-level interactive flexible load control method for ports, aiming to construct a complete technical framework of "feature extraction - time series modeling - window division - coupling modeling - policy optimization - performance improvement". At the feature extraction level, the graph attention mechanism and the cross-layer attention mechanism are introduced to achieve dynamic association and feature fusion of the data of the three layers of the ship layer, port equipment layer, and power grid layer. Through node feature embedding preprocessing and multi-level attention calculation, the problems of inter-layer feature mismatch and information loss are effectively solved.

[0149] In terms of time series modeling, the solution combines probability distribution estimation and tensor decomposition techniques to probabilistically model the dynamic characteristics of the system. In particular, a data densification processing mechanism is introduced to effectively solve the data sparsity problem in traditional methods. In the time window division step, a dynamic programming algorithm is used to calculate the optimal division points, and combined with a cross-scale attention mechanism, it realizes the adaptive capture of features at different time scales, breaking through the limitations of traditional fixed window methods.

[0150] In terms of control strategy optimization, the solution constructs a multi-time scale objective function, comprehensively considering the control requirements of millisecond-level frequency modulation, second-level voltage regulation, and minute-level peak shaving. The improved NSGA-III algorithm is used for multi-objective optimization, and a convergence compensation mechanism is introduced to effectively solve the convergence problem under high-dimensional constraint conditions. In particular, in the strategy optimization stage, a deep reinforcement learning method is introduced, and through an experience replay mechanism and a soft update algorithm, the continuous optimization and performance improvement of the control strategy are realized.

[0151] The technical advantages of the present invention are mainly reflected in: First, through multi-layer graph feature extraction and probabilistic time series modeling, the ability to capture the dynamic characteristics of the system is improved; second, through adaptive time window division and multi-scale feature fusion, the response ability to load changes at different time scales is enhanced; third, through the combination of multi-objective optimization and reinforcement learning, the continuous optimization and performance improvement of the control strategy are realized. The solution shows good adaptability and control effect in practical applications. Especially when dealing with complex scenarios such as sudden changes in ship power demand, it demonstrates significant performance advantages.

[0152] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-level interactive control method for port flexible loads oriented to ship-port network collaboration, characterized in that: The steps include: Step S1, obtaining a pre-stored basic data set, wherein the basic data set includes a ship data set, a port equipment data set, and a power grid topology data set, and constructing adjacency matrices of a ship layer, a port equipment layer, and a power grid layer based on the basic data set, respectively, and using a graph attention mechanism to obtain intra-layer feature weights and inter-layer feature weights, and generating a multi-layer graph feature tensor; Step S2, obtaining and calculating the node feature probability distribution of the multi-layer graph feature tensor and the time series change probability distribution of the historical time series data based on the multi-layer graph feature tensor and the pre-stored historical time series data, constructing a multi-dimensional probability tensor, and using a tensor decomposition method to obtain a time series feature matrix; Step S3, obtaining a time series feature matrix, using a dynamic programming algorithm to determine the time window division points, generating a multi-scale time window sequence, and obtaining a multi-scale dynamic feature set based on a cross-scale attention mechanism; Step S4, obtaining pre-stored physical constraint data and logical constraint data, combining with a multi-scale dynamic feature set, constructing an inter-layer static coupling matrix and a dynamic coupling matrix, and then generating a coupling feature matrix; Step S5, obtaining the coupling characteristic matrix and pre-stored system constraint data, constructing a multi-time scale objective function and constraint condition set, and using a multi-objective optimization algorithm to obtain a collaborative control strategy set; Step S6: Obtain the collaborative control strategy set and pre-stored evaluation index data, calculate the performance index value and perform strategy optimization to generate the optimal control strategy.

2. The multi-level interactive control method of port flexible load for ship-port network collaboration according to claim 1 is characterized in that: Step S1 is specifically as follows: Step S11, obtaining a pre-stored ship data set, extracting the ship identification, power demand and time window information therein; obtaining a pre-stored port equipment data set, extracting the equipment identification, capacity and operating status information therein; obtaining a pre-stored power grid topology data set, extracting the node information and edge information therein; generating an initial feature data set according to the ship identification, power demand and time window information, the equipment identification, capacity and operating status information, and the node information and edge information; Step S12: Based on the ship power demand data in the initial feature data set, a ship-level adjacency matrix is ​​calculated; Based on the port equipment operation status information in the initial feature data set, the port equipment layer adjacency matrix is ​​calculated; Based on the grid node and edge information in the initial feature data set, the grid layer adjacency matrix is ​​calculated; Generate a ship-port association matrix based on the connection relationship between ships and port equipment, and generate a port-grid association matrix based on the connection relationship between port equipment and power grid nodes; Step S13, the ship layer adjacency matrix, the port equipment layer adjacency matrix and the power grid layer adjacency matrix are processed respectively using the graph attention mechanism to obtain intra-layer feature weight data; The ship-port association matrix and the port-network association matrix are processed using a cross-layer attention mechanism to obtain the inter-layer feature weight data; the intra-layer feature weight data and the inter-layer feature weight data are combined to generate a multi-layer graph feature tensor.

3. The multi-level interactive control method of port flexible load for ship-port network collaboration according to claim 1 is characterized in that: Step S2 is specifically as follows: Step S21, obtaining node feature data in the multi-layer graph feature tensor, and using a probability distribution estimation method to calculate the node feature probability distribution data; obtaining pre-stored historical time series data, and using a probability distribution estimation method to calculate the time series change probability distribution data; combining the node feature probability distribution data and the time series change probability distribution data to generate a multi-dimensional probability tensor; Step S22: Process the multidimensional probability tensor using the Tucker decomposition method to obtain a core tensor and a factor matrix group; extract the main feature pattern data based on the core tensor, combine it with the factor matrix group, and generate a time series feature matrix.

4. The multi-level interactive control method of port flexible load for ship-port network collaboration according to claim 1 is characterized in that: Step S3 is specifically as follows: Step S31, obtaining a time series feature matrix, calculating the time series correlation between features to obtain a correlation matrix; processing the correlation matrix using a dynamic programming algorithm to obtain an optimal window partition point set; generating a multi-scale time window sequence according to the optimal window partition point set; Step S32: segment the temporal feature matrix based on the multi-scale time window sequence to obtain a local feature matrix set; aggregate the local feature matrix set using a cross-scale attention mechanism to generate a multi-scale dynamic feature set.

5. The multi-level interactive control method of port flexible load for ship-port network collaboration according to claim 1 is characterized in that: Step S4 is specifically as follows: Step S41, obtaining pre-stored physical constraint data and constructing a physical constraint matrix; obtaining pre-stored logical constraint data and constructing a logical constraint matrix; combining the physical constraint matrix and the logical constraint matrix to generate a static coupling matrix; Step S42, obtaining a multi-scale dynamic feature set, calculating the inter-layer time-varying coupling strength to obtain a dynamic coupling matrix; The static coupling matrix and the dynamic coupling matrix are fused to generate a coupling feature matrix.

6. The multi-level interactive control method of port flexible load for ship-port network collaboration according to claim 1 is characterized in that: Step S5 is specifically as follows: Step S51, obtaining pre-stored frequency modulation index data, constructing a millisecond-level objective function; obtaining pre-stored voltage modulation index data, constructing a second-level objective function; obtaining pre-stored peak modulation index data, constructing a minute-level objective function; combining and processing these three objective functions to generate a multi-scale objective function set; Step S52: Obtain the equipment operation limit values ​​in the pre-stored system constraint data to generate an equipment constraint set; obtain the network security parameters in the pre-stored system constraint data to generate a network constraint set; obtain the timing correlation data in the coupling characteristic matrix to generate a timing constraint set; combine and process these three constraint sets to generate a comprehensive constraint condition set; Step S53: construct a multi-objective optimization model based on the multi-scale objective function set and the comprehensive constraint condition set; solve the model using the improved NSGA-III algorithm to generate a collaborative control strategy set.

7. The multi-level interactive control method of port flexible load for ship-port network collaboration according to claim 1 is characterized in that: Step S6 is specifically as follows: Step S61: obtaining the accuracy requirements in the collaborative control strategy set and the pre-stored evaluation index data, and calculating the control accuracy index value; obtaining the time requirements in the collaborative control strategy set and the pre-stored evaluation index data, and calculating the response time index value; Obtain the stability requirements in the collaborative control strategy set and the pre-stored evaluation index data, and calculate the stability index value; combine the three types of index values ​​to generate a comprehensive evaluation index set; Step S62: constructing an evaluation function based on the comprehensive evaluation index set; The reinforcement learning method is used to iteratively optimize the collaborative control strategy set to generate the optimal control strategy.

8. The multi-level interactive control method of port flexible load for ship-port-network collaboration according to claim 2 is characterized in that: Step S11 is specifically as follows: Step S111, obtaining a pre-stored ship data set, encoding the ship identification using a data segmentation algorithm to obtain identification coding data, segmenting the power demand data using a sliding time window method to obtain segmented power data, mapping the time window information using a time series encoder to obtain time feature data, and fusing the identification coding data, segmented power data and time feature data to obtain initial ship feature data; Step S112, obtaining a pre-stored data set of port equipment, using an adaptive quantization method to discretize the equipment capacity data to obtain discrete capacity data, encoding the equipment operation status data to obtain status encoding data, and fusing the discrete capacity data and the status encoding data to obtain initial port characteristic data; Step S113, obtaining a pre-stored power grid topology data set, using a depth-first search algorithm to traverse the node information to obtain node sequence data, using a minimum spanning tree algorithm to simplify the edge information to obtain simplified edge set data, and combining the node sequence data and the simplified edge set data to obtain the initial power grid characteristic data; Step S114: input the initial data of ship characteristics, the initial data of port characteristics and the initial data of power grid characteristics into the autoencoder for dimensionality reduction processing to obtain an initial feature data set.

9. The multi-level interactive control method of port flexible load for ship-port network collaboration according to claim 3 is characterized in that: Step S21 is specifically as follows: Step S211, obtaining a multi-layer graph feature tensor, using an adaptive kernel density estimation method to perform probability distribution calculation on the node feature data to obtain node probability initial data, using a Bayesian network to perform conditional probability reasoning on the node probability initial data to obtain node conditional probability data, and using an information entropy criterion to optimize the node conditional probability data to obtain node feature probability distribution data; Step S212, obtaining pre-stored historical time series data, using a dynamic time warping algorithm to align the historical time series data to obtain aligned time series data, using a hidden Markov model to calculate the state transition probability of the aligned time series data to obtain transition probability data, and using a sequence prediction model to extract evolution features of the transition probability data to obtain time series change probability distribution data; Step S213, obtain node feature probability distribution data and time series change probability distribution data, use a tensor network to perform high-dimensional mapping processing on the probability distribution data to obtain mapping probability data, use an adaptive probability fusion algorithm to integrate the mapping probability data to obtain fused probability data, reconstruct the fused probability data based on the tensor construction criterion, and generate a multidimensional probability tensor.

10. The multi-level interactive control method of port flexible load for ship-port-network collaboration according to claim 4 is characterized in that: Step S31 is specifically as follows: Step S311, obtaining a time series feature matrix, using a local binary pattern algorithm to perform feature encoding processing on the time series data to obtain feature encoding data, using a mutual information criterion to perform correlation measurement processing on the feature encoding data to obtain correlation measurement data, and using an adaptive threshold method to screen the correlation measurement data to obtain a correlation matrix; Step S312, obtaining a correlation matrix, using an interval graph construction algorithm to perform graph structure conversion processing on the correlation data to obtain interval graph data, using a shortest path algorithm to perform optimization traversal processing on the interval graph data to obtain path feature data, and using a dynamic programming algorithm to calculate the partition points of the path feature data to obtain an optimal window partition point set; Step S313, obtain the optimal window division point set, use a multi-scale decomposition algorithm to perform hierarchical division processing on the division point data to obtain hierarchical sequence data, use an adaptive window generation algorithm to perform window construction processing on the hierarchical sequence data to obtain window construction data, use a time series alignment algorithm to perform boundary optimization processing on the window construction data, and generate a multi-scale time window sequence.

Images