Multi-modal-based ocean ranching wind and light and heat storage multi-energy complementary scheduling method and system

Abstract

The application belongs to the technical field of optimal scheduling, and provides a marine pasture wind-light-diesel storage multi-energy complementary scheduling method and system based on multi-modal. The environmental modeling layer predicts the future state of the system according to the space-time correlation of meteorological data, the physical law of equipment operation and the periodic mode of load change. The reward function of the strategy generation and optimization layer is a multi-objective weighted reward, at least including a power supply reliability reward based on load loss probability, an energy economy reward based on diesel fuel consumption and energy abandonment cost, a device life reward based on device start-stop and loss, and a penalty term for frequency, voltage deviation and constraint violation. Through the construction of a perception, decision and execution closed-loop strategy, the fusion perception of multi-modal data and the dynamic intelligent decision under multi-objective constraints are realized, and the flexible collaborative control of multi-energy equipment is realized, so that the power supply reliability, operation economy, equipment service life and environmental adaptability of the island marine pasture energy system are improved.

Description

Technical Field

[0001] This invention belongs to the field of optimization scheduling technology, and in particular relates to a multimodal multi-energy complementary scheduling method and system for marine ranching based on wind, solar, diesel and storage. Background Technology

[0002] Traditional marine ranching energy systems mostly adopt a wind-solar-diesel-storage architecture, and their scheduling strategies mainly rely on empirical rules, simple logic control, or offline optimization algorithms.

[0003] The wind speed, sunlight, waves, and aquaculture load in the marine environment are highly random, intermittent, and periodic. Traditional preset rules are difficult to match the ever-changing energy supply and demand relationship in real time, leading to frequent power shortages or waste of renewable energy. In isolated mode, disconnected from the main power grid, frequent start-stop of diesel generators and drastic fluctuations in wind and solar power can easily cause system frequency and voltage instability, threatening the safety of aquaculture equipment. Existing methods struggle to achieve a dynamic optimal balance among multiple conflicting objectives such as ensuring power supply reliability, reducing diesel consumption, and extending the lifespan of critical equipment. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a multimodal, wind-solar-diesel-storage multi-energy complementary scheduling method and system for marine ranches. By constructing a closed-loop strategy for perception, decision-making, and execution, this invention achieves fusion perception of multimodal data, dynamic intelligent decision-making under multi-objective constraints, and flexible collaborative control of multiple energy devices. Ultimately, it improves the power supply reliability, operational economy, equipment lifespan, and environmental adaptability of the energy system for isolated island marine ranches, and can achieve a dynamic optimal balance among multiple conflicting objectives such as ensuring power supply reliability, reducing diesel consumption, and extending the lifespan of key equipment.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: In a first aspect, the present invention provides a multimodal, wind-solar-diesel-storage multi-energy complementary scheduling method for marine ranching, comprising: Acquire meteorological and environmental data, equipment status data, and load demand data for the marine ranch energy system; Based on the meteorological environment data, the equipment status data, and the load demand data, a hierarchical multimodal reinforcement learning model is used to generate scheduling strategies for each energy unit, including wind, solar, diesel, and energy storage. The hierarchical multimodal reinforcement learning model includes an environment modeling layer for learning system dynamics and a strategy generation and optimization layer for generating optimization strategies. The environment modeling layer predicts the future state of the system based on the spatiotemporal correlation of meteorological data, the physical laws of equipment operation, and the periodic patterns of load changes. The reward function of the strategy generation and optimization layer is a multi-objective weighted reward, with objectives including at least a power supply reliability reward based on the probability of load loss, an energy economy reward based on diesel fuel consumption and energy curtailment costs, an equipment lifespan reward based on equipment start-up, shutdown, and wear, and penalties for frequency, voltage deviations, and constraint violations.

[0006] Furthermore, the meteorological environmental data includes wind speed, light intensity, temperature, and wave height; the equipment status data includes diesel generator output power, operating hours, and cooling water temperature; the state of charge, charging and discharging current, battery temperature, and terminal voltage of the energy storage battery; and the conversion efficiency of the wind power converter and photovoltaic inverter; the load demand data includes the total power consumption of the aquaculture vessel, feeding system, oxygenation equipment, lighting, and monitoring instruments.

[0007] Furthermore, the environmental modeling layer is constructed based on a graph neural network to model the spatiotemporal correlation of meteorological data, the physical laws of equipment operation, and the periodic patterns of load changes, in order to predict the future state of the system and provide a virtual simulation environment for reinforcement learning training. The environmental modeling layer includes a dynamic adaptive graph structure learning mechanism, a physical constraint embedding layer, a multi-scale spatiotemporal attention mechanism, and a heterogeneous graph information aggregation module. The policy generation and optimization layer adopts a reinforcement learning algorithm based on deep deterministic policy gradient. The state space of the policy generation and optimization layer is composed of multi-source real-time measurement data and the global state code output by the environmental modeling layer. The action space of the policy generation and optimization layer includes the power regulation coefficients of wind turbines and photovoltaics, the start-stop and output commands of diesel generators, and the charging and discharging power commands of the energy storage system.

[0008] Furthermore, before the actions in the action space are executed, they need to be processed by an action corrector to perform physical constraint verification and correction based on real-time measurement values ​​obtained from multimodal data. The physical constraints include upper and lower limits of renewable energy power and ramp rate constraints, minimum start-stop time and output limit constraints of diesel generators, and state of charge and charging / discharging power constraints of energy storage systems.

[0009] Furthermore, the action space of the strategy generation and optimization layer is as follows: ; The physical constraints on the action include: Renewable energy power constraints: ; Diesel generator constraints: ; ; ; Constraints of energy storage systems: ; ; ; in, for The actual output power of renewable energy at any given time; for Available renewable energy power at any given time; for Real-time renewable energy reference power / dispatch instructions; for Maximum ramp rate limit for renewable energy; for The output power of the diesel generator at any given time; For diesel generator start / stop state variables; and These refer to the minimum and maximum technical output of the diesel generator, respectively. The duration of continuous operation and downtime of the diesel generator, respectively; and These are the minimum operating and downtime times for the diesel generator, respectively. This represents the maximum gradeability of the diesel generator. for The charging and discharging power of the energy storage system at all times. A positive value indicates discharge power. A negative value indicates charging power; and These represent the maximum charging / discharging power of the energy storage system; SOC t for Battery state of charge (SOC) at all times. min Minimum allowable SOC; SOC max Maximum SOC allowed; SOC t+1 The predicted SOC value for the next time step; η ch For charging efficiency; η dis For discharge efficiency; This refers to the rated capacity of the energy storage system.

[0010] Furthermore, the reward function for: ; ; ; Among them, R reliability Reliability reward; As an economic reward; Incentives based on equipment lifespan; This is a hard constraint penalty item; for Total power supply capacity at any time; Δf t λ1 is the frequency deviation penalty coefficient; λ2 is the voltage deviation penalty coefficient; for Power deficit at any moment; for Total load demand at any given time.

[0011] Furthermore, the energy economic incentive minimizes fuel costs and curtailment costs as follows: ; Among them, R economy For economic rewards; c fuel The unit cost of diesel fuel; for At any given time, the diesel generator output power is denoted as 'a'; the diesel generator fuel consumption rate is 'b'; the diesel generator starting fuel consumption is 'c'; and the constant term is 'no-load loss'. This is the operation and maintenance cost coefficient; This is the energy curtailment penalty coefficient; for Available renewable energy power at any given time; for Actual output power of renewable energy at any given time; for Constantly abandoning renewable energy power; equipment lifespan rewards and penalties include diesel engine start-stop, accelerated battery aging, and drastic power fluctuations: ; Among them, R lifetime Incentives based on equipment lifespan; This is the start-stop penalty coefficient; for The number of times it is started at any given moment; This is the energy storage lifetime penalty coefficient; K1 is the basic penalty coefficient; K2 is the exponential growth coefficient. for Current of the energy storage system at all times; This represents the absolute value deviating from the median charge level. This is the power fluctuation penalty coefficient; This represents the absolute value of the power change.

[0012] Furthermore, the objective function is: ; Among them, P wt ( P represents the real-time output power of the fan; ref This is a reference value for active power; w( ) represents the real-time rotational speed of the fan rotor; w opt ρ1 is the optimal rotational speed; ρ2 is the rotational speed deviation weighting coefficient; ρ2 is the pitch angle variation weighting coefficient.

[0013] Furthermore, a feedforward compensation term based on the battery model and the command change rate is added. : ; ; in, K represents the inverter control output voltage at time t. p K is the proportional gain coefficient; e(t) is the current tracking error; i Here, ∫e(t) dt is the integral gain coefficient; ∫e(t) dt is the error integral term. For feedforward control; R bat i is the battery's equivalent internal resistance; ref (t) represents the reference current at time t, and the expected charging / discharging current value; L_bat is the battery's equivalent inductance; The reference current change rate; t is the battery open-circuit voltage; SOC(t) is the battery state of charge.

[0014] Secondly, the present invention also provides a multimodal marine ranching wind-solar-diesel-storage multi-energy complementary scheduling system, comprising: The data acquisition module is configured to acquire meteorological and environmental data, equipment status data, and load demand data of the marine ranch energy system. The scheduling module is configured to: generate scheduling strategies for each energy unit (wind, solar, diesel, and storage) using a hierarchical multimodal reinforcement learning model based on the meteorological environment data, the equipment status data, and the load demand data; wherein the hierarchical multimodal reinforcement learning model includes an environment modeling layer for learning system dynamics and a strategy generation and optimization layer for generating optimization strategies; the environment modeling layer predicts the future state of the system based on the spatiotemporal correlation of meteorological data, the physical laws of equipment operation, and the periodic patterns of load changes; the reward function of the strategy generation and optimization layer is a multi-objective weighted reward, the objectives of which include at least a power supply reliability reward based on the probability of load loss, an energy economy reward based on diesel fuel consumption and energy curtailment costs, an equipment lifespan reward based on equipment start-up, shutdown, and losses, and penalties for frequency, voltage deviations, and constraint violations.

[0015] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the multimodal marine ranching wind-solar-diesel-storage multi-energy complementary scheduling method described in the first aspect.

[0016] Fourthly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the steps of the multimodal marine ranching wind-solar-diesel-storage multi-energy complementary scheduling method described in the first aspect.

[0017] Fifthly, the present invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the multimodal marine ranching wind-solar-diesel-storage multi-energy complementary scheduling method described in the first aspect.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention generates scheduling strategies for each energy unit (wind, solar, diesel, and storage) based on meteorological environmental data, equipment status data, and load demand data using a hierarchical multimodal reinforcement learning model. The hierarchical multimodal reinforcement learning model includes an environmental modeling layer for learning system dynamics and a strategy generation and optimization layer for generating optimization strategies. The environmental modeling layer predicts the future state of the system based on the spatiotemporal correlation of meteorological data, the physical laws of equipment operation, and the periodic patterns of load changes. The reward function of the strategy generation and optimization layer is a multi-objective weighted reward, with objectives including at least a power supply reliability reward based on the probability of load loss, an energy economy reward based on diesel fuel consumption and energy curtailment costs, an equipment lifespan reward based on equipment start-up, shutdown, and wear, and penalties for frequency, voltage deviations, and constraint violations. By constructing a closed-loop strategy encompassing perception, decision-making, and execution, this invention achieves fusion perception of multimodal data, dynamic intelligent decision-making under multi-objective constraints, and flexible collaborative control of multiple energy devices. Ultimately, it improves the power supply reliability, operational economy, equipment lifespan, and environmental adaptability of the isolated island marine ranch energy system, achieving a dynamic optimal balance among multiple conflicting objectives such as ensuring power supply reliability, reducing diesel consumption, and extending the lifespan of critical equipment. Attached Figure Description

[0019] The accompanying drawings, which form part of this embodiment, are used to provide a further understanding of this embodiment. The illustrative embodiments and their descriptions are used to explain this embodiment and do not constitute an improper limitation of this embodiment.

[0020] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention; Figure 2 This is a diagram of the overall architecture of a multi-energy complementary scheduling system for marine ranches, including wind, solar, diesel, and storage. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0023] Multi-energy complementary dispatch: By coordinating the joint operation of multiple energy sources (wind power, solar power, diesel power generation, and energy storage equipment), efficient energy utilization and stable energy supply can be achieved.

[0024] Island operation system: A power supply system that is independent of the main power grid and needs to rely on its own energy configuration to meet the electricity demand in the region.

[0025] Example 1: As marine ranches expand into deeper waters and on a larger scale, a stable, economical, and green energy supply has become a key bottleneck for their continued operation. Traditional marine ranch energy systems mostly adopt a wind-solar-diesel-storage architecture, and their scheduling strategies mainly rely on empirical rules, simple logic control, or offline optimization algorithms.

[0026] However, as described in the background section, wind speed, sunlight, waves, and aquaculture loads in marine environments exhibit strong randomness, intermittency, and periodicity. Traditional preset rules struggle to match the ever-changing energy supply and demand relationships in real time, leading to frequent power shortages or waste of renewable energy. In isolated island mode, disconnected from the main power grid, frequent start-ups and shutdowns of diesel generators and drastic fluctuations in wind and solar power can easily cause system frequency and voltage instability, threatening the safety of aquaculture equipment. Existing methods struggle to achieve a dynamic optimal balance among multiple conflicting objectives, such as ensuring power supply reliability, reducing diesel consumption, and extending the lifespan of critical equipment (e.g., diesel engines, energy storage batteries). Harsh conditions such as high humidity, high salt corrosion, and long-distance communication delays place higher demands on the robustness and fault tolerance of the control system.

[0027] S1, Multimodal Data Perception: S1.1 Meteorological and Environmental Data Sensing: Optionally, deploy anemometers, irradiance meters, temperature and humidity sensors, and wave monitors to collect wind speed data in real time. Light intensity ,temperature Wave height Data such as these are used to predict the available power of renewable energy sources.

[0028] S1.2, Equipment Status Data Sensing: Monitoring diesel generator output power Operating hours, cooling water temperature; monitoring the state of charge of the energy storage battery. Charging and discharging current Battery temperature Terminal voltage; monitor the operating status and conversion efficiency of wind power converters and photovoltaic inverters.

[0029] S1.3 Load demand data perception: Optionally, a smart meter and monitoring system can be used to collect the total power consumption of various loads, such as aquaculture boats, feeding systems, oxygenation equipment, lighting, and monitoring instruments, in real time. And its changing trends.

[0030] S2, Reinforcement Learning Intelligent Decision Making: Optionally, a hierarchical multimodal reinforcement learning framework (hierarchical multimodal reinforcement learning model) is adopted, taking multi-source data such as meteorological environmental data, equipment status data, and load demand data as input, and outputting the optimal scheduling command. The multi-source data is processed by the hierarchical multimodal reinforcement learning framework to output the optimal scheduling command. Specifically: S2.1 State Space Construction: The collected real-time data is integrated into a high-dimensional state vector. Specifically, the data used includes renewable energy data (current wind turbine output, current photovoltaic output, maximum available wind and solar power), energy storage data (current SOC, SOH health status, charging and discharging power), diesel engine data (current output, cumulative operating time, start-stop status), load data (current total load), grid data (current frequency, frequency deviation, bus voltage), and meteorological data (current wind speed, solar irradiance, temperature). In addition, a global state code z(t) output from the PCE-STGAT environmental modeling layer is received. This code is a high-dimensional feature vector generated by the graph neural network after spatiotemporal correlation and physical constraint fusion of all the above data. Finally, the state space S(t) is composed of the original measurements and the graph neural network code.

[0031] S2.2, Action Generation: The policy network takes the constructed state space S(t) as its sole input, performs forward computation through a deep neural network, and outputs the original action vector. The original actions include the renewable energy power regulation coefficients α(t) and β(t), the diesel engine output P_diesel(t) and start / stop state u_d(t), and the energy storage charging / discharging power P_ess(t). During the training phase, additional exploration noise is added to enhance the policy's exploration capabilities.

[0032] S2.3, Action Correction: Based on the original action, real-time measurements are directly obtained from multimodal data for physical constraint verification. The specific constraint verification data used includes the current SOC value of energy storage and its upper and lower limits, the current start-stop status and minimum operating time of the diesel engine, and the minimum / maximum output limits, the current grid frequency and voltage, the maximum available power of wind and solar power (used to calculate the dynamic lower limit of α_min), and the current load demand. The action corrector trims or adjusts the original action based on these real-time measurements to ensure that the corrected action strictly meets all safety and physical constraints, and then outputs the corrected feasible action A_final(t) to the execution module.

[0033] S2.4 Reward Calculation: Based on the actual execution results reported by the execution module, the system state after execution is obtained from multimodal data. Combined with the action instructions output by the action corrector, the comprehensive reward R(t) is calculated. The reward function comprehensively considers four dimensions: power supply reliability (quantified by LOLP load failure probability), energy economy (diesel fuel consumption and equipment operation and maintenance costs), equipment lifespan (energy storage charging and discharging loss costs), and system stability (frequency deviation squared term). This reward value is used for subsequent experience playback and policy network updates.

[0034] S2.5, Experience Storage and Policy Update: The current cycle's state S(t), action A_final(t), reward R(t), and next state S(t+1) are stored as a complete experience in the experience replay buffer. When the buffer's data volume reaches a set threshold, a batch of experiences is randomly sampled to calculate the loss function and update the policy network's parameters. The next state S(t+1) in the experience data is predicted by the PCE-STGAT environment modeling layer, enabling the agent to continuously learn and optimize in an offline environment. Its core design is as follows: S2.5.1, Environment Modeling Layer: Using graph neural networks to construct dynamic models of systems, such as Figure 1 As shown, this dynamic model can effectively learn the spatiotemporal correlation of meteorological data, the physical laws of equipment operation, and the periodic patterns of load changes, providing an accurate virtual simulation environment for reinforcement learning. The input of the graph neural network is current and historical multi-device data + inter-device connection relationships, and the output is future predicted values ​​+ system state codes.

[0035] The "future predictions" output by the graph neural network are predictions of key system state quantities at the next moment (or multiple future steps), specifically including the following three categories: Power grid operation state prediction: power grid frequency f(t+1), key bus voltage V(t+1), and power flow of each branch. Equipment state prediction: energy storage state of charge (SOC(t+1), actual output of each generation unit, and trends in equipment health status. External environment prediction: future short-term wind speed, solar intensity, and load demand trends.

[0036] The "future prediction" is passed to two downstream modules: the training environment for the reinforcement learning agent, where a graph neural network acts as a "virtual simulator" to replace the real environment. The agent takes action a(t) in the current state s(t), and the graph neural network predicts the next state s(t+1) and the immediate reward r(t), which the agent uses for policy learning. This allows the agent to undergo millions of trial-and-error training cycles in an offline environment without actually running physical equipment. An online MPC optimizer is also included.

[0037] The MPC optimizer uses a graph neural network to make rolling predictions of the system state over H time steps (e.g., one hour in advance, with each step being 5 minutes). Based on these predictions, it then solves an optimization problem within a finite time domain to find the optimal control sequence, enabling proactive control. For example, if it predicts a sudden increase in wind speed in 10 minutes, the MPC can proactively reduce diesel engine output and increase energy storage charging power to actively address disturbances rather than reacting passively.

[0038] (1) Dynamic adaptive graph structure learning mechanism: Traditional graph neural networks are usually based on fixed topology, while the equipment connection relationship and energy flow path of marine ranch energy system will change due to equipment failure, communication interruption and other situations.

[0039] This embodiment introduces a dynamic adjacency matrix learning module, which automatically learns the connection weights between nodes based on real-time system status (device availability, communication link status, etc.). Its input is the current features of each node, and its output is the connection weight matrix between nodes, used in graph convolution to determine how information is transmitted. The implementation formula is: ; in, Indicates time node With nodes Connection weights; For node features; For predefined edge attributes (such as electrical distance, communication delay). It is a learnable neural network.

[0040] (2) Physical Constraint Embedding Layer: During message passing, physical constraints of the energy system, such as the law of conservation of energy and equipment operating limits, are explicitly embedded. For power generation nodes, their output power is constrained to not exceed their rated capacity; for energy storage nodes, their charge / discharge rate and SOC range are constrained; for load nodes, supply and demand balance is ensured. Specifically, constraint satisfaction terms are added to the node state update function. The input is its old state and information from its neighbors, and the output is the new state, which is sent to the next layer or the final output layer, such as: ; in, Let i be the physical constraint mask for node i.

[0041] (3) Multi-scale spatiotemporal attention mechanism: To address the multi-scale characteristics of marine meteorological data (second-level fluctuations, minute-level trends, daily cycles, etc.), a multi-scale temporal convolution module is designed to extract features at different time scales. A spatial attention mechanism is introduced, allowing the model to focus on the regions and equipment that have the greatest impact on the system at the current moment. Spatiotemporal feature fusion: ; In this system, GAT stands for Graph Attention Network, and TCN stands for Temporal Convolutional Network. The input to the spatiotemporal feature fusion is the spatial result of the graph convolution and the temporal result of the temporal network, and the output is the fused features, which are finally fed into the prediction layer and the readout layer.

[0042] (4) Heterogeneous graph information aggregation: The marine ranch energy system contains various types of nodes (meteorological nodes, power generation nodes, energy storage nodes, load nodes, etc.) and various types of relationships (spatial proximity, electrical connection, energy flow, etc.). A heterogeneous message passing mechanism is designed, using different weight matrices for different types of nodes and relationships: ,in This represents the relationship type between nodes i and j. Its inputs are neighbor features and relationship type, and its output is relationship-specific weighted information, which is ultimately sent to the current node for summary and update. (5) Uncertainty Modeling and Robustness Enhancement: To address the uncertainty and noise in marine environmental data, Bayesian deep learning is introduced into the graph neural network to probabilistically model node representations and edge weights. Monte Carlo Dropout and other techniques are used to perform multiple samplings during testing to obtain the predicted distribution rather than a single-point estimate, thus enhancing the model's robustness. Its input is noisy data, and its output is a probability distribution (mean + variance). Its input is multi-source sensor data with measurement noise, communication errors, and missing data; its output is the probability distribution (mean and variance) of each predicted quantity (frequency, SOC, load, wind and solar power output, etc.). The decision module uses these distributions to make risk-sensitive conservative decisions. Finally, the data is sent to the decision module for risk assessment.

[0043] S2.5.2, Strategy Generation and Optimization Layer: An intelligent agent is constructed based on algorithms such as deep deterministic policy gradient or proximal policy optimization. Its key elements are carefully designed, and its state space is as follows:

[0044]

[0045] ; A comprehensive representation system at time The operational status.

[0046] The motion space is: ; This refers to the reference power command for each energy device. These actions are subject to strict physical constraints: Renewable energy power constraints: ; Diesel generator constraints (to prevent frequent start-stop and inefficient operation): ; ; ; Constraints of energy storage systems (ensuring safety and lifespan): ; ; ; in, for The actual output power of renewable energy at any given time; for Available renewable energy power at any given time; for Real-time renewable energy reference power / dispatch instructions; for Maximum ramp rate limit for renewable energy; for The output power of the diesel generator at any given time; This refers to the start / stop state variables (0 / 1 variables) of the diesel generator. =1 indicates that the diesel generator is in It is always running. =0 indicates that the diesel generator is in It is constantly in a shutdown state; and These refer to the minimum and maximum technical output of the diesel generator, respectively. The duration of continuous operation and downtime of the diesel generator, respectively; and These are the minimum operating and downtime times for the diesel generator, respectively. This represents the maximum gradeability of the diesel generator. for The charging and discharging power of the energy storage system at all times. A positive value indicates the discharge power (supply to the grid). A negative value indicates charging power (absorbing electrical energy from the grid), and the unit is usually kW or MW; and These represent the maximum charging / discharging power of the energy storage system; SOC t for Battery state of charge (SOC) at any given time (unit: %), range 0-100%, representing the percentage of remaining charge; min The minimum permissible SOC (in %) is typically 10%-20% to prevent deep discharge from damaging battery life; SOC max The maximum permissible SOC (in %) is typically 80%-95% to prevent overcharging from damaging battery life; SOCt+1 The predicted SOC value for the next time step (in %); η ch The charging efficiency (0 < η_ch ≤ 1, typically 0.85-0.95) includes AC / DC conversion losses and battery internal resistance losses; η dis The discharge efficiency is (0 < η_dis ≤ 1, typically 0.85-0.95). The rated capacity of an energy storage system is its nominal energy capacity, typically expressed in kWh or MWh; reward function. The design is a multi-objective weighted sum that guides the agent to learn the optimal policy.

[0047] ; ; ; Among them, R reliability Reliability bonus (negative values ​​represent penalties); Economic incentives reflect indicators that reflect operating costs, such as fuel costs and maintenance costs; For equipment lifespan rewards, indicators that take into account equipment degradation and extend service life are considered; This is a hard constraint penalty, a direct penalty for violating hard constraints (such as exceeding the SOC limit); for Total power supply capacity at any given time, i.e., the total power that all power sources (renewable + diesel + energy storage discharge) can provide, in kW / MW; Δf t For frequency deviation, Δft = ft - ft + 1, the difference between the actual frequency and the rated frequency is Δf(t) = f(t) - f_nominal; λ1 is the frequency deviation penalty coefficient, a weighting parameter that determines the importance of frequency fluctuation; λ2 is the voltage deviation penalty coefficient, the larger the value, the higher the requirement for frequency stability. for Power deficit at any time, P lack，t =max(0, P) load，t P supply，t ); for Real-time total load demand, directly penalizing instability and power outages, is the total power required by all users, reflecting electricity demand, in kW / MW.

[0048] The energy economics incentive minimizes fuel costs and energy curtailment costs, specifically: ; Among them, R economy For economic incentives (negative values ​​represent costs); c fuelThe unit cost of diesel fuel (unit: yuan / liter) includes procurement cost, transportation cost, and storage cost; for The output power of the diesel generator at time t is the actual output of the diesel generator at time t, in kW or MW; a is the diesel generator fuel consumption rate (in liters / kWh), which represents the amount of diesel fuel consumed to generate 1 kWh of electricity, usually 0.2-0.3 L / kWh, varying with the load rate; b is the diesel generator starting fuel consumption (in liters / cycle), including the extra consumption during preheating and starting, usually the fuel amount for 1-2 hours of rated operation; c is the constant term / no-load loss, which represents the basic fuel consumption of the generator when running under no-load conditions. There are losses even when no electricity is generated, and usually c>0. This is the operation and maintenance cost coefficient, representing the operation and maintenance cost per unit of electricity generated, including lubricating oil, regular maintenance, labor, etc., in yuan / kWh. This is the energy curtailment penalty coefficient, representing the penalty cost for curtailing 1 kWh of renewable energy. It is usually set relatively high to encourage consumption. The unit is yuan / kWh. It is usually set relatively high and takes precedence over fuel costs. for The available power of renewable energy at any given time is the maximum power that can theoretically be generated based on weather conditions, determined by wind and solar resources, and is expressed in kW / MW. for Real-time renewable energy output power, and the actual renewable energy power used by the system during dispatch, in kW / MW; for The amount of renewable energy power that is constantly being abandoned represents available but unused renewable energy sources, measured in kW / MW.

[0049] Equipment lifespan rewards and penalties include diesel engine start-stop, accelerated battery aging (simplified model example), and dramatic power fluctuations, specifically: ; Among them, R lifetime R is the equipment lifespan bonus (a negative value indicates lifespan loss). lifetime The larger the value (the smaller the negative value): the less impact on equipment lifespan, and the more gentle the operation mode; This is the start-stop penalty coefficient, in yuan / time, representing the lifespan loss cost corresponding to each start-up of the diesel generator; for The number of starts at time t indicates whether the diesel generator was started at time t, usually 0 or 1 (1 indicates start). This is the energy storage lifetime penalty coefficient, in yuan, which adjusts the overall weight of the energy storage lifetime penalty. K1 is the base penalty coefficient, the base of the penalty function, which adjusts the minimum value of the penalty; K2 is the exponential growth coefficient, which controls the rate at which the penalty increases with the depth of charge and discharge. The larger K2 is, the faster the penalty increases with the depth of charge and discharge. for Current of the energy storage system at any time, in A, representing the magnitude of the charging and discharging current and reflecting the charging and discharging rate; The absolute value of the deviation from the median charge level is 0, which represents SOC = 50% (optimal), and 0.5 represents SOC = 0% or 100% (worst). This is the power fluctuation penalty coefficient, in yuan / kW, representing the lifetime loss cost per unit of power change. This represents the absolute value of power change, reflecting the severity of power fluctuations.

[0050] S2.5.3, Online Learning and Rolling Optimization Layer: The intelligent agent continuously updates its decision network parameters online through interaction with the real environment or a high-fidelity simulation model. Combining the concept of model predictive control, it performs rolling optimization over a finite time domain starting from the current state, thereby achieving a rapid and robust response to sudden disturbances (such as equipment failure or sudden weather changes).

[0051] S3, Multi-functional Complementary Execution: This module receives instructions from the decision-making module and translates them into specific equipment actions: Renewable Energy Flexible Access Unit: Through the intelligent converter, according to... Performs precise power control, achieving maximum power point tracking and active power smoothing. Diesel generator coordinated control unit: based on... The generator controller adjusts fuel injection quantity and speed according to start / stop commands to ensure efficient and stable operation. Energy storage system optimization management unit: Based on... The power converter is controlled by the battery management system to perform charging and discharging operations that meet health constraints (such as current and SOC range).

[0052] This module is a layered, distributed, adaptive collaborative execution system that not only forwards instructions but also possesses local intelligence, fault-tolerant reconfiguration, and dynamic coordination capabilities. Its architecture includes: an instruction parsing and distribution layer, a local intelligent controller layer, and a device driver and protection layer.

[0053] S3.1, Instruction Parsing and Dispatch Layer: This layer receives a global optimization instruction set from the reinforcement learning decision module. Its core function is to perform instruction feasibility verification and dynamic decomposition. Dynamic security boundary verification: Before distribution, the instruction is compared with the real-time monitored system security boundaries. These boundaries are periodically reported by the local controller, taking into account equipment status (such as battery SOH, diesel engine coolant temperature) and environmental conditions (such as whether the real-time wind speed exceeds the fan cut-out limit). If the instruction exceeds the current dynamic security boundary, the instruction softening and reallocation algorithm is activated. .in, It is the feasible region formed by the dynamic safety boundaries of all devices at the current moment. This algorithm can compute an absolutely safe alternative instruction set that is closest to the original instruction within milliseconds, ensuring execution safety. Heterogeneous instruction compilation: This compiles unified optimization variables into heterogeneous control instructions that can be recognized by different devices.

[0054] For example: For a wind turbine, it is compiled as a torque setpoint. and propeller pitch angle reference value and satisfy Dynamic correlation constraints ensure the coordination of pneumatic-mechanical-electrical control. For photovoltaic inverters, this is compiled into an inner current loop reference value. It also embeds seamless switching logic between the maximum power point (MPPT) and the scheduled power point (SPPT) to achieve a smooth transition. For the energy storage PCS, it is compiled into a cascaded reference value of the power outer loop and the current inner loop, and an adaptive current limiting based on the battery's state of equilibrium (SOH) is added. .

[0055] S3.2, Local Intelligent Controller Layer: Each major energy unit (wind turbine, photovoltaic array, diesel generator, energy storage system) is equipped with a Local Intelligent Controller (LIC). The LIC is no longer a simple instruction executor, but an intelligent agent with model predictive control (MPC) capabilities and cooperative communication functions.

[0056] Local Model Predictive Control (Local-MPC): Each LIC contains a simplified and accurate model of the controlled device (such as a wind turbine drive train model, a diesel engine speed control system model, or a second-order RC equivalent circuit model of a battery). After receiving the setpoint command from the upper layer, it does not output directly, but instead runs a local MPC optimization once within a very short time window (such as the next 2-5 seconds) at a higher frequency (such as 100Hz).

[0057] The objective function (taking wind turbine LIC as an example) is: ; Constraints include safe operating speed range, pitch angle change rate limit, and generator torque limit. Among these, P... wt ( P represents the real-time output power of the wind turbine. At time t, it is the actual active power generated by the wind turbine, expressed in kW or MW. The selectable value range is 0 ≤ P. wt (t) ≤ P wt_max (t); P wt_max (t) represents the maximum available power at the current wind speed, P wt_max (t) = 0.5 * ρ * A * C p_max * V(t)³ is the variable that needs to be determined for MPC optimization; P ref The active power reference value is the power setpoint issued by the upper-level dispatching system, in kW or MW. It originates from the reinforcement learning decision output of the central dispatching layer and can be calculated based on system power balance and economic optimization. The update cycle is 1-5 minutes. ref ≤ P wt_max (t) (must be within the wind turbine's capacity); w( ) represents the real-time rotational speed of the wind turbine rotor, and the angular velocity of the wind turbine rotation, in rad / s or rpm; minimum speed w min (Typically 30%-40% of the rated speed), rated speed w rated Or maximum speed w max (Typically 110%-120% of rated speed), and its relationship with power is: P wt ( ) = T( ) * w( ), where T( ) represents the generator torque; w opt The optimal rotational speed, representing the best rotational speed for achieving maximum wind energy capture (MPPT) at the current wind speed, is expressed in rad / s or rpm. opt (t) =(λ opt * V(t)) / R, where λ opt For the optimal tip speed ratio (usually 6-8), V(t) is the real-time wind speed at the hub height (m / s), and R is the rotor radius (m), which is adjusted in real time according to the wind speed. Δβ is the change in pitch angle relative to the previous moment, expressed in degrees (°). ) = β( ) - β( -1), limiting mechanical stress, reducing wear of the pitch mechanism, smoothing power output, and reducing impact on the power grid. The rate of change constraint is usually limited to ±2-4° / s. ρ1 is the speed deviation weighting coefficient, which is the importance of the balanced speed optimization term in the objective function. It is dimensionless, and a typical value can be 0.1-0.3. Under normal circumstances, ρ1 = 0.2. When the wind speed is stable, increase ρ1 to emphasize MPPT. When the wind speed fluctuates greatly, decrease ρ1 to prioritize power tracking. ρ2 is the pitch angle change weighting coefficient, which limits the intensity of pitch angle change. It is dimensionless, and a typical value can be 0.05-0.15. ρ2 is smaller for new wind turbines (0.05-0.08), allowing for larger changes. ρ2 is larger for older wind turbines or when protection is required (0.1-0.15). Increase ρ2 in severe weather to strengthen mechanical protection.

[0058] Local-MPC achieves optimized and refined device-level control, which can smooth out disturbances at the second or even sub-second level (such as the impact of gusts on wind turbines), and transform macro-schedule commands into high-quality, high-dynamic-performance local device control signals. This greatly reduces the real-time computing pressure on the central decision-making level and improves the operating quality and lifespan of individual devices.

[0059] Each LIC exchanges critical status information (such as actual output and available adjustment margin) through a low-latency local communication network (such as industrial Ethernet based on time-sensitive networking). When an abnormal response or limit is detected in a device (such as energy storage), the LICs of adjacent devices with adjustment capabilities (such as diesel engines or another set of photovoltaics) can fine-tune their respective outputs within microseconds through a predefined consensus protocol, achieving local power balance self-healing without waiting for instructions from the cloud.

[0060] S3.3, Device Driver and Protection Layer: This layer is the final execution unit, responsible for converting the optimized control signals output by the LIC into power electronic switching signals or mechanical actuator actions. Its core innovation lies in the integration of feedforward adaptive closed-loop control and cross-device protection coordination.

[0061] Adaptive feedforward-feedback composite control: Taking the control of energy storage PCS as an example, its current loop controller not only adopts traditional PI feedback, but also adds a feedforward compensation term based on the battery model and the command change rate. : ; ; in, The inverter control output voltage at time t, in volts (V), is the control voltage command output by the converter, driving power switching devices such as IGBTs; K pK is the proportional gain coefficient, the proportional term coefficient of the feedback control, which is dimensionless (or V / A, depending on the unit of e(t)). Typical values ​​are 0.1-2.0. Increasing K... p Improving response speed may cause overshoot; reducing K... p Improved stability, but slower response; e(t) is the current tracking error, the difference between the reference current and the actual current, in amperes (A), e(t) = i ref (t) - i actual (t), where i ref (t) represents the reference current (from upper-level control), i actual (t) represents the actual measured current; K i is the integral gain coefficient, the integral term coefficient of the feedback control, with units of 1 / second (or V / (A·s)), and a typical selectable value of 10-100. It eliminates steady-state error and improves control accuracy. Integral saturation problems require anti-saturation treatment; ∫e(t) dt is the error integral term, the historical cumulative value of the difference, with units of ampere-second (A·s), in discrete form: ∫e(t)dt ≈ Σ e(k) * Δt, where Δt is the sampling period; This is a feedforward control (FF) term that compensates for known disturbances in advance, reducing the burden on feedback control. bat The equivalent internal resistance of a battery is the internal resistance exhibited by the battery during charging and discharging, measured in ohms (Ω). Typical values ​​are: lithium-ion batteries: 10-50 mΩ (cell); lead-acid batteries: 20-100 mΩ (cell). The internal resistance increases with decreasing temperature and with aging: internal resistance increases with aging. ref (t) is the reference current at time t, the expected charging and discharging current value, in amperes (A), positive values ​​are the discharging current, negative values ​​are the charging current, and it comes from the MPC optimization output of the LIC layer; L_bat is the battery equivalent inductance (equivalent series inductance), the equivalent inductance of the battery and its connecting lines, in Henry (H), the typical value can be in the microhenry level (μH), reflecting the back electromotive force effect when the current changes; , where is the reference current change rate, the expected rate of change of current over time, expressed in amperes per second (A / s), and is calculated using the following formula: =[i ref (t) - i ref The current (t-1)] / Δt is usually limited to 100-500 A / s to avoid excessive current surges; V is the open-circuit voltage of the battery, the terminal voltage of the battery under no-load conditions, in volts (V), and is a non-linear function of SOC. For lithium-ion batteries: V oc (SOC) = V min + (V max - Vmin ) * SOC^n, where V min The minimum discharge voltage (typically 2.5-3.0V), V max The maximum charging voltage is denoted by n (typically 4.1-4.2V), and n is the curve shape parameter (0.5-1.5). SOC(t) is the battery state of charge, the percentage of remaining battery capacity relative to the total capacity, expressed as a percentage (%). The update formula is: SOC(t) = SOC(t-1) - [i...]. bat [(t) * η] / (3600 * C), where i bat (t) represents the actual current (A), η represents the charge / discharge efficiency, and C represents the battery capacity (Ah). This significantly improves the tracking accuracy and dynamic response speed for fast power commands.

[0062] Cross-device protection coordination mechanism: Traditional protection is isolated; this invention designs system-level protection logic. For example, when a sharp drop in grid frequency is detected, not only will the diesel generators act according to the governor, but all wind turbines and photovoltaic LICs will simultaneously receive a "frequency emergency support" broadcast signal, based on a preset droop coefficient. Immediately increase active power output proportionally (for photovoltaic systems, this means temporarily exiting MPPT mode): ; in, For the first Emergency power regulation of the equipment refers to the additional or reduced active power that the equipment needs to provide during frequency emergencies, expressed in kilowatts (kW) or megawatts (MW). To indicate an increase in power generation or a decrease in load, A negative value indicates a reduction in power generation or an increase in load, with a response time requirement of ≤100 milliseconds; For the first The droop factor (also known as the frequency regulation factor) of a device represents the amount of power regulation caused by a unit frequency change. The larger the value, the greater the power adjustment for the same frequency change; the unit is kW / Hz or MW / Hz. Δf(t) is the system frequency deviation, the difference between the actual frequency and the rated frequency; a negative value indicates a frequency decrease, and a positive value indicates a frequency increase; the unit is Hertz (Hz). Δf(t) = f(t) - f n (t), the detection accuracy is usually required to be ±0.01Hz, and the dead zone is set to not operate within ±0.05Hz to avoid frequent adjustments. This global coordinated fast frequency response increases the equivalent inertia of the system by several times, which is a revolutionary means to ensure the transient stability of isolated power grids.

[0063] Workflow and fault-tolerant refactoring of the execution module: Normal mode: Cloud command → Command layer verification and distribution → Each LIC receives and executes Local-MPC → Device layer precise tracking.

[0064] Communication Interruption Mode: If a LIC loses contact with the cloud, it will switch to local autonomous mode. Based on the last valid command and local sensor data, a simplified rolling optimization scheduler will run to continue maintaining the device in a safe and economical range until communication is restored.

[0065] Equipment Failure Mode: If a device (such as a diesel engine) fails and exits, its LIC will immediately broadcast the failure information. The instruction parsing layer will dynamically reconstruct the system's feasible domain. The system will then quickly recalculate the emergency dispatch plan. Meanwhile, the LICs of other devices will temporarily assume more adjustment responsibilities through consistency coordination.

[0066] This embodiment realizes intelligent optimization under safety constraints. By encoding complex equipment physical constraints and operating conditions into the action space and reward function of reinforcement learning, the generated scheduling strategy inherently meets engineering safety requirements, solving the reliability and security problems of AI algorithms in critical infrastructure applications.

[0067] A multi-objective dynamic optimal balance has been achieved. The designed composite reward function quantifies and unifies multiple objectives such as reliability, economy, and equipment lifespan within a single optimization framework. The system can automatically adjust its strategy focus according to different operating conditions (such as typhoon season and peak aquaculture season), achieving comprehensive optimization both globally and in the long term. It is expected to increase the renewable energy consumption rate by more than 15% and reduce diesel consumption by 20%-30%.

[0068] It significantly improves the lifespan of critical equipment by proactively managing diesel engine start-stop and battery health status through a reward function, achieving "preventative" scheduling of equipment wear and tear. It can effectively reduce ineffective diesel engine start-stops by more than 50% and extend the cycle life of energy storage batteries by 15%-20%, significantly reducing operation and maintenance costs.

[0069] With strong environmental adaptability and robustness, the system can continuously adapt to environmental changes and equipment aging in marine ranches based on online learning and rolling optimization mechanisms. The modular design also facilitates the integration of new equipment such as wave energy and fuel cells, and its strong scalability makes it particularly suitable for deep-sea applications with harsh environments and ever-changing requirements.

[0070] The system's reliability and availability are greatly enhanced. The distributed consistency coordination and cross-device protection and collaboration mechanism enable the system to maintain basic operation through the autonomous cooperation of other devices when some devices fail or communication is partially interrupted. This achieves decentralized robustness and completely changes the vulnerability of traditional centralized control systems, which are prone to complete system failure.

[0071] It achieves flexible and intelligent execution. The execution module has a series of intelligent functions such as dynamic security verification, instruction softening, and fault-tolerant reconfiguration, ensuring that any instruction from the decision-making layer is implemented in the safest and smoothest way. This allows the upper-level reinforcement learning algorithm to explore a wider policy space more "boldly" without being overly concerned about execution risks, thereby discovering a globally superior scheduling strategy.

[0072] In one specific embodiment, a heterogeneous fusion system with three levels of collaboration and a two-way closed loop—cloud-based intelligent optimization, edge-based collaborative control, and terminal-based precise execution—was constructed. For example... Figure 2 The diagram shown is the overall architecture of the marine ranch wind-solar-diesel-storage multi-energy complementary scheduling system. The main components are: The system's hierarchical structure and functional positioning adopt a vertically layered and horizontally collaborative design approach, decomposing complex global optimization problems into processing units with different time scales and spatial ranges, thus achieving a balance between "centralized decision-making for global optimality" and "distributed execution for speed and reliability".

[0073] The cloud-based intelligent layer (decision optimization center) is the system's "intelligent brain," deployed in a shore-based data center or a central control platform with highly reliable communication guarantees. Its core responsibility is to perform global non-real-time optimization across devices and time periods. This layer does not directly deal with millisecond-level fluctuations in devices, but rather performs strategy calculations based on macroscopic states on a minute-level cycle.

[0074] Hardware platform: This typically consists of a high-performance server cluster, including compute nodes (equipped with GPUs for deep model training and inference), data storage nodes (for storing massive amounts of historical operational data and model parameters), and communication gateway nodes. Considering potential long-distance communication bottlenecks in marine ranches, this layer can also be deployed on a "regional edge cloud" closer to the ranch to reduce latency.

[0075] The core software runs the core algorithm engine of this invention, including: a multimodal data fusion engine: receiving and cleaning massive, heterogeneous, and asynchronous raw data from lower layers. For example, it timestamps, unifies units, and verifies the quality of wind speed data sampled at the second level, light intensity data sampled at the minute level, and power grid frequency data sampled at the millisecond level, eliminating noise and outliers to form a consistent system state snapshot.

[0076] The improved PCE-STGAT dynamic environment model is a "digital twin" of the system. It maps physical entities such as wind turbines, photovoltaic arrays, diesel engines, energy storage batteries, load points, and meteorological monitoring points to nodes in a graph neural network, and maps electrical connections, energy flow, and spatial correlations to edges. Through "physical constraint embedding" and "dynamic graph structure learning," this model can accurately simulate the spatiotemporal propagation effects of wind and solar resources, the coupled evolution of equipment states, and the power flow characteristics of the power grid, providing a high-fidelity simulation environment for upper-level decision-making.

[0077] Constrained Integrated Deep Reinforcement Learning Agent: Through interaction with the aforementioned environmental model, this agent learns the optimal scheduling strategy. Its action space directly corresponds to the control commands of each energy device (such as the wind turbine output coefficient α), and its reward function comprehensively quantifies multiple objectives such as stability, economy, and reliability. The key action corrector module ensures that any strategy it outputs pre-satisfies hard physical and engineering constraints such as the minimum start-stop time of the diesel engine and the safe range of energy storage SOC, thereby guaranteeing the inherent feasibility of the decision.

[0078] Model Predictive Control Online Optimizer: As the final optimization stage, it uses the policy output by the reinforcement learning agent as the initial value and performs higher-precision rolling optimization within a short time window (e.g., the next hour). Utilizing the predictive capabilities of PCE-STGAT, it proactively considers recent trends in wind, solar, and load changes, fine-tuning scheduling commands to achieve a smoother and more economical control trajectory.

[0079] The edge control layer (regional collaborative execution hub) is a key innovation in the architecture, serving as a bridge between "intelligence" and "physicality." It consists of intelligent region controllers distributed near the energy nodes of the marine ranch. Each LIC is an industrial-grade embedded device with strong local computing, storage, and communication capabilities.

[0080] Hardware features: Wide temperature range, corrosion resistant, high protection rating (e.g., IP67), suitable for high humidity and high salinity marine environments. Built-in multi-core processor, supporting real-time operating system.

[0081] Core Function: Command Translation and Fine-grained Execution: Receives minute-level macro-level commands from the cloud (e.g., "Adjust the total output of the wind turbine group to 500kW"), and decomposes and translates them into millisecond-level micro-level commands executable by local equipment (e.g., specific pitch angles and generator torque settings for each wind turbine). This process is not a simple forwarding; rather, it utilizes a built-in local model predictive control algorithm to perform secondary optimization at the equipment level, ensuring optimal dynamic response to track commands from higher levels and mitigating second-level local disturbances not considered by the higher-level model (e.g., gusts of wind encountered by a single wind turbine).

[0082] Distributed consensus coordination: Each LIC communicates directly with other LICs via a low-latency local industrial network (such as Ethernet based on time-sensitive networking). When a LIC detects that the equipment it manages has reached its output limit or has failed, it can quickly negotiate with neighboring LICs through a pre-defined consensus protocol to autonomously adjust the output distribution of adjacent areas, achieving instantaneous rebalancing of local power. This "edge-to-edge collaboration" mechanism constitutes the first line of self-healing defense for the system in the event of interruption or delay of cloud commands.

[0083] Data acquisition and preprocessing: High-frequency acquisition of local sensor data, filtering, compression and feature extraction, and then uploading to the cloud to reduce the pressure on core network bandwidth.

[0084] The physical device layer (energy conversion and sensing terminals) is the "limbs" and "sensors" of the system, including all physical devices that generate, store, and consume energy, as well as the sensor network that senses the environment.

[0085] Wind turbine generator set: Equipped with an advanced pitch and speed control system, it can receive power or speed commands from the LIC and achieve seamless switching from maximum power point tracking mode to power limiting mode.

[0086] Photovoltaic power generation array: Through intelligent inverters, it can achieve flexible control of maximum power point tracking and constant power output.

[0087] Diesel generator set: integrates digital speed governor and automatic synchronization device, which can quickly adjust output or execute start and stop according to instructions.

[0088] Energy storage unit: Battery energy storage system, the core of which is the battery management system and power conversion system. BESS can not only execute charging and discharging power commands, but also report its health status to the system, providing a basis for cloud-based optimization of lifespan cost calculation.

[0089] Load Units: These encompass various electrical equipment used in aquaculture production (such as feeders and aerators), environmental control, and logistical support. Critical loads may have simple communication interfaces to report their operational plans or priorities.

[0090] Multi-dimensional sensor network: This is the physical foundation for the system to achieve "multimodal perception." It includes: A meteorological sensor network: measuring wind speed, wind direction, light intensity, temperature, humidity, waves, etc., providing environmental input for wind and solar power prediction and safe system operation. An electrical measurement network: installing high-precision power metering devices at key busbars, lines, and equipment outlets to monitor voltage, current, power, frequency, harmonics, etc. in real time. An equipment condition monitoring network: monitoring the mechanical and electrical health status of major equipment such as wind turbine gearboxes, diesel engines, and transformers through vibration, temperature, and acoustic sensors.

[0091] Closed-loop design of data flow, control flow, and coordination flow; uplink data flow (sensing flow): indicated by blue solid arrows, with the direction from bottom to top.

[0092] Path: Raw data generated by sensors and smart meters at the physical device layer → collected by the nearest LIC and pre-processed (such as compression and packaging) → uploaded to the cloud data fusion platform via the backbone communication network (fiber optic / 5G).

[0093] Content includes raw sampling data (such as instantaneous wind speed), equipment status information (such as diesel engine operating hours), alarm signals (such as energy storage overheating), etc. This data is the sole source of information for building accurate environmental models, assessing system status, and performing optimization calculations in the cloud.

[0094] Downward control flow (decision flow): Represented by dashed arrows, the direction is from top to bottom.

[0095] Path: The cloud-based intelligent layer calculates and generates an optimized instruction set for the next scheduling cycle → This set is then sent to the corresponding regional LICs via the communication network → The LICs parse the instructions, perform local optimizations based on local real-time information, and then drive the specific actuators (such as frequency converters, circuit breakers, and speed controllers) to operate.

[0096] Content: The instructions are highly structured, containing information such as target values, execution time, priority, and safety boundaries. For example, an instruction sent to the wind turbine LIC might be "Maintain the average active power at 320kW for the next 5 minutes, and the instantaneous power is not allowed to exceed 350kW."

[0097] Horizontal collaborative flow (autonomous flow): Indicated by a purple double arrow, it occurs inside the edge control layer.

[0098] Path and Content: This is the key design element for achieving high reliability in this architecture. Adjacent LICs (such as the LIC managing photovoltaics and the LIC managing energy storage) directly exchange critical status information (such as "the output of photovoltaics in this area is about to drop sharply" or "the SOC of this energy storage is approaching its limit") via the local network. Based on a pre-defined distributed consensus algorithm, they can autonomously negotiate and adjust their output within milliseconds, collaboratively responding to localized sudden changes without waiting for a slower response from the cloud (typically on the order of seconds or minutes). This effectively solves the system vulnerability problem of centralized control in the event of communication interruptions or cloud failures.

[0099] The system workflow and typical scenario demonstration: Within a standard scheduling cycle (e.g., 5 minutes), the system works collaboratively according to the following process: Cycle Start-up and Data Aggregation: At the start of a new cycle, all LICs synchronously collect detailed operating data and sensor readings from local devices over the past 5 minutes, package them, and upload them to the cloud.

[0100] Global Status Assessment and Prediction: The cloud-based data fusion platform completes data cleaning and alignment. The PCE-STGAT model uses this data to update the "digital twin" status of the entire system and predict wind and solar power and load trends for the next few hours.

[0101] Intelligent Decision Generation: Based on the latest system state and historical experience, the DRL agent proposes a preliminary scheduling strategy. After the physical feasibility of this strategy is verified by the "action corrector," it is handed over to the MPC optimizer for rolling optimization, ultimately generating an optimized instruction sequence that balances economy, stability, and executability.

[0102] Command issuance and distributed execution: Optimize command issuance to each LIC. Instead of rigid execution, the LIC activates its local fast MPC controller to transform the 5-minute power target value into the optimal control trajectory for the devices in the next few seconds (such as a smooth power ramp-up curve), and drives the devices to execute. Simultaneously, the LIC continuously monitors local conditions and collaborates with its neighbors.

[0103] Closed-loop feedback and learning: The actual results after the device executes the commands, along with any anomalies that occur during the process, are treated as new data and fed back to the cloud in the next cycle. This data is stored in an experience base and used to periodically (e.g., every night) retrain the DRL agent offline, enabling its strategy to continuously evolve and adapt to device aging and seasonal changes.

[0104] Typical scenario – a sudden surge in wind power caused by strong winds: Sensing: When the weather sensors and the wind turbine itself detect a sharp increase in wind speed, the relevant LIC immediately reports the "power surge warning" as high-priority data to the cloud, and at the same time notifies the nearby energy storage LIC and load LIC through horizontal collaborative flow.

[0105] Edge autonomous response (millisecond level): Upon receiving an early warning and coordination request, the energy storage LIC immediately activates its maximum power charging mode to absorb excess power; nearby adjustable load LICs (if any) temporarily increase some load according to the agreement. This quickly mitigates the power surge in the local power grid.

[0106] Cloud-based intelligent rescheduling (seconds / minutes): The cloud receives alerts almost simultaneously, and the PCE-STGAT model quickly simulates the impact of the event. The DRL agent, combining the current energy storage charging status and load adjustment capabilities, recalculates and issues new global optimization commands within seconds. For example, it commands some wind turbines to actively limit power (adjusting the α coefficient), gradually reducing the energy storage charging power to an economical level and restoring the original load.

[0107] System stabilized: Through the close cooperation of "edge rapid autonomy" and "cloud optimization and readjustment", the system smoothly weathered the disturbance, avoiding protection actions caused by excessive frequency or overvoltage, and making the most of wind energy.

[0108] The core advantages and engineering value of the architecture Figure 2 The architecture presented is not simply a collection of components, but embodies the following deep engineering principles: A balance between decoupling and aggregation: Decoupling computationally intensive, non-real-time global optimization (cloud) from computationally lightweight, high-real-time local control (edge), and connecting them through standardized interfaces, reduces system complexity and communication bandwidth requirements. Resilient design: By endowing the edge layer with considerable autonomy, the system possesses multi-level resilience, enabling "cloud failure, edge autonomy; single point of failure, neighbor assistance," making it particularly suitable for the harsh communication conditions of deep-sea environments. Elastic expansion: Adopting a modular design. When the marine ranch expands and new power generation units are added, only a new LIC needs to be deployed and connected to the network, and the corresponding node added to the cloud model; the overall system architecture does not need to be completely redesigned. Technical compatibility: The edge controller (LIC) can interface with existing equipment of various brands and models through standard industrial protocols, protecting existing user investments and enabling smooth intelligent upgrades.

[0109] This embodiment explicitly embeds physical and engineering constraints such as energy conservation, upper and lower limits of equipment power, and minimum diesel engine operating time into the message passing and loss function of the PCE-STGAT model. This makes the learned dynamic model and decision-making strategy inherently conform to physical laws, fundamentally ensuring the safety and feasibility of the decision results, and solving the pain points of "unexplainable and untrustworthy" traditional AI methods in critical infrastructure. The improved graph neural network can adapt to changes in system topology (such as equipment switching and communication interruptions) and continuously adapt to equipment aging and environmental changes through online fine-tuning. Compared with traditional methods based on fixed models, the generalization ability and long-term adaptability of this system are significantly improved.

[0110] This embodiment introduces a dynamic power regulation lower limit (α_i^min(t)) and a spinning reserve constraint, enabling the system to proactively maintain regulation margin when wind and solar power are abundant, transforming the traditional "generation first, regulation later" approach into "generation and regulation coordination." This is equivalent to providing the system with a built-in, zero-response preventative buffer capability, suppressing fluctuation impacts at the source. The cross-device protection coordination mechanism in the execution module can instantly coordinate wind turbines, photovoltaics, diesel engines, and energy storage to provide power support during sudden frequency drops, increasing the system's equivalent inertia and primary frequency regulation capability to several times that of traditional solutions, and revolutionizing the transient stability of the islanded power grid.

[0111] This embodiment unifies power supply reliability (LOLP), operational economy (fuel and maintenance costs), and equipment lifespan loss (battery degradation cost C_loss) into a comprehensive goal that can be maximized. The agent automatically balances "generating more power now" with "extending the equipment's lifespan for several more years" during decision-making. By establishing and updating the lifespan loss model of the energy storage equipment online (related to charge / discharge power and depth), and incorporating it as a dynamic cost item into real-time optimization, proactive maintenance of energy storage health is achieved. Simultaneously, by optimizing the diesel engine's operating range and reducing its inefficient operation and start-stop frequency, the diesel engine's lifespan is extended.

[0112] The distributed local intelligent controller (LIC) in this embodiment possesses edge computing and local autonomy capabilities. In the event of a cloud communication interruption, it can maintain safe operation based on the last instruction and local information through a built-in simplified MPC, achieving a leap in architectural resilience from "central cerebral palsy leading to total paralysis" to "partial brain damage still allowing for self-care." The system features dynamic security boundary verification and instruction soft redistribution mechanisms. When a device failure is detected or a safety limit is reached, the feasible domain can be immediately reconstructed and the output of each device redistributed. Combined with distributed consistency coordination among LICs, second-level self-healing is achieved, ensuring continuous power supply under extreme disturbances.

[0113] In summary, this invention reconstructs the technical chain of "precise physical modeling - constraint integrated decision-making - adaptive collaborative execution," achieving a paradigm shift from "passive, local, single-objective, and fragile traditional control" to "active, global, multi-objective, and resilient intelligent autonomous scheduling." This precisely meets the stringent requirements of deep-sea marine ranches for efficient, stable, reliable, and long-life operation of energy systems.

[0114] Example 2: This embodiment provides a multimodal marine ranching wind-solar-diesel-storage multi-energy complementary scheduling system, including: The data acquisition module is configured to acquire meteorological and environmental data, equipment status data, and load demand data of the marine ranch energy system. The scheduling module is configured to: generate scheduling strategies for each energy unit (wind, solar, diesel, and storage) using a hierarchical multimodal reinforcement learning model based on the meteorological environment data, the equipment status data, and the load demand data; wherein the hierarchical multimodal reinforcement learning model includes an environment modeling layer for learning system dynamics and a strategy generation and optimization layer for generating optimization strategies; the environment modeling layer predicts the future state of the system based on the spatiotemporal correlation of meteorological data, the physical laws of equipment operation, and the periodic patterns of load changes; the reward function of the strategy generation and optimization layer is a multi-objective weighted reward, the objectives of which include at least a power supply reliability reward based on the probability of load loss, an energy economy reward based on diesel fuel consumption and energy curtailment costs, an equipment lifespan reward based on equipment start-up, shutdown, and losses, and penalties for frequency, voltage deviations, and constraint violations.

[0115] The working method of the system is the same as that of the multimodal marine ranching wind-solar-diesel-storage multi-energy complementary scheduling method in Example 1, and will not be repeated here.

[0116] Example 3: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the multimodal marine ranching wind-solar-diesel-storage multi-energy complementary scheduling method described in Embodiment 1.

[0117] Example 4: This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the program, it implements the steps of the multimodal marine ranching wind-solar-diesel-storage multi-energy complementary scheduling method described in Embodiment 1.

[0118] Example 5: This embodiment provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the steps of the multimodal marine ranching wind-solar-diesel-storage multi-energy complementary scheduling method described in Embodiment 1.

[0119] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A multimodal multi-energy complementary scheduling method for marine ranching based on wind, solar, diesel, and storage, characterized in that, include: Acquire meteorological and environmental data, equipment status data, and load demand data for the marine ranch energy system; Based on the meteorological environment data, the equipment status data, and the load demand data, a hierarchical multimodal reinforcement learning model is used to generate scheduling strategies for each energy unit, including wind, solar, diesel, and energy storage. The hierarchical multimodal reinforcement learning model includes an environment modeling layer for learning system dynamics and a strategy generation and optimization layer for generating optimization strategies. The environment modeling layer predicts the future state of the system based on the spatiotemporal correlation of meteorological data, the physical laws of equipment operation, and the periodic patterns of load changes. The reward function of the strategy generation and optimization layer is a multi-objective weighted reward, with objectives including at least a power supply reliability reward based on the probability of load loss, an energy economy reward based on diesel fuel consumption and energy curtailment costs, an equipment lifespan reward based on equipment start-up, shutdown, and wear, and penalties for frequency, voltage deviations, and constraint violations.

2. The multimodal multi-energy complementary scheduling method for marine ranching based on wind, solar, diesel, and storage as described in claim 1, characterized in that, The meteorological environment data includes wind speed, light intensity, temperature, and wave height; the equipment status data includes diesel generator output power, operating hours, and cooling water temperature, energy storage battery state of charge, charging and discharging current, battery temperature, and terminal voltage, as well as the conversion efficiency of wind power converters and photovoltaic inverters; the load demand data includes the total power consumption of aquaculture vessels, feeding systems, oxygenation equipment, lighting, and monitoring instruments.

3. The multimodal multi-energy complementary scheduling method for marine ranching based on wind, solar, diesel, and storage as described in claim 1, characterized in that, The environmental modeling layer is constructed based on a graph neural network to model the spatiotemporal correlation of meteorological data, the physical laws of equipment operation, and the periodic patterns of load changes, in order to predict the future state of the system and provide a virtual simulation environment for reinforcement learning training. The environmental modeling layer includes a dynamic adaptive graph structure learning mechanism, a physical constraint embedding layer, a multi-scale spatiotemporal attention mechanism, and a heterogeneous graph information aggregation module. The policy generation and optimization layer adopts a reinforcement learning algorithm based on deep deterministic policy gradient. The state space of the policy generation and optimization layer is composed of multi-source real-time measurement data and the global state code output by the environmental modeling layer. The action space of the policy generation and optimization layer includes the power regulation coefficients of wind turbines and photovoltaics, the start-stop and output commands of diesel generators, and the charging and discharging power commands of energy storage systems.

4. The multimodal multi-energy complementary scheduling method for marine ranching based on wind, solar, diesel, and storage as described in claim 3, is characterized in that, Before execution, actions in the action space must pass through an action corrector to perform physical constraint verification and correction based on real-time measurement values ​​obtained from multimodal data. The physical constraints include upper and lower limits of renewable energy power and ramp rate constraints, minimum start-stop time and output limit constraints of diesel generators, and state of charge and charging / discharging power constraints of energy storage systems.

5. The multimodal multi-energy complementary scheduling method for marine ranching based on wind, solar, diesel, and storage as described in claim 1, characterized in that, The action space of the strategy generation and optimization layer is: ； The physical constraints on the action include: Renewable energy power constraints: ； Diesel generator constraints: ； ； ； Constraints of energy storage systems: ； ； ； in, for The actual output power of renewable energy at any given time; for Available renewable energy power at any given time; for Real-time renewable energy reference power / dispatch instructions; for Maximum ramp rate limit for renewable energy; for The output power of the diesel generator at any given time; For diesel generator start / stop state variables; and These refer to the minimum and maximum technical output of the diesel generator, respectively. The duration of continuous operation and downtime of the diesel generator, respectively; and These are the minimum operating and downtime times for the diesel generator, respectively. This represents the maximum gradeability of the diesel generator. for The charging and discharging power of the energy storage system at all times. A positive value indicates discharge power. A negative value indicates charging power; and These represent the maximum charging / discharging power of the energy storage system; SOC t for Battery state of charge (SOC) at all times. min Minimum allowable SOC; SOC max Maximum SOC allowed; SOC t+1 The predicted SOC value for the next time step; η ch For charging efficiency; η dis For discharge efficiency; This refers to the rated capacity of the energy storage system.

6. The multimodal multi-energy complementary scheduling method for marine ranching based on wind, solar, diesel, and storage as described in claim 5, is characterized in that, reward function for: ； ； ； Among them, R reliability Reliability reward; As an economic reward; Incentives based on equipment lifespan; This is a hard constraint penalty item; for Total power supply capacity at any time; Δf t λ1 is the frequency deviation penalty coefficient; λ2 is the voltage deviation penalty coefficient; for Power deficit at any moment; for Total load demand at any given time.

7. The multimodal multi-energy complementary scheduling method for marine ranching based on wind, solar, diesel, and storage as described in claim 6, characterized in that, The energy economic incentive minimizes fuel costs and curtailment costs as follows: ； Among them, R economy For economic rewards; c fuel The unit cost of diesel fuel; for At any given time, the diesel generator output power is denoted as 'a'; the diesel generator fuel consumption rate is 'b'; the diesel generator starting fuel consumption is 'c'; and the constant term is 'no-load loss'. This is the operation and maintenance cost coefficient; This is the energy curtailment penalty coefficient; for Available renewable energy power at any given time; for Actual output power of renewable energy at any given time; for Constantly abandoning renewable energy power; equipment lifespan rewards and penalties include diesel engine start-stop, accelerated battery aging, and drastic power fluctuations: ； Among them, R lifetime Incentives based on equipment lifespan; This is the start-stop penalty coefficient; for The number of times it is started at any given moment; This is the energy storage lifetime penalty coefficient; K1 is the basic penalty coefficient; K2 is the exponential growth coefficient. for Current of the energy storage system at all times; This represents the absolute value deviating from the median charge level. This is the power fluctuation penalty coefficient; This represents the absolute value of the power change.

8. The multimodal multi-energy complementary scheduling method for marine ranching based on wind, solar, diesel, and storage as described in claim 1, characterized in that, The objective function is: ； Among them, P wt ( P represents the real-time output power of the fan; ref This is a reference value for active power; w( ) represents the real-time rotational speed of the fan rotor; w opt ρ1 is the optimal rotational speed; ρ2 is the rotational speed deviation weighting coefficient; ρ2 is the pitch angle variation weighting coefficient.

9. The multimodal multi-energy complementary scheduling method for marine ranching based on wind, solar, diesel, and storage as described in claim 1, characterized in that, Add a feedforward compensation term based on the battery model and command change rate. : ； ； in, K represents the inverter control output voltage at time t. p K is the proportional gain coefficient; e(t) is the current tracking error; i Here, ∫e(t) dt is the integral gain coefficient; ∫e(t) dt is the error integral term. For feedforward control; R bat i is the battery's equivalent internal resistance; ref (t) represents the reference current at time t, and the expected charging / discharging current value; L_bat is the battery's equivalent inductance; The reference current change rate; t is the battery open-circuit voltage; SOC(t) is the battery state of charge.

10. A multimodal marine ranching wind-solar-diesel-storage multi-energy complementary scheduling system, characterized in that, include: The data acquisition module is configured to acquire meteorological and environmental data, equipment status data, and load demand data of the marine ranch energy system. The scheduling module is configured to: generate scheduling strategies for each energy unit (wind, solar, diesel, and storage) using a hierarchical multimodal reinforcement learning model based on the meteorological environment data, the equipment status data, and the load demand data; wherein the hierarchical multimodal reinforcement learning model includes an environment modeling layer for learning system dynamics and a strategy generation and optimization layer for generating optimization strategies; the environment modeling layer predicts the future state of the system based on the spatiotemporal correlation of meteorological data, the physical laws of equipment operation, and the periodic patterns of load changes; the reward function of the strategy generation and optimization layer is a multi-objective weighted reward, the objectives of which include at least a power supply reliability reward based on the probability of load loss, an energy economy reward based on diesel fuel consumption and energy curtailment costs, an equipment lifespan reward based on equipment start-up, shutdown, and losses, and penalties for frequency, voltage deviations, and constraint violations.

Images