FSRU re-gasification hybrid heat source scheduling method based on load prediction and rule learning

Abstract

This invention relates to the field of intelligent scheduling technology for FSRU heat sources, specifically disclosing a method for scheduling FSRU regasification hybrid heat sources based on load forecasting and rule learning. The method includes: collecting basic operational data of the FSRU, marine environmental data, and coupled correlation data, and performing data cleaning and standardization preprocessing; inputting the preprocessed data into a three-dimensional coupled load forecasting model; dynamically switching the network structure complexity of the model based on real-time collected marine environmental parameters; and outputting the load forecast value and deviation probability. This invention, by constructing a three-dimensional coupled environment-load-equipment model and introducing a dynamic switching mechanism, can intelligently select simplified or complex network structures according to real-time sea conditions, significantly improving the prediction accuracy for nonlinear environmental impacts while ensuring computational efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent scheduling technology for FSRU heat sources, and in particular to a scheduling method for FSRU regasification hybrid heat sources based on load forecasting and rule learning. Background Technology

[0002] Floating storage and regasification units (FSRUs) are key equipment for receiving and processing liquefied natural gas (LNG) at sea, operating for extended periods in complex marine environments characterized by high salt spray, high humidity, and variable wind and waves. Their regasification process relies on a hybrid heat source system, such as steam, gas, or combined heat and power (CHP), to provide thermal energy to convert liquefied natural gas into gaseous form and transport it to the pipeline network.

[0003] Existing FSRU heat source scheduling typically employs static scheduling logic based on terrestrial industrial standards or simple PID feedback control. This means that heat source output is passively adjusted solely based on current natural gas export demand, lacking in-depth coupling analysis of the critical variable of the marine environment. This traditional real-time response model often assumes that the external environment is relatively stable or exists as a simple disturbance, ignoring the nonlinear effects of sea state on pipeline pressure, equipment corrosion rate, and thermal efficiency.

[0004] However, in actual operation, FSRU faces a highly challenging technical conflict between dynamic environmental impacts and the mismatch between static scheduling strategies.

[0005] Specifically, the marine environment exhibits significant non-stationarity and concept drift characteristics. Parameters such as wind waves, swells, and temperature variations not only fluctuate dramatically, but their statistical distribution patterns also fundamentally change with seasonality or sudden weather events. Existing fixed models or single control strategies cannot maintain robustness under unknown extreme sea conditions, often leading to severe distortions in load forecasts.

[0006] More importantly, heat source equipment has huge thermal inertia, while marine environmental data is often accompanied by high-frequency turbulent noise. This leads to a sharp contradiction between the need for rapid safety response and the protection against thermal fatigue of equipment: if the system adjusts sensitively to environmental fluctuations, the equipment life will be drastically reduced due to frequent temperature changes; if the system is too smooth and lagging, it may not be able to respond in time when sudden wind and waves impact the pipeline network, causing safety accidents. Summary of the Invention

[0007] This invention aims to at least partially solve one of the technical problems in related technologies. Therefore, the objective of this invention is to propose a FSRU regasification hybrid heat source scheduling method based on load forecasting and rule learning, in order to achieve comprehensive optimization of FSRU operational safety, equipment durability, and energy utilization efficiency in complex marine environments.

[0008] To achieve the above objectives, a first aspect of the present invention proposes a method for scheduling FSRU regasification hybrid heat sources based on load forecasting and rule learning, comprising the following steps:

[0009] Collect basic operational data, marine environmental data, and coupled correlation data of the FSRU, and perform data cleaning and standardization preprocessing;

[0010] The preprocessed data is input into the three-dimensional coupled load prediction model. Based on the real-time collected marine environmental parameters, the network structure complexity of the model is dynamically switched, and the load prediction value and deviation probability are output.

[0011] An environmental state vector containing current marine environmental parameters and equipment status is input into a deep Q-network adapted to the marine environment to generate heat source scheduling rules.

[0012] Based on the load forecast, deviation probability, and heat source scheduling rules, and combined with the reliability of the remaining power of new energy sources, advance scheduling and real-time scheduling decisions are executed.

[0013] The dynamic switching process of the three-dimensional coupled load prediction model includes:

[0014] In response to real-time collected wind and wave level data and salt spray concentration data, the current marine environment category is determined. If the wind and wave level is less than or equal to a first intensity threshold and the salt spray concentration is less than a first concentration threshold, a simplified structural model is invoked for inference. If the wind and wave level is greater than the first intensity threshold, or the salt spray concentration is greater than or equal to the first concentration threshold, or the ocean temperature change rate exceeds a preset temperature change threshold, a complex structural model is invoked for inference. The complex structural model adopts a weighted time decay residual structure long short-term memory network combined with a graph attention network architecture.

[0015] To achieve the above objectives, a second aspect of the present invention proposes a FSRU regasification hybrid heat source scheduling system based on load forecasting and rule learning, comprising:

[0016] The data acquisition and preprocessing module is used to acquire basic operational data, marine environmental data, and coupled related data of the FSRU, and to perform data cleaning and standardization preprocessing.

[0017] The dynamic load forecasting module is used to input preprocessed data into a three-dimensional coupled load forecasting model, dynamically switch the network structure complexity of the model based on real-time collected marine environmental parameters, and output load forecast values ​​and deviation probabilities.

[0018] The intelligent rule generation module is used to input the environmental state vector containing the current marine environment parameters and equipment status into a deep Q-network adapted to the marine environment to generate heat source scheduling rules.

[0019] The hybrid scheduling execution module is used to perform pre-schedule and real-time scheduling decisions based on the load forecast value, deviation probability and heat source scheduling rules, combined with the reliability of the remaining power of new energy sources.

[0020] The dynamic load forecasting module is configured to perform the following operations when performing dynamic model switching:

[0021] In response to real-time collected wind and wave level data and salt spray concentration data, the current marine environment category is determined. If the wind and wave level is less than or equal to a first intensity threshold and the salt spray concentration is less than a first concentration threshold, a simplified structural model is invoked for inference. If the wind and wave level is greater than the first intensity threshold, or the salt spray concentration is greater than or equal to the first concentration threshold, or the ocean temperature change rate exceeds a preset temperature change threshold, a complex structural model is invoked for inference. The complex structural model adopts a weighted time decay residual structure long short-term memory network combined with a graph attention network architecture.

[0022] To achieve the above objectives, a third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the above-described FSRU regasification hybrid heat source scheduling method based on load forecasting and rule learning.

[0023] The FSRU regasification hybrid heat source scheduling method based on load forecasting and rule learning in this invention constructs a three-dimensional coupled model of environment-load-equipment and introduces a dynamic switching mechanism. It can intelligently select simplified or complex network structures according to real-time sea conditions, which significantly improves the prediction accuracy of nonlinear environmental impacts while ensuring computational efficiency.

[0024] In response to the non-stationary nature of the marine environment, this invention utilizes an online dynamic calibration mechanism based on distributed drift detection, which enables the system to automatically adjust and optimize target weights when encountering unseen extreme conditions, thus ensuring the robustness of the system.

[0025] Furthermore, by implementing an asymmetric peak-locked damping control strategy, the system successfully resolved the control conflict between high-frequency environmental fluctuations and equipment thermal inertia. While ensuring rapid pre-emptive response to severe sea conditions to guarantee supply continuity, it effectively suppressed damage to the equipment caused by ineffective oscillations, achieving the overall optimization of FSRU's operational safety, equipment durability, and energy utilization efficiency in complex marine environments. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the FSRU regasification hybrid heat source scheduling method based on load forecasting and rule learning provided by the present invention.

[0027] Figure 2This is a schematic diagram illustrating the implementation of the FSRU regasification hybrid heat source scheduling system based on load forecasting and rule learning provided by the present invention.

[0028] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0029] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0030] The following description, with reference to the accompanying drawings, describes an embodiment of the FSRU regasification hybrid heat source scheduling method, system, and electronic equipment based on load prediction and rule learning.

[0031] Example 1:

[0032] This embodiment provides a method for scheduling a hybrid heat source for FSRU (Floating Storage Regasification Unit) regasification based on load forecasting and rule learning. This method is configured for use in the FSRU's energy management controller, which serves as the core computing unit and can comprehensively manage the shipboard steam heat source, gas heat source, cogeneration system, and connected renewable energy systems. This embodiment aims to address the problem that traditional scheduling methods cannot simultaneously consider environmental non-stationarity, equipment thermal inertia, and load forecasting accuracy in complex marine environments.

[0033] like Figure 1 As shown, the method in this embodiment includes the following steps:

[0034] Specifically, the method in this embodiment first performs the following steps: collecting basic operational data of the FSRU, marine environmental data, and coupled correlation data, and then performing data cleaning and standardization preprocessing. This process is the foundation for building a high-precision prediction model.

[0035] For example, regarding the acquisition of basic operational data, the energy management controller acquires historical regasification load data, steam heat source supply capacity parameters, gas heat source combustion efficiency parameters, real-time output power of the cogeneration heat source, pipeline topology data, and operational status data of key equipment in real time through the industrial control bus within the FSRU. The pipeline topology data includes pipeline connections, valve status, and the physical attributes of each node. Equipment operational status data covers the inlet and outlet temperatures of the intermediate medium vaporizer (IFV), seawater pump speed, etc.

[0036] For example, this embodiment introduces a high-frequency sampling environmental sensing system for collecting marine environmental data. This system includes a shipborne weather station, marine environmental monitoring buoys deployed around the FSRU, and dedicated marine-grade sensors installed on key parts of the ship's hull. The collected data specifically includes wave intensity, instantaneous wind speed, salt spray concentration, ocean surface temperature, equipment corrosion current density, and pipeline strain values. Among these, the equipment corrosion current density is monitored in real time by an electrochemical corrosion sensor to characterize the degree of corrosion of metal components by a high-humidity, high-salt environment.

[0037] For example, the collection of coupled and correlated data mainly refers to the mapping data between wind waves and pipeline pressure, and the mapping data between salt spray and equipment failure rate generated in real time through edge computing nodes.

[0038] It is important to note that, to ensure data transmission stability under harsh sea conditions, this embodiment employs a specially designed data transmission link. Specifically, the multi-dimensional data is collected using a lightweight messaging protocol. The preferred lightweight messaging protocol used here is the MQTT-SN protocol, which maintains high data throughput even in low-bandwidth and unstable network environments. Furthermore, a dual-link backup channel is utilized for transmission, simultaneously establishing a satellite communication link and a maritime broadband communication link. Cyclic redundancy check bits are added to the core data during encapsulation, and when the primary link signal weakens, it automatically and seamlessly switches to the backup link, ensuring zero data loss.

[0039] After data acquisition is completed, the energy management controller will perform data cleaning and standardization preprocessing.

[0040] The first step in preprocessing is time alignment. Because different sensors have different sampling frequencies—for example, meteorological data might be on the order of seconds, while equipment operation data might be on the order of minutes—the system performs time alignment on the collected data, unifying data from different sampling periods to a standard time interval. This standard time interval can be set to 1 minute. Through timestamp matching and linear interpolation algorithms, time discrepancies between multi-source data are eliminated.

[0041] The second step in preprocessing is to generate the coupling characteristics of wind and waves and pipeline pressure. These coupling characteristics are calculated based on wind and wave intensity and pipeline strain values, and are used to characterize the degree of influence of the marine environment on pipeline pressure. Specifically, the coupling characteristic is defined as the product of wind and wave intensity and a first environmental adaptation coefficient, plus the product of pipeline strain value and a second environmental adaptation coefficient. This characteristic value can intuitively reflect the physical pressure fluctuations caused by external sea state impacts on the internal pipeline network.

[0042] The third step in preprocessing is anomaly removal. Based on statistical anomaly detection criteria and a marine environmental trend judgment mechanism, the system removes invalid outliers and retains valid abrupt changes caused by wind, waves, or salt spray. The preferred statistical anomaly detection criterion here is the 3-Sigma criterion, used to identify outliers deviating from a normal distribution. However, considering that sudden increases in wind and waves in the marine environment are valid abrupt changes rather than sensor malfunctions, this embodiment introduces a marine environmental trend judgment mechanism. This mechanism combines the wind speed change rate within adjacent time windows; if the data abrupt change conforms to meteorological evolution patterns, it is retained, thereby avoiding the accidental deletion of crucial environmental characteristic information.

[0043] After data preprocessing is completed, the method execution steps of this embodiment are as follows: input the preprocessed data into the three-dimensional coupled load prediction model, dynamically switch the network structure complexity of the model based on the real-time collected marine environmental parameters, and output the load prediction value and the deviation probability.

[0044] Specifically, the core of this step lies in the dynamic switching mechanism to balance computational efficiency and prediction accuracy. The energy management controller responds to real-time collected wind and wave level data and salt spray concentration data to determine the current marine environmental category.

[0045] For example, the system presets a first intensity threshold (e.g., level 3 wind and waves), a first concentration threshold (e.g., 30 milligrams per cubic meter), and a preset temperature change threshold (e.g., a change of 5 degrees Celsius every ten minutes).

[0046] The specific switching logic is as follows: If the real-time monitored wind and wave level is less than or equal to the first intensity threshold, and the salt spray concentration is less than the first concentration threshold, the system determines that the current environment is stable, and then calls the simplified structural model for inference. This simplified structural model can be a pruned Long Short-Term Memory (LSTM) network, whose input features only include historical load, temperature, and equipment status, resulting in fast inference speed and low energy consumption.

[0047] Specifically, if the wind and wave level is greater than the first intensity threshold, such as a force 8 wind and wave; or the salt spray concentration is greater than or equal to the first concentration threshold, such as 80 milligrams per cubic meter; or the ocean temperature change rate exceeds the preset temperature change threshold, the system determines that the current environment is an extreme environment and then calls the complex structure model for reasoning.

[0048] It is important to note that the complex structure model employs a weighted time decay residual structure Long Short-Term Memory (WTD-RS-LSTM) network combined with a Graph Attention Network (GAT) architecture. WTD-RS-LSTM is used to process time-series data; the weighted time decay mechanism is introduced to amplify the impact of recent extreme weather data on the prediction results, avoiding over-reliance on long-term historical stable data. The residual structure addresses the gradient vanishing problem in deep networks. The Graph Attention Network (GAT) is used to process the spatial characteristics of the pipeline network topology data. Through an attention mechanism, it dynamically calculates the mutual influence weights between nodes in the pipeline network, such as valves and pump stations, thereby accurately capturing the impact of local pipeline network pressure fluctuations on the overall load.

[0049] It is also important to note that the training process of the three-dimensional coupled load prediction model is a key technical aspect of this embodiment. The training process includes: constructing an input feature set containing historical load data, marine environmental characteristics, equipment health characteristics, and pipeline topology characteristics. Subsequently, the model is trained using a loss function. This loss function consists of a mean squared error term for the predicted values, a prediction bias term caused by wind and waves, and a prediction weight term corresponding to equipment corrosion risk.

[0050] Specifically, the mathematical expression of this loss function is as follows: ;

[0051] in, This represents the predicted load value. This represents the actual load value. Mean square error, Indicates the wind and wave level. This represents the absolute value of the prediction deviation. Indicates the corrosion current density of the equipment. Indicates a specific penalty factor, and These are the corresponding weighting coefficients.

[0052] By introducing a prediction bias term caused by wind and waves, i.e. The model is strongly constrained by strong wind and wave data during training, thus being forced to learn the nonlinear disturbance of sea state on the load; by introducing a prediction weight term corresponding to the equipment corrosion risk, i.e. The model tends to generate predictions that reduce the load on equipment with high corrosion risk.

[0053] Furthermore, to determine the proportions of the aforementioned items in the loss function, this embodiment uses historical marine environmental data to determine the weighting coefficients of the prediction bias term and the prediction weight term. The system performs offline analysis of meteorological data and FSRU operation logs from the past few years for this sea area, and uses regression analysis to determine the influence weights of wind, waves, and corrosion factors on load prediction accuracy in different seasons, thereby setting the initial values ​​of the aforementioned weighting coefficients.

[0054] Based on accurate load forecasting, the method in this embodiment then proceeds to the following steps: inputting an environmental state vector containing current marine environmental parameters and equipment status into a deep Q-network (DQN) adapted to the marine environment to generate heat source scheduling rules.

[0055] Specifically, the first step is to construct a multidimensional environmental state vector. The state vector contains at least:

[0056] ;

[0057] in, For wind and wave levels, Salt spray concentration, For equipment health, Forecasted load values This refers to the remaining electricity generated by new energy sources.

[0058] Device health It is a comprehensive indicator, calculated based on salt spray concentration, cumulative operating time, and measured corrosion current density. For example:

[0059] ;

[0060] in, Initial health value, This is the salt spray erosion coefficient. For runtime, As a corrosion current correction term, this state vector comprehensively reflects the current condition and status of the FSRU.

[0061] Specifically, the action space of DQN is defined. This action space includes pipeline pressure pre-regulation, wind and wave heat source combination switching, high-corrosion equipment load limiting, and new energy and grid coordinated energy replenishment actions. Pipeline pressure pre-regulation refers to adjusting the opening of pressure reducing valves in advance based on predicted load fluctuations; wind and wave heat source combination switching refers to prioritizing the switching to a more stable steam heat source during periods of high wind and wave intensity, reducing reliance on vibration-sensitive gas turbines; high-corrosion equipment load limiting refers to proactively reducing the flow setpoint of a gasifier when excessive corrosion current is detected, thus slowing the corrosion process; and new energy and grid coordinated energy replenishment refers to deciding when to utilize shipborne photovoltaic or wind power, and when to connect to shore power or start a diesel generator.

[0062] To guide DQN in learning the optimal strategy, a carefully designed reward function is required. This embodiment calculates the reward value of scheduling actions based on the reward function. The reward function is configured to be negatively correlated with load deviation rate, energy cost, equipment corrosion aggravation value, and pipeline pressure fluctuation value, and positively correlated with the supply continuity coefficient. This means that if the strategy generated by DQN leads to a large deviation between actual load and demand, high fuel costs, rapid equipment corrosion, or severe pipeline pressure fluctuations, the model will receive a significant negative penalty; conversely, if the strategy can ensure a continuous and stable supply of natural gas under extreme conditions, i.e., a high supply continuity coefficient, the model will receive a positive reward. Through extensive iterative training, the DQN agent will learn how to minimize costs and equipment damage while ensuring supply continuity under harsh sea conditions.

[0063] After generating the scheduling rules and obtaining the load forecast values, the method execution steps in this embodiment are as follows: based on the load forecast values, deviation probability and heat source scheduling rules, combined with the reliability of the remaining power of new energy sources, perform pre-scheduling and real-time scheduling decisions.

[0064] For example, in advance scheduling, the system aims to leverage the system's thermal inertia or response delay characteristics to proactively counteract environmental fluctuations. Specific operations include: acquiring marine weather forecast data for a preset time period, such as 1 to 6 hours in the future; subsequently, determining the execution time of the advance operations based on the mapping relationship between wind and wave levels and scheduling lead time; and simultaneously, determining the buffer margin for heat source output based on the mapping relationship between wind and wave levels and heat source output buffer capacity.

[0065] The mapping relationship is set as follows: as the wind and wave level increases, the time length of the scheduling lead time and the proportion of the heat source output buffer increase. For example, when the forecast wind and wave level is 3, the scheduling lead time may be set to 5 minutes, and the heat source output buffer, i.e., the reserved extra heat power reserve, may be set to 5%; while when the forecast wind and wave level increases to 8, the scheduling lead time increases to 20 minutes, and the heat source output buffer increases to 20%. This positively correlated mapping mechanism ensures that the system has preheated sufficient heat sources and increased pipeline pressure before severe sea conditions arrive, preventing supply interruptions due to response lag.

[0066] For example, this embodiment emphasizes reliability assessment for the utilization of new energy sources. System execution steps: Combining the reliability of the remaining power of the new energy source, pre-scheduling and real-time scheduling decisions are performed. Specifically, historical remaining power data of the new energy source and marine environmental parameters are input into a gated cyclic unit (GRU) model to calculate the reliability of the remaining power of the new energy source. The GRU model excels at capturing long-term dependencies in time series data. Based on current sunlight, wind speed, and the historical discharge curve of the battery, it can predict the stability of new energy power supply over a future period and output a probability value between 0 and 1, which is the reliability.

[0067] The specific decision-making logic is as follows: if the confidence level is greater than or equal to the confidence threshold, for example, 0.7, then the system generates a scheduling instruction to prioritize the use of new energy sources to drive deep heat source transmission, that is, to use more clean energy to drive seawater pumps or electric heaters; if the confidence level is less than the confidence threshold, it indicates that the supply of new energy sources is unstable, and the system generates a scheduling instruction to activate grid backup, that is, to start backup diesel generators in advance or apply for increased shore power to fill potential energy gaps.

[0068] Considering that FSRUs typically exist in cluster or multi-node form, or to leverage the experience of similar vessels, the method in this embodiment also includes distributed parameter aggregation and model optimization through a marine-adapted federated learning module.

[0069] Specifically, to adapt to the communication-constrained marine environment, the system dynamically adjusts the aggregation period based on the current wind and wave levels and salt spray concentration. In stable environments, a first aggregation period (e.g., 12 hours) is set to reduce communication overhead; in harsh or extreme environments, a second aggregation period shorter than the first (e.g., 3 hours) is set. The purpose of shortening the aggregation period is to enable faster sharing of coping strategies learned by each FSRU terminal when the environment changes rapidly, thus achieving rapid model iteration.

[0070] Specifically, during parameter transmission, a differentiated aggregation strategy is adopted for the embedded parameters of the graph neural network nodes in the complex structural model. This strategy retains the weight distribution of highly correlated edges with transmission efficiency exceeding a preset proportion. This means that only the network node parameters that have the greatest impact on the prediction results are transmitted, while the edge parameters with less impact are sparsified, thereby significantly reducing the amount of data transmitted.

[0071] After model parameter aggregation, the system performs a consistency check. A robust loss function-based consistency check mechanism is used to verify the aggregated model parameters. This robust loss function combines KL divergence with an environmental shock bias term. Local model parameters are only updated when the aggregated global model outperforms the local model on the local validation set, and no serious bias due to environmental differences occurs. This prevents malicious or faulty data from a single node from contaminating the global model.

[0072] Finally, as the bottom line for the safety of the entire scheduling system, the method in this embodiment also includes performing lightweight classification and graded response for anomalies specific to the marine environment.

[0073] Specifically, the system uses a decision tree model to identify anomaly types. Compared to deep learning models, decision trees offer significantly higher reasoning speed and interpretability. The anomaly types include environmental mutations (such as sudden increases in wind and waves, or sudden drops in temperature), equipment corrosion (such as gradual corrosion or sudden leaks), and coupled faults (such as abnormal pipeline pressure caused by wind and waves combined with equipment failure).

[0074] Based on the identified anomaly type, the system executes a tiered response:

[0075] If a high-priority anomaly is identified, such as a pipeline leak or a surge in wind and wave intensity exceeding level 3 within one minute, the system will trigger an emergency heat source combination and adjust the opening of pipeline valves within a first preset time (e.g., 100 milliseconds). This emergency heat source combination may include activating all standby boilers at full power, while valve adjustment involves rapidly isolating the damaged pipe section.

[0076] If a medium-priority anomaly is identified, such as a sudden increase in salt spray concentration leading to a greater risk of equipment corrosion, the system will reduce the load on high-risk equipment and activate redundant equipment within a second preset time (e.g., 1 to 5 seconds), which is longer than the first preset time. This provides the system with a buffer period to smoothly switch over and avoid unnecessary downtime.

[0077] It's also important to note that the system implements closed-loop feedback: it records the mapping relationship between environmental parameters, anomaly types, scheduling actions, and processing effects, and updates the anomaly response strategy library at preset intervals, such as every 24 hours. This allows the system to learn from each anomaly event and continuously improve its decision tree classification rules and response strategies.

[0078] In summary, this embodiment constructs a complete FSRU regasification heat source scheduling system through refined multi-dimensional data acquisition and processing, dynamic prediction using a three-dimensional coupled model adapted to the marine environment, intelligent rule generation based on reinforcement learning, hybrid scheduling execution considering equipment and environmental constraints, and federated learning optimization and anomaly hierarchical response. This method not only solves the failure problem of static scheduling strategies in dynamic marine environments but also effectively balances the contradiction between rapid response and equipment protection, possessing extremely high industrial application value.

[0079] Example 2:

[0080] This embodiment, based on Embodiment 1, further refines the optimization strategy for the three-dimensional coupled load prediction model. Specifically, this embodiment focuses on the technical challenges that may be encountered in actual operation of the step of determining the weighting coefficients of the prediction deviation term and the prediction weight term through fitting historical marine environmental data, and elaborates in detail on the specific implementation process of the online dynamic calibration step based on distribution drift detection.

[0081] It is important to note that in Example 1, the system primarily relies on static weighted coefficients derived offline from historical data to balance prediction accuracy with environmental constraints. This method performs well when the marine environment is in a normal distribution or undergoes periodic fluctuations. However, the real marine environment exhibits significant non-stationary characteristics and concept drift. For example, during a severe typhoon, although the wave intensity may be the same as a historical storm, its statistical distribution characteristics, such as swell period, wave shear, or gust frequency, may be completely different. This unknown operating condition with similar values ​​but different distributions can cause the static coefficients determined based on historical experience to become invalid, preventing the model from perceiving the potential high risks of the current environment and blindly pursuing the minimum mean square error of the load forecast, thus creating potential safety hazards. To address this, this example introduces an online dynamic calibration mechanism, aiming to enable the FSRU scheduling system to have self-examination and crisis response capabilities.

[0082] Specifically, after determining the weighting coefficients of the prediction bias term and prediction weight term by fitting historical marine environmental data, the method of this embodiment will initiate an independently running background process to perform an online dynamic calibration step based on distribution drift detection. The core logic of this step is to measure the difference between the current environment and the historical experience environment in real time using statistical distance, and dynamically adjust the weight allocation in the loss function accordingly.

[0083] First, in order to capture changes in environmental distribution, the system first performs the operation of establishing a real-time observation window containing data from the most recent preset time period, and a benchmark reference window containing historical typical working condition data.

[0084] For example, the real-time observation window is defined as a first-in, first-out (FIFO) sliding data queue that stores high-frequency marine environmental data collected by the FSRU within the most recent hour. The data update frequency of this window is consistent with the sensor's sampling frequency to ensure the capture of the latest environmental dynamics. For instance, this window stores wave level data and its first derivative, i.e., the wave level change rate, from the most recent 3,600 sampling points. The wave level change rate is chosen as the core monitoring indicator because, compared to wave height alone, the change rate better reflects the turbulence and destructive power of sea conditions, and is also a major cause of model prediction failure.

[0085] For example, the baseline reference window is a relatively static dataset that stores typical operating condition data accumulated by the FSRU during its historical operation. To ensure the effectiveness of the comparison, the system retrieves several data segments from the historical database that are most similar to current seasonal information and basic sea state categories, and aggregates these data segments into the baseline reference window. This window represents the normal distribution of marine environments that the model has seen and adapted to during the training phase.

[0086] After establishing two comparison windows, the system performs an operation to calculate the distribution deviation index of the rate of change of wind and wave levels between the real-time observation window and the benchmark reference window using the bulldozer distance algorithm.

[0087] It's also important to note that the bulldozer distance algorithm used here is commonly known in mathematics as the Wasserstein distance. Compared to traditional distribution difference measures such as KL divergence or JS divergence, the Wasserstein distance still provides a meaningful distance metric when dealing with probability distributions where two support sets do not overlap or have minimal overlap. In extreme marine environments, the real-time observed wind and wave rate of change distribution may deviate completely from the historical baseline distribution. In such cases, KL divergence might result in infinite values ​​or become incalculable, while the Wasserstein distance can accurately quantify the minimum work required to transform the current distribution pattern into the historical baseline distribution pattern, i.e., the sum of the mass of movement multiplied by the distance moved.

[0088] Specifically, the system first performs histogram statistics on the wind and wave level change rate data within the real-time observation window and the baseline reference window, constructing two discrete probability distributions. Let the probability distribution generated by the real-time observation window be the real-time distribution law P, and the probability distribution generated by the baseline reference window be the baseline distribution law Q. The system calls a linear programming solver or approximation algorithm to calculate the Wasserstein distance between these two distributions, and the result is defined as the distribution deviation index. This index is a non-negative real number; the smaller the value, the closer the statistical characteristics of the current sea state are to historical experience; the larger the value, the more unfamiliar the current sea state, the more severe the concept drift phenomenon, and the higher the unknown risks faced by the model.

[0089] After obtaining the distribution bias index, the system needs to convert it into specific adjustment instructions for the model weights. Therefore, the system performs the operation of constructing an adaptive gain function based on the distribution bias index and calculating the real-time gain factor.

[0090] Specifically, the adaptive gain function is designed to implement a non-linear triggering mechanism. Under normal sea conditions, the distribution deviation index typically fluctuates within a low range, at which point the system should remain silent to avoid interfering with the normal operation of the model. However, once the deviation index exceeds the safety limit, the gain function should rapidly output a high-gain signal. To this end, this embodiment sets a safe drift threshold. This threshold is determined based on the distribution deviation statistics during the FSRU's historical fault-free operation, for example, by taking the historical deviation mean plus three standard deviations.

[0091] For example, the adaptive gain function It is configured to output a value greater than one when the distribution deviation index exceeds the safe drift threshold. The mathematical logic of this function can be described as follows:

[0092] ;

[0093] in, The input is the distribution bias index. For the safe drift threshold, Here, the sensitivity coefficient is... It is a preset positive number used to control the intensity of the system's response to environmental anomalies.

[0094] Specifically, the output value calculated by this function is called the real-time gain factor. This factor is a dynamically changing scalar, and its physical meaning represents the risk multiplication factor of the current environment relative to historical experience environments. For example, in a long-period surge event that has never been recorded before, the calculated distribution deviation index may far exceed the safe drift threshold, causing the real-time gain factor to reach 1.5 or even higher.

[0095] After obtaining the real-time gain factor, the method in this embodiment enters the core execution stage, that is, using the real-time gain factor to perform real-time multiplication correction on the weighting coefficients of the prediction deviation term and the prediction weight term to generate dynamic execution coefficients.

[0096] Specifically, recalling the weighting coefficients mentioned in Example 1, these coefficients were determined through fitting historical data during the offline phase, and we call them static baseline coefficients. They reflect the proportions that wind and wave deviations and equipment corrosion risks should account for in the loss function in an average sense. However, during the online operation phase, the system no longer directly uses these static baseline coefficients, but instead uses parameters corrected by the real-time gain factor.

[0097] The system multiplies the static baseline coefficient with the real-time gain factor calculated at the current moment, and the result is the dynamic execution coefficient. If the current environment is stable and the real-time gain factor is 1, the dynamic execution coefficient is equal to the static baseline coefficient, and the model maintains its original optimization strategy; if the current environment is abnormal and the real-time gain factor is greater than 1, the dynamic execution coefficient will be significantly greater than the static baseline coefficient.

[0098] Finally, the system feeds back the dynamic execution coefficients to the loss function, dynamically increasing the weight of the prediction bias caused by wind and waves and the risk of equipment corrosion in the model optimization objective when the FSRU encounters an unstable marine environment that does not appear in historical data.

[0099] Specifically, this feedback mechanism fundamentally changes the behavior of the three-dimensional coupled load prediction model during the fine-tuning or online inference phases. When the weights of the wind and wave deviation term and the corrosion risk term in the loss function are dynamically increased, the model will find that reducing the prediction error caused by wind and waves and reducing the corrosion risk can bring greater benefits than simply reducing the load numerical error when performing gradient descent or path search, i.e., the loss value decreases more.

[0100] Optionally, this mechanism will produce the following specific technical effects in actual operation: When encountering unknown extreme sea conditions, the model will proactively sacrifice some numerical accuracy of load prediction, instead generating more conservative or robust prediction results. For example, the model may predict a load value slightly higher than the actual demand and suggest reducing the operating power of equipment with high corrosion risk. Although the mean square error may increase slightly from a purely numerical perspective, this prediction result can force the scheduling system to reserve more safety margin, avoiding equipment shutdowns or pipeline overpressure accidents caused by ignoring current extreme environmental shocks due to overfitting historical data.

[0101] In summary, this embodiment successfully constructs a dynamic calibration closed loop capable of coping with black swan events by introducing distribution drift detection based on Wasserstein distance and adaptive gain adjustment. This mechanism solves the model degradation problem caused by the conceptual drift of the marine environment during long-term operation of FSRU, ensuring that the hybrid heat source scheduling system not only operates efficiently in familiar normal environments but also maintains a safety baseline in unknown extreme environments, fully demonstrating the significant progress of this invention in robustness and intelligence.

[0102] Example 3:

[0103] This embodiment, based on Embodiments 1 and 2, further refines the execution logic for hybrid heat source scheduling. Specifically, this embodiment focuses on the execution stability of the mapping relationship between the time length of the scheduling lead and the proportion of the heat source output buffer as the wind and wave level increases in the actual physical system, and elaborates on the specific implementation process of the asymmetric peak-locked damping control step.

[0104] It is important to note that in the description of Example 1, the system establishes a linear or monotonic mapping mechanism based on wind and wave levels to determine the output buffer of the heat source. Theoretically, this mechanism ensures that the heat source supply increases with the severity of the environment. However, in practical marine engineering applications, the raw wind and wave data directly collected by the energy management controller is often not a smooth curve, but rather a high-frequency fluctuating signal superimposed with a large amount of random noise. For example, gusts, breaking waves, or instantaneous swells can cause sensor readings to jump drastically in a short period. If the dispatch system responds completely in real-time to these high-frequency signals—that is, immediately increasing the load when the wind and waves are high and immediately decreasing the load when the wind and waves are slightly lower—it will cause the actuators of the FSRU, such as steam regulating valves, gas turbine throttles, and seawater pump frequency converters, to fall into a state of frequent oscillation. This oscillation not only fails to improve the system's response speed, but also, due to the huge thermal inertia of the intermediate medium vaporizer (IFV) or regasification boiler, causes severe thermal stress fatigue inside the equipment, and may even lead to divergence and instability of the control system.

[0105] Therefore, after determining the buffer margin of the heat source output based on the mapping relationship between wind and wave levels and the heat source output buffer capacity, this embodiment introduces a dedicated damping control logic, namely, executing an asymmetric peak-locked damping control step. The core design concept of this step is rapid rise and slow fall, that is, the response to environmental deterioration is immediate and sensitive, while the response to environmental improvement is delayed and controlled, thereby suppressing equipment thermal fatigue caused by high-frequency environmental fluctuations.

[0106] Specifically, in order to accurately identify whether the current sea state is in a high-frequency turbulent state that requires damping control, the system first performs the operation of calculating the environmental fluctuation index.

[0107] For example, the environmental fluctuation index is a key indicator for measuring the dispersion of marine environmental parameters. This index is generated based on the variance of wind and wave level data within a sliding time window. The energy management controller maintains a fixed-length sliding time window in memory, covering wind and wave level sampling points over the past several minutes, such as five to ten minutes. The system calculates the arithmetic mean of all sampling points within this window in real time, then calculates the square of the difference between each sampling point and the arithmetic mean, and finally takes the average of these squared values ​​to obtain the variance value. To eliminate the influence of dimensions and facilitate subsequent comparisons, the system can further normalize this variance value, and the final generated value is the environmental fluctuation index.

[0108] Optionally, the environmental volatility index physically characterizes the roughness of the sea surface. When the sea surface is mainly dominated by long-period swells, the wind and wave levels change relatively smoothly, and the environmental volatility index is low; however, when the sea surface is disturbed by strong gusts or turbulence, the wind and wave level readings change drastically in a short period of time, resulting in a significant increase in the environmental volatility index.

[0109] After obtaining the environmental fluctuation index, the system performs an operation to compare the environmental fluctuation index with a preset turbulence threshold. The turbulence threshold is a preset empirical constant whose value depends on the seasonal characteristics of the sea area where the FSRU is located and the upper limit of the mechanical adjustment frequency response of the heat source equipment.

[0110] If the calculated environmental fluctuation index exceeds the turbulence threshold, the system determines that the current environment has entered a high-frequency turbulence state. At this point, if the direct mapping relationship in Example 1 is continued, the operating frequency of the regulating valve will inevitably exceed its design life limit. Therefore, the system suspends the real-time execution of the mapping relationship and activates the peak lockout mode. This suspension does not mean stopping scheduling, but rather cutting off the signal path for real-time wind and wave data to directly control the heat source output, and instead allowing a more robust lockout logic to take over control.

[0111] It should also be noted that once the peak lock mode is entered, the system's scheduling logic will undergo a fundamental change, shifting from servo control to conservative anchoring.

[0112] Specifically, in the peak locking mode, the system performs the operation of forcibly anchoring the scheduling advance and the heat source output buffer to historical peak parameters.

[0113] For example, the so-called historical peak parameter refers to the scheduling parameter corresponding to the maximum wind and wave level monitored within the sliding time window. The energy management controller traverses all historical wind and wave level data within the sliding time window and filters out the maximum value. Subsequently, the system uses the mapping relationship of Embodiment 1 to query the scheduling lead and heat source output buffer corresponding to the maximum wind and wave level. Regardless of whether the real-time wind and wave level has subsided at the current moment, the system forcibly maintains the setpoint of the actuator at the level corresponding to this maximum value.

[0114] Optionally, this design embodies a high degree of safety. In turbulent conditions, a decrease in wave readings at a given moment may simply be a temporary lull in wind or a trough, not an indication of the end of severe weather. By anchoring to historical peak parameters, the system effectively assumes that the worst-case scenario in the near future will at least reach the highest level of the past period. For example, if the wave level reached level eight in the past ten minutes, even if the current instantaneous reading drops to level six, the system still reserves 20% of the heat source buffer capacity according to the level eight standard and executes a 20-minute advance scheduling allowance. This strategy ensures that the FSRU always has sufficient thermal energy reserves to cope with potential rebounds in wind and waves, effectively avoiding the risk of supply disruptions caused by prematurely releasing energy reserves due to misjudging improved conditions.

[0115] While peak lockout mode ensures safety, it also increases energy consumption. To ensure a timely return to normal economic operation while maintaining safety, this embodiment designs a stringent exit mechanism, namely, executing delayed release logic.

[0116] It is important to note that the core of this logic is to prevent the system from repeatedly entering and exiting the lockout mode when environmental parameters fluctuate around the threshold, that is, to eliminate the ping-pong effect.

[0117] Specifically, the logic first includes activating the thermal relaxation protection timer. The thermal relaxation protection timer is a countdown or forward timing module, and its set duration is called the thermal relaxation time. This time setting is closely related to the physical thermal time constant of the heat source equipment. For example, for large gas-fired boilers or intermediate medium gasifiers, the time required for the pipe wall temperature to naturally cool from a high-load state to a low-load equilibrium state may be as long as thirty minutes or even longer. Therefore, the timer's set value is usually set to cover this physical cooling process to prevent thermal shock caused by human intervention exceeding the equipment's physical response speed.

[0118] The system is only allowed to reduce the output of the heat source when extremely stringent combined conditions are met. Specifically, the release operation is only allowed when the moving average of the real-time wind and wave levels remains below the hysteresis threshold for a duration exceeding the set value of the thermal relaxation protection timer.

[0119] This introduces the concept of a retracement hysteresis threshold, which is typically set below the wave level that triggered peak locking, or a dynamically calculated baseline. For example, if peak locking is triggered by wave level 8, the retracement hysteresis threshold might be set to level 6 or lower.

[0120] Specifically, the moving average is introduced to further filter out high-frequency noise. The system no longer looks at instantaneous values, but rather at the arithmetic average trend of wind and wave levels over a period of time, such as one minute.

[0121] The entire judgment logic is as follows: Assume the system is currently locked at level 8 wind and wave intensity. At this point, the wind and waves begin to weaken. The system detects that the moving average of the real-time wind and wave intensity has dropped below level 6. The system does not act immediately but triggers a thermal relaxation protection timer. For the next thirty minutes, the system continuously monitors the wind and wave data. If, during this process, the moving average of the wind and wave intensity rebounds back above level 6, the timer immediately resets to zero, and the system continues to maintain the level 8 lockout. Only when the wind and wave intensity is consistently and continuously below level 6 for more than thirty minutes does the system confirm that the environment has truly improved, not just experienced a temporary fluctuation, and that the thermal stress on the equipment has been fully released.

[0122] It should also be noted that even if all the above stringent release conditions are met, the method of this embodiment absolutely does not allow a precipitous drop in parameters. The system performs the operation of gradually reducing the scheduling advance and the heat source output buffer amount with a preset stepped decay step size.

[0123] Specifically, the step decay step size is a predefined parameter descent gradient. For example, the amount of heat source output buffer is reduced by one percent or two percent each time, and the time for each reduction in scheduling lead is limited to one minute.

[0124] When the system decides to exit peak lockout mode, the energy management controller generates a series of discrete control commands. Assuming the current buffer level is 20%, and the target is to reduce it to 5%, the system will not send the 5% command all at once. Instead, it will send a new setpoint at preset time intervals, such as every two minutes: 19%, 18%, 17%, and so on, until the target value is reached.

[0125] Optionally, this stepped release process simulates the natural cooling or unloading curve of physical equipment. For the core component of the regasification system, the intermediate medium vaporizer (IFV), the temperature field change of its propane or ethylene glycol medium must be gradual. If the heat source input is drastically reduced in a short period, such as by suddenly shutting down a boiler, the temperature gradient within the vaporizer will be drastically distorted, resulting in a huge difference in thermal expansion and contraction between the heat exchange tube bundle and the shell, which can lead to flange leaks or weld cracks. By limiting the stepped decay step size, the scheduling system artificially smooths the control commands, allowing the thermal stress inside the equipment to be released in a controlled manner, greatly extending the service life of critical equipment.

[0126] In summary, this embodiment details the entire process of the asymmetric peak-locked damping control mechanism. This mechanism accurately identifies high-frequency turbulent states through the environmental fluctuation index, establishes a safety barrier under severe sea conditions using the peak-locked mode, constructs a robust exit defense line relying on the thermal relaxation protection timer and the fallback hysteresis threshold, and finally achieves a flexible parametric soft landing through a stepped decay step size.

[0127] Specifically, this technical solution has achieved significant beneficial effects in actual operation: it fundamentally solves the control challenge of FSRUs requiring both rapid response and stable operation under complex sea conditions. During the rising edge of approaching waves, the system responds with almost zero delay and locks into a high-energy state, ensuring the continuity of natural gas output and preventing tripping accidents caused by pressure fluctuations. During the falling edge of receding waves, the system patiently waits for the environment to completely stabilize and slowly adjusts parameters, avoiding frequent start-ups and shutdowns and thermal fatigue damage to the equipment. This asymmetric control intelligence enables the FSRU's hybrid heat source system to maintain higher reliability and lower maintenance costs throughout its entire life cycle, providing important guidance for the intelligent operation of offshore energy equipment.

[0128] Example 4:

[0129] like Figure 2 As shown, this embodiment provides a FSRU regasification hybrid heat source scheduling system based on load forecasting and rule learning. Specifically, this system is configured to be embedded in or run within the energy management controller of a Floating Storage and Regasification Unit (FSRU). As the core hub of the entire FSRU energy scheduling, this energy management controller is physically connected to various onboard sensors, actuators, and communication equipment via a high-frequency data bus. It aims to address the load forecasting distortion caused by the non-stationarity of the marine environment, as mentioned in the background art, and the safety hazards caused by the mismatch between static scheduling strategies and dynamic environmental impacts.

[0130] Specifically, the system in this embodiment includes:

[0131] The scheduling system described in this embodiment mainly comprises four core functional modules in its logical architecture: a data acquisition and preprocessing module, a dynamic load forecasting module, an intelligent rule generation module, and a hybrid scheduling execution module. These four modules are coupled together to form a closed-loop control circuit from environmental perception to decision execution.

[0132] For example, the data acquisition and preprocessing module is the system's sensing front end. Its main function is to collect basic operational data of the FSRU, marine environmental data, and coupled correlation data, and to perform data cleaning and standardization preprocessing. In actual operation, this module reads wind and wave levels, salt spray concentrations, and pipeline pressure and flow data from weather stations in real time through hardware interfaces. Addressing the issue of high-frequency noise often accompanying marine environmental data mentioned in the background section, this module incorporates a time alignment algorithm and anomaly removal mechanism. It not only unifies data from different sampling frequencies to a standard time axis but also uses statistical criteria to remove invalid outliers caused by sensor malfunctions, ensuring that the data input to subsequent modules has a very high signal-to-noise ratio and spatiotemporal consistency. Furthermore, this module is also responsible for generating the coupling characteristics of wind and wave and pipeline pressure, providing key input dimensions for subsequent models to characterize the physical impact of the environment.

[0133] For example, the dynamic load forecasting module plays a crucial role in accurately predicting energy demand. This module is configured to input preprocessed data into a three-dimensional coupled load forecasting model and dynamically switch the network structure complexity of the model based on real-time collected marine environmental parameters, thereby outputting load forecast values ​​and deviation probabilities.

[0134] It is important to note that this module's design directly addresses the challenges posed by dynamic environmental impacts in the background technology. To strike a balance between computational efficiency and prediction robustness, this module is configured to execute a rigorous set of discriminative logic when performing dynamic model switching:

[0135] Specifically, the module first responds to real-time collected wind and wave level data and salt spray concentration data to determine the current marine environment category. If the monitored wind and wave level is less than or equal to a first intensity threshold and the salt spray concentration is less than a first concentration threshold, the system determines that the current sea state is stable, and the module automatically calls a simplified structural model for inference. This simplified model has low computational load and can quickly respond to normal load fluctuations, effectively reducing the computing power overhead of the energy management controller.

[0136] Conversely, if the monitoring data indicates that the wave intensity is greater than the first intensity threshold, or the salt spray concentration is greater than or equal to the first concentration threshold, or the ocean temperature change rate exceeds the preset temperature change threshold, the module will immediately switch to the complex structure model for reasoning.

[0137] Specifically, to address nonlinear disturbances under extreme operating conditions, the complex structural model integrated in this embodiment employs a weighted time-decayed residual structure long short-term memory network combined with a graph attention network architecture. The weighted time decay mechanism makes the model more attentive to recent meteorological changes, the residual structure solves the degradation problem of deep networks, and the graph attention network captures the spatial dependencies between nodes in the pipeline network topology. This deep composite architecture ensures that the system can still output high-confidence load forecasts and deviation probabilities even in extremely harsh environments, thereby achieving the beneficial effect of improving prediction accuracy under the influence of nonlinear environments.

[0138] After obtaining the prediction results, the intelligent rule generation module begins operation. This module inputs an environmental state vector containing current marine environmental parameters and equipment status into a deep Q-network adapted to the marine environment to generate heat source scheduling rules. Internally, this module maintains a multi-dimensional environmental state vector space, covering key indicators such as equipment health and wave levels. Through deep reinforcement learning algorithms, this module can learn optimal strategies from tens of thousands of simulations, generating scheduling rules that balance cost and safety, such as proactively limiting the load on corrosive equipment when salt spray concentrations are high.

[0139] Ultimately, all decision-making information converges to the hybrid scheduling execution module. This module is used to perform pre-scheduled and real-time scheduling decisions based on the load forecast, deviation probability, and heat source scheduling rules, combined with the reliability of remaining renewable energy power.

[0140] Specifically, this module serves as the execution terminal for the system's interaction with physical equipment. It not only adjusts the heat source output based on predicted values ​​but also executes the asymmetric peak-locked damping control logic detailed in Example 3. When the dynamic load forecasting module warns of environmental deterioration, the hybrid scheduling execution module intelligently determines whether to prioritize utilizing the shipborne photovoltaic-driven pump unit or activate grid-based backup power based on the reliability of the remaining renewable energy capacity. Simultaneously, by implementing rapid-rise and slow-fall damping control, this module effectively resolves the contradiction between the need for rapid and safe response and the protection against equipment thermal fatigue in the background technology, ensuring that the FSRU maximizes the service life of critical heat source equipment while guaranteeing a continuous natural gas supply.

[0141] In summary, the FSRU regasification hybrid heat source scheduling system provided in this embodiment, through the close collaboration and intelligent design of various functional modules, successfully transforms an abstract scheduling method into specific hardware control logic, providing strong technical support for the efficient and safe operation of FSRUs in complex marine environments.

[0142] Example 5:

[0143] Corresponding to the above embodiments, the present invention also proposes an electronic device.

[0144] like Figure 3 The diagram shows a structural schematic of an electronic device according to the present invention. The electronic device 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the electronic device 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one unit, and the structure of this electronic device 100 does not constitute a limitation on the embodiments of the present invention.

[0145] Processor 101 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this disclosure. Processor 101 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0146] Bus 102 may include a pathway for transmitting information between the aforementioned components. Bus 102 may be a PCI bus or an EISA bus, etc. Bus 102 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0147] The memory 103 stores a computer program corresponding to the FSRU regasification mixed heat source scheduling method based on load forecasting and rule learning according to the above embodiments of the present invention. The computer program is controlled and executed by the processor 101. The processor 101 executes the computer program stored in the memory 103 to implement the content shown in the aforementioned method embodiments.

[0148] Among them, electronic devices 100 include, but are not limited to: mobile terminals such as laptops and PADs (tablet computers) and fixed terminals such as desktop computers. Figure 3 The electronic device 100 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0149] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for scheduling FSRU regasification hybrid heat sources based on load forecasting and rule learning, characterized in that, include: Collect basic operational data, marine environmental data, and coupled correlation data of the FSRU. The coupled correlation data includes mapping data between wind and waves and pipeline pressure, and mapping data between salt spray and equipment failure rate. Perform data cleaning and standardization preprocessing. The preprocessed data is input into the three-dimensional coupled load prediction model. Based on the real-time collected marine environmental parameters, the network structure complexity of the model is dynamically switched, and the load prediction value and deviation probability are output. An environmental state vector containing current marine environmental parameters and equipment status is input into a deep Q-network adapted to the marine environment. The network is trained based on a reward function that is negatively correlated with load deviation rate, energy cost, equipment corrosion aggravation value and pipeline pressure fluctuation value, and positively correlated with supply continuity coefficient. The heat source scheduling rules are generated, including pipeline pressure pre-regulation, wind and wave resistant heat source combination switching, high corrosion equipment load limitation and new energy and grid coordinated energy supplementation actions. Based on the load forecast, deviation probability, and heat source scheduling rules, and combined with the reliability of the remaining power of new energy sources, advance scheduling and real-time scheduling decisions are executed. In this process, historical remaining power data of new energy sources and marine environmental parameters are input into the gated cyclic unit model, and the probability value between 0 and 1 is calculated as the credibility of the remaining power of the new energy sources. The pre-scheduling is as follows: acquiring marine meteorological forecast data for a future preset time period, determining the execution time of the pre-scheduling operation based on the mapping relationship between wind and wave level and scheduling lead time, and determining the buffer margin of heat source output based on the mapping relationship between wind and wave level and heat source output buffer capacity, so as to perform system preheating and increase pipeline pressure in advance. The real-time scheduling is as follows: when the confidence level of the remaining power of the new energy source is greater than or equal to the confidence threshold, a scheduling instruction is generated to prioritize the use of new energy sources to drive the transmission of deep heat sources; when the confidence level is less than the confidence threshold, a scheduling instruction is generated to activate the grid energy replenishment backup. The dynamic switching process of the three-dimensional coupled load prediction model includes: In response to real-time collected wind and wave level data and salt spray concentration data, the current marine environment category is determined. If the wind and wave level is less than or equal to a first intensity threshold and the salt spray concentration is less than a first concentration threshold, a simplified structural model is invoked for inference. If the wind and wave level is greater than the first intensity threshold, or the salt spray concentration is greater than or equal to the first concentration threshold, or the ocean temperature change rate exceeds a preset temperature change threshold, a complex structural model is invoked for inference. The complex structural model adopts a weighted time decay residual structure long short-term memory network combined with a graph attention network architecture.

2. The method according to claim 1, characterized in that, The process involves collecting basic operational data of the FSRU, marine environmental data, and coupled correlation data, and performing data cleaning and standardization preprocessing, including: The data is collected using a lightweight messaging protocol and transmitted using a dual-link backup channel. Time alignment is performed on the collected data to unify data from different sampling periods to a standard time interval; The coupling characteristics between wind and waves and pipeline pressure are generated. These coupling characteristics are calculated based on wind and wave levels and pipeline strain values, and are used to characterize the degree of influence of the marine environment on pipeline pressure. Based on statistical anomaly detection criteria and marine environmental trend judgment mechanism, invalid outliers are removed and valid mutation data caused by wind, waves or salt spray are retained.

3. The method according to claim 1, characterized in that, The training process of the three-dimensional coupled load prediction model includes: Construct an input feature set that includes historical load data, marine environmental characteristics, equipment health characteristics, and pipeline topology characteristics; The model is trained using a loss function, which consists of the mean square error of the predicted values, the prediction bias caused by wind and waves, and the prediction weights corresponding to the equipment corrosion risk. The weighting coefficients of the prediction bias term and the prediction weight term are determined by fitting historical marine environmental data.

4. The method according to claim 3, characterized in that, After determining the weighting coefficients of the prediction bias term and the prediction weight term by fitting historical marine environmental data, the method further includes performing an online dynamic calibration step, specifically: Establish a real-time observation window containing data from the most recent preset time period, and a benchmark reference window containing historical typical working condition data; The distribution deviation index of the rate of change of wind and wave levels between the real-time observation window and the benchmark reference window is calculated using the bulldozer distance algorithm; An adaptive gain function is constructed based on the distribution deviation index, and a real-time gain factor is calculated. The adaptive gain function is configured to output a value greater than one when the distribution deviation index exceeds the safe drift threshold. The weighting coefficients of the prediction deviation term and the prediction weight term are corrected by real-time multiplication using the real-time gain factor to generate dynamic execution coefficients; The dynamic execution coefficients are fed back into the loss function, and when the FSRU encounters an unstable marine environment that does not appear in historical data, the weight of the prediction bias caused by wind and waves and the risk of equipment corrosion in the model optimization objective is dynamically increased.

5. The method according to claim 1, characterized in that, The environmental state vector input to the deep Q-network includes at least wind and wave levels, salt spray concentration, equipment health, load forecast, and remaining renewable energy power.

6. The method according to claim 1, characterized in that, The mapping relationship between wind and wave level and scheduling advance, and the mapping relationship between wind and wave level and heat source output buffer are specifically set as follows: as the wind and wave level increases, the duration of the scheduling advance and the proportion of the heat source output buffer are increased.

7. The method according to claim 6, characterized in that, After increasing the time length of the scheduling lead and the proportion of the heat source output buffer as the wind and wave level increases, the process further includes an asymmetric peak-locked damping control step, specifically: Calculate the environmental fluctuation index, which is generated based on the variance of wind and wave level data within a sliding time window; The environmental fluctuation index is compared with a preset turbulence threshold. If the turbulence threshold is exceeded, the real-time execution of the mapping relationship is suspended and the peak lock mode is activated. In the peak locking mode, the scheduling advance and the heat source output buffer are forcibly anchored to historical peak parameters, which correspond to the maximum wind and wave level monitored within the sliding time window. The execution of the hysteresis release logic includes activating the thermal relaxation protection timer. Only when the moving average of the real-time wind and wave level is continuously lower than the hysteresis threshold and the duration exceeds the set value of the thermal relaxation protection timer, is it allowed to gradually reduce the scheduling advance and the heat source output buffer amount with a preset stepped decay step size.

8. The method according to claim 1, characterized in that, It also includes distributed parameter aggregation and model optimization through a federated learning module adapted for oceanization, including: The polymerization cycle is dynamically adjusted according to the current wind and wave level and salt spray concentration. The first polymerization cycle is set in a stable environment, and the second polymerization cycle, which is shorter than the first polymerization cycle, is set in a harsh environment. For the node embedding parameters of the graph neural network in the complex structural model, the weight distribution of highly associated edges with transmission efficiency higher than a preset ratio is retained. The aggregated model parameters are validated using a consistency verification mechanism based on a robust loss function.

9. The method according to claim 1, characterized in that, It also includes lightweight classification and graded response for anomalies specific to the marine environment, including: Anomaly types are identified using a decision tree model, including environmental mutations, equipment corrosion, and coupled faults. If a high-priority anomaly is identified, an emergency heat source combination will be triggered and the opening of pipeline valves will be adjusted within the first preset time. If a medium-priority anomaly is identified, the load on high-risk equipment will be reduced and redundant equipment will be activated within a second preset time period that is longer than the first preset time. Record the mapping relationship between environmental parameters, exception types, scheduling actions and processing effects, and update the exception response strategy library within a preset period.

Images