A multi-source energy collaborative management method applied to an unmanned sail vehicle

Abstract

This invention discloses a multi-source energy collaborative management method for unmanned sailing vehicles. The method includes receiving a set of future forecast parameters from a cloud-based weather forecasting system; preprocessing the data in the forecast parameter set to obtain standardized forecast time-series data; acquiring real-time status data and historical operating condition data of the unmanned sailing vehicle and inputting them into a pre-trained energy consumption and endurance prediction model to obtain multivariate time-series prediction results in the prediction time domain; dynamically constructing a set of time-domain inequality constraints; and inputting the standardized forecast time-series data, time-domain inequality constraints, real-time status data, and multivariate time-series prediction results into an MPC controller to obtain a future time-domain continuous optimal control sequence. This invention can effectively avoid the risk of energy storage power depletion, fully utilize surplus clean energy, support energy planning for long-range voyages, and improve the accuracy of voyage time forecasts.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, specifically to a multi-source energy collaborative management method applied to unmanned sailing vehicles. Background Technology

[0002] At present, with the continuous enrichment of marine in-situ environmental monitoring methods and the continuous growth of the marine unmanned vehicle industry, unmanned sailing vehicles rely on natural wind energy as the main propulsion power, using green and clean energy, which greatly improves the endurance of the vehicle in the ocean and increases the working distance and working events of the vehicle.

[0003] However, the following problems exist in the energy management and control of sailing vehicles: 1. Traditional energy management uses static control methods with fixed thresholds and fixed power ratios, which cannot dynamically adjust the power supply strategy according to the real-time changes in load energy consumption. This easily leads to insufficient power supply under high load conditions and wasted energy during low load standby, and cannot achieve autonomous and precise energy allocation in unmanned scenarios. 2. Existing energy consumption modeling relies solely on real-time sensor data, lacks historical time-series data correction, and lacks prior support from future meteorological and hydrological data. It cannot be combined with time-series operating data accumulated over long-term flight of the aircraft, making it difficult to achieve real-time correction of the power consumption model and resulting in low accuracy of energy management. 3. Traditional underwater vehicles can only perceive the marine environment based on the sensors on board, lacking the ability to predict the environment in advance and coordinate energy and navigation. They cannot grasp the temporal changes in wind speed, wind direction, current, and solar radiation in the operating area in advance, and cannot adjust navigation and power supply strategies in advance to match wind and solar energy resources, resulting in low utilization of clean energy and the need for manual adjustment of navigation operation modes. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a multi-source energy collaborative management method applicable to unmanned sailing vehicles.

[0005] To achieve the above objectives, the present invention provides a multi-source energy collaborative management method for an unmanned sailing vehicle, wherein the unmanned sailing vehicle is equipped with a photovoltaic array, an adjustable wing sail mechanism, an electric auxiliary propulsion system, and an energy storage battery, and the method includes: Receive the set of future forecast parameters sent by the cloud-based weather forecasting system; The data in the forecast parameter set are preprocessed to obtain standardized forecast time series data; Acquire real-time status data and historical operating condition data of unmanned sailing vehicles; The standardized forecast time series data, real-time status data, and historical operating condition data are input into the pre-trained energy consumption and range prediction model to obtain multivariate time series prediction results in the prediction time domain. A set of time-domain inequality constraints is dynamically constructed, and the standardized forecast time series data, time-domain inequality constraints, real-time state data, and multivariate time series prediction results are input to the MPC controller. The MPC controller uses the objective function of stable state of charge of energy storage battery, minimum total power consumption of system, suppression of frequent fluctuations in propulsion power, and ensuring stable power supply to load as the objective function to solve for the future continuous optimal control sequence in the time domain.

[0006] Furthermore, the set of forecast parameters is as follows: ; in, for The set of forecast parameters sent at each time point, To forecast wind speed timing, To predict wind direction and timing, To forecast the timing of solar irradiance over the sea surface, To predict the timing of tidal currents; The specific real-time status data is as follows: ; in, for Real-time status data at any given moment. For real-time wind speed, Solar irradiance, The remaining power of the energy storage battery, The actual angle of attack of the adjustable wing sail mechanism. For real-time load power, This represents the real-time photovoltaic power generation of the photovoltaic array. This is to enable the adjustable wing sail mechanism to capture the equivalent mechanical power of wind energy in real time.

[0007] Furthermore, the energy consumption range prediction model includes: The input layer is used to set the length of the time-series sliding window and to concatenate the standardized forecast time-series data, real-time status data and historical operating condition data to obtain a multi-dimensional input matrix. The feature fusion layer is used to perform nonlinear mapping and fusion on the multidimensional input matrix output by the input layer to obtain fused features; The LSTM temporal feature extraction layer is used to predict the dynamic evolution trend of future wind and solar power based on the fused features and output the hidden features at the current moment. The output layer is used to output the multivariate time series prediction results in the prediction time domain based on the hidden features at the current time.

[0008] Furthermore, the multidimensional input matrix is ​​represented as follows: ; in, The multidimensional input matrix obtained by splicing, For splicing operations, For standardized forecast time series data, Historical operating data; The feature fusion layer performs nonlinear mapping fusion in the following way: ; in, As a feature of fusion, For activation function, This is the fusion weight matrix of the feature fusion layer. This is the bias vector.

[0009] Furthermore, the hidden features at the current time point output by the LSTM temporal feature extraction layer are: ; ; ; ; ; ; in, for Hidden features of time, for The output vector of the output gate at each time step. For Hadama accumulation, The hyperbolic tangent activation function is used. for Cellular state at all times It is the Sigmoid activation function. This is the weight matrix of the output gate. for The fusion characteristics of moments for Hidden features of time, This is a vector concatenation symbol. This is the bias vector for the output gate. This is the output vector of the forget gate. for Cellular state at all times The output vector of the input gate. for The state of candidate cells at any given time. Here is the weight matrix for the candidate states. This is the bias vector for the candidate state. Here is the weight matrix of the input gate. This is the bias vector for the input gate. Here is the weight matrix for the forget gate. This is the bias vector for the forget gate.

[0010] Furthermore, the multivariate time series prediction result is expressed as follows: ; in, For multivariate time series prediction results, for The predicted power generation of the photovoltaic array at any given time. for The time-adjustable wing sail mechanism captures wind energy equivalent to predicted mechanical power. for Predicted power consumption for various loads at all times. for The remaining self-continuous flight time at that moment, , This is the prediction time domain for the energy consumption and range prediction model.

[0011] Furthermore, the loss function used during training of the energy consumption and range prediction model... for: ; ; ; ; in, The overall prediction mean square error, This is the absolute error term for photovoltaic power prediction. This is the absolute error term for wind power prediction. For adjustable weight hyperparameters, for Observations of a multivariate time series at time 10:00. for Photovoltaic power generation at any given time for Real-time wind energy capture measured power. It is the absolute value symbol.

[0012] Furthermore, the time-domain inequality constraints are as follows: ; in, for The power generation of the photovoltaic array at any time for The maximum power output of the photovoltaic array at any given time. for Forecast of sea surface solar irradiance at specific times. for The adjustable wing sail mechanism captures the equivalent mechanical power of wind energy in real time. for The adjustable wing sail mechanism captures the maximum value of the equivalent mechanical power corresponding to wind energy in real time. for Forecast wind speed and direction at specific times. for The output power of the electric auxiliary thruster at any given time, , These are the minimum and maximum output power of the electric auxiliary thruster, respectively. for The remaining charge of the energy storage battery at any given time. , These are the minimum and maximum remaining capacity of the energy storage battery, respectively. for Load power at any given time , These are the minimum and maximum values ​​of the load power, respectively. , This represents the prediction time domain for the MPC controller.

[0013] Furthermore, the objective function is expressed as: ; in, For control sequence Find the objective function The minimum value, for Total system power consumption at any given time. The target value for the remaining capacity of the energy storage battery. for The remaining power of the energy storage battery at any given time. for The change in the electric auxiliary propulsion load power at any given time. for The change in load power at any given time. , , , All are weighting coefficients.

[0014] Furthermore, the MPC controller also adjusts the forecast of sea surface solar irradiance time series. And solar irradiance determination threshold The pattern determination is performed as follows: When using an electric-assisted propulsion system in conjunction with an adjustable wing sail mechanism, if the predicted time period threshold exceeds [a certain threshold] within the prediction window... and If the conditions are met, the electric auxiliary thruster will continue to operate, and excess electricity generated by the photovoltaic array will be used to charge the energy storage battery; otherwise, the electric auxiliary thruster will not operate.

[0015] Beneficial effects: 1. Based on meteorological and hydrological forecast data obtained from the cloud every day, the system introduces future environmental prior information to achieve proactive energy regulation in advance, reduce the shortcomings of traditional feedback control, and thus plan energy allocation, load management and navigation strategies in advance, effectively avoid the risk of energy storage power loss, and fully utilize surplus clean energy. 2. A three-type time-series fusion input architecture is adopted, and the weather forecast is used as an independent feature input to a lightweight DNN-LSTM network; a composite loss function with wind and solar power constraints is designed accordingly; compared with the prediction scheme that relies solely on historical data, the prediction error of future wind and solar power output is significantly reduced when facing sudden changes in marine weather, solving the problem of long-term time-series forecast drift and supporting energy planning for long-range voyages. 3. By dynamically switching propulsion modes according to the predicted solar irradiance, such as starting the thrusters in high light environment to increase the spacecraft's range while ensuring that solar energy powers its own batteries; and shutting down the thrusters in low light environment to save energy storage; thereby achieving integrated optimization of motion control and energy management; and achieving deep synergy between navigation control strategy and energy scheduling, thereby extending the navigation time and improving the accuracy of navigation time prediction. 4. Forecast time series data simultaneously serve the energy consumption and endurance prediction model and the MPC controller; compared with the schemes used in a single link of forecast information, the optimized feasible domain is more in line with the real future sea conditions, and the energy allocation scheme has the characteristics of global optimization on a long time scale. 5. By continuously collecting local operating condition data and periodically fine-tuning network parameters, and by utilizing historical data, real-time data, and weather forecast data in combination, the energy consumption and endurance prediction model can continuously adapt to local weather and sea condition characteristics during long-term cruises in different sea areas, thus continuously improving long-term operational stability. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the multi-source energy collaborative management method applied to unmanned sailing vehicles according to an embodiment of the present invention. Detailed Implementation

[0017] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. These embodiments are implemented based on the technical solutions of the present invention, and it should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0018] like Figure 1 As shown, this embodiment of the invention provides a multi-source energy collaborative management method for an unmanned sailing vehicle. The unmanned sailing vehicle is equipped with a photovoltaic array, an adjustable wing sail mechanism, an electric auxiliary propulsion system, and an energy storage battery. The method includes: The set of future forecast parameters received from the cloud-based weather forecasting system can be specifically represented as: ; in, for The set of forecast parameters sent at each time point, To forecast wind speed timing, To predict wind direction and timing, To forecast the timing of solar irradiance over the sea surface, To predict the timing of tidal currents (velocity and direction).

[0019] The data in the forecast parameter set are preprocessed to obtain standardized forecast time-series data. The preprocessing methods include data parsing and time alignment. The resulting standardized forecast time-series data is represented as follows: .

[0020] Acquire real-time status data and historical operating condition data of the unmanned sailing vehicle. The specific real-time status data is as follows: ; in, for Real-time status data at any given moment. For real-time wind speed, Solar irradiance, The remaining power of the energy storage battery, The actual angle of attack of the adjustable wing sail mechanism. For real-time load power, This represents the real-time photovoltaic power generation of the photovoltaic array. To estimate the equivalent mechanical power of wind energy captured in real time by the adjustable wing sail mechanism, it can be done in the following way: ; ; in, Standard air density, , The effective windward area of ​​the wing sail in the adjustable wing sail mechanism. The relative wind speed is measured by sensors at the ship's onboard weather station. The effective thrust coefficient in the heading direction. The real-time angle of attack of the wing sail is collected in real time by the wing sail servo motor angle encoder. Relative wind direction angle, output from weather station sensors. Effective thrust coefficient for heading. It is composed of the lift coefficient and the drag coefficient, as detailed below: ; in, The lift coefficient of the wing sail. This is the drag coefficient of the wing sail, which depends on the sail itself. Additionally, the historical operating data mentioned above includes historical actual photovoltaic power generation, wind power utilization, and changes in remaining battery capacity.

[0021] Standardized forecast time-series data, real-time status data, and historical operating condition data are input into a pre-trained energy consumption and endurance prediction model to obtain multivariate time-series prediction results in the prediction time domain. The aforementioned energy consumption and endurance prediction model is a lightweight DNN-LSTM, specifically comprising an input layer, a feature fusion layer, an LSTM time-series feature extraction layer, and an output layer. The input layer is used to set the time-series sliding window length and concatenates the standardized forecast time-series data, real-time status data, and historical operating condition data to obtain a multidimensional input matrix, as detailed below: ; in, The multidimensional input matrix obtained by splicing, For splicing operations, For standardized forecast time series data, This is historical operating data.

[0022] The feature fusion layer is primarily used to perform nonlinear mapping and fusion of three types of time-series features from different sources and with different dimensions to obtain fused features. This establishes the relationships between irradiance and power generation, wind speed and direction and wind energy capture efficiency, and tidal current information and navigation force characteristics. Furthermore, by unifying the feature scale, it improves the convergence speed of subsequent time-series networks. The nonlinear mapping and fusion method used by the feature fusion layer is as follows: ; in, As a feature of fusion, For activation function, This is the fusion weight matrix of the feature fusion layer. This is the bias vector.

[0023] The LSTM temporal feature extraction layer alleviates the gradient vanishing problem during ultra-long temporal training and is primarily used to capture the long-term dependencies between marine environment and temporal energy data. Furthermore, the fused features obtained through the feature fusion layer... To further predict the future dynamic evolution trend of wind and solar power, the hidden features at the current moment are output. Specifically, the hidden features at the current moment output by the LSTM temporal feature extraction layer are: ; ; ; ; ; ; in, for Hidden features of time, for The output vector of the output gate at each time step. For Hadama accumulation, The hyperbolic tangent activation function is used. for Cellular state at all times It is the Sigmoid activation function. This is the weight matrix of the output gate. for The fusion characteristics of moments for Hidden features of time, This is a vector concatenation symbol. This is the bias vector for the output gate. This is the output vector of the forget gate. for Cellular state at all times The output vector of the input gate. for The state of candidate cells at any given time. Here is the weight matrix for the candidate states. This is the bias vector for the candidate state. Here is the weight matrix of the input gate. This is the bias vector for the input gate. Here is the weight matrix for the forget gate. This is the bias vector for the forget gate.

[0024] The output layer is used to output the multivariate time series prediction results in the prediction time domain based on the hidden features at the current time, as follows: ; in, For multivariate time series prediction results, for The predicted power generation of the photovoltaic array at any given time. for The time-adjustable wing sail mechanism captures wind energy equivalent to predicted mechanical power. for Predicted power consumption for various loads at all times. for The remaining self-continuous flight time at that moment, , This is the prediction time domain for the energy consumption and range prediction model.

[0025] In the loss function section, considering that traditional single MSE loss is insufficient to constrain the prediction error of energy components, this invention designs a composite loss function to focus on constraining the prediction accuracy of wind and solar power generation and suppressing long-term time-series drift. The loss function is: ; ; ; ; in, The overall prediction mean square error, This is the absolute error term for photovoltaic power prediction. This is the absolute error term for wind power prediction. For adjustable weight hyperparameters, for Observations of a multivariate time series at time 10:00. for Photovoltaic power generation at any given time for Real-time wind energy capture measured power. It is the absolute value symbol.

[0026] A set of time-domain inequality constraints is dynamically constructed, and the standardized forecast time series data, time-domain inequality constraints, real-time state data, and multivariate time series prediction results are input to the MPC controller. The MPC controller uses the objective function of stabilizing the state of charge of the energy storage battery, minimizing the total power consumption of the system (energy consumption of all sensors and control units), suppressing frequent fluctuations in propulsion power, and ensuring stable power supply to the load. It solves for the future continuously optimal control sequence in the time domain, specifically including the future continuously optimal charging and discharging power of the energy storage battery, the motor control frequency of the sail control, the thruster duty cycle, the servo control frequency, and the operation control of the functional sensor modules. The above objective function can be expressed as: ; in, For control sequence Find the objective function The minimum value, for Total system power consumption at any given time. The target value for the remaining capacity of the energy storage battery. for The remaining power of the energy storage battery at any given time. for The change in the electric auxiliary propulsion load power at any given time. It is the absolute value symbol. for The change in load power at any given time. , , , All are weighting coefficients.

[0027] The above time-domain inequality constraint can be expressed as: ; in, for The power generation of the photovoltaic array at any time for The maximum power output of the photovoltaic array at any given time. for Forecast of sea surface solar irradiance at specific times. for The adjustable wing sail mechanism captures the equivalent mechanical power of wind energy in real time. for The adjustable wing sail mechanism captures the maximum value of the equivalent mechanical power corresponding to wind energy in real time. for Forecast wind speed and direction at specific times. for The output power of the electric auxiliary thruster at any given time, , These are the minimum and maximum output power of the electric auxiliary thruster, respectively. for The remaining power of the energy storage battery at any given time. , These are the minimum and maximum remaining capacity of the energy storage battery, respectively. for Load power at any given time , These are the minimum and maximum values ​​of the load power, respectively. , This is the prediction time domain for the MPC controller. The upper limit of photovoltaic power generation is dynamically updated based on the predicted irradiance time series; the upper limit of wind energy capture is corrected in real time based on the predicted wind speed and direction; at the same time, tidal current forecasts are introduced to correct the basic power consumption boundary corresponding to navigation resistance.

[0028] In addition, the MPC controller also adjusts the forecast of sea surface solar irradiance time series. And solar irradiance determination threshold The pattern determination is performed as follows: When using an electric-assisted propulsion system in conjunction with an adjustable wing sail mechanism, if the predicted time period threshold (e.g., half of the predicted window) exceeds... and Then, continue to control the electric auxiliary thruster to operate, and control the excess power generated by the photovoltaic array to charge the energy storage battery; When external energy sources are scarce, i.e., for most of the forecast window Then set: The system is configured with zero power supply from the electric auxiliary propulsion system, relying solely on the adjustable wing sail mechanism for navigation. This maximizes the conservation of stored electrical energy, while the acquired solar energy is prioritized for use by observation sensors and satellite communication equipment. Furthermore, the MPC controller continuously optimizes the sail angle-of-attack sequence based on forecasted wind direction to maximize wind energy capture efficiency.

[0029] Ultimately, the vehicle's energy control system continuously collects real-time operational data and writes it to a local time-series database. The main control unit periodically extracts the latest time-series samples, combines them with the real-time collected data, and uses a lightweight DNN-LSTM network to predict errors. Furthermore, a gradient descent algorithm is employed to fine-tune the network weights online. ; in, The network weights after fine-tuning. The network weights before fine-tuning. For learning rate, This represents the gradient of the loss function with respect to the weights. Through continuous online fine-tuning, the energy consumption and endurance prediction model can adapt to different sea conditions and lighting conditions, achieving continuous optimization of prediction performance under long-term operation.

[0030] The training method for the energy consumption and range prediction model in this embodiment of the invention is as follows: The training dataset consists of three parts: computer simulation data, laboratory tank test data, and test data during sea navigation. The computer simulation data is obtained by testing the vehicle in a simulated navigation environment built with MATLAB Simulink. The laboratory tank test data consists of power consumption and power parameters of various components tested in a real water environment within a laboratory tank. The sea navigation test data consists of actual parameters such as power consumption, speed, and heading obtained by sensors onboard the vehicle during its sea navigation.

[0031] The training set during the training process consisted of: pool experimental data, simulated virtual data, and 30% of the in-situ ocean test data. The test set consisted of the remaining 70% of the in-situ ocean test data.

[0032] The feature extraction process involves directly concatenating the cloud-based forecast time series, real-time sensing time series, and historical operating condition time series into an input matrix after simple preprocessing. It is fed into the network for training.

[0033] During training, the training batch size was set to 64, the training epochs to 300, and the loss function was set to... During training, if the loss function changes with the number of training iterations, and the loss on the training set decreases slightly while the loss on the test set remains the same or even increases, the model training is considered to have stopped early, at which point the model is considered to have converged.

[0034] Specific examples: The aircraft in this case is equipped with a flexible photovoltaic array, an adjustable wing sail mechanism, an electric auxiliary propulsion system, and a large-capacity lithium iron phosphate battery energy storage system. The airborne hardware adopts an embedded main control unit and is equipped with a Beidou satellite communication module. It is equipped with multiple environmental sensors, a high-precision power acquisition module, and local solid-state storage media for persistent storage of time-series data.

[0035] At the data processing level, the server first transmits a 72-hour forecast sequence of the patrol area to the vehicle daily via BeiDou data. Parameters include wind speed, wind direction, sea surface solar irradiance, and tidal current speed and direction. Upon receiving the data, the vehicle's control unit parses the forecast message, aligns the timestamps, and obtains standardized forecast time-series data. Simultaneously, the main control unit collects real-time sensor information: real-time wind speed, light intensity, remaining battery power, sail angle of attack, thruster power consumption, and power consumption of various marine observation sensors; it also reads the local solid-state storage time-series database and retrieves historical time-series samples. .

[0036] In addition, in the energy forecasting network part, a 24-hour time-series sliding window is constructed, and the cloud forecast time series, real-time sensing time series, and historical operating condition time series are concatenated into an input matrix. The data is fed into the network. First, it enters a DNN feature fusion layer to complete nonlinear feature mapping; the fused features are then fed into two cascaded LSTM units to mine temporal correlations; the network extrapolates the photovoltaic power generation, captureable wind power, and total system energy consumption within the future prediction window, and calculates the estimated self-sustaining flight time of the aircraft. The composite loss function proposed in this invention is used during the model training phase. Strengthen the constraints on wind and solar power output forecasting and suppress long-term forecast drift.

[0037] At the spacecraft energy control level, the cloud-based forecast time series is directly used as the measurable external disturbance sequence for the MPC controller, and a set of time-domain inequality constraints is dynamically constructed. The upper limit of photovoltaic power generation is updated in real time with the forecast irradiance, and the upper limit of wind energy capture is dynamically adjusted with the forecast wind speed and direction. The MPC uses stable battery remaining power, optimal energy consumption, and stable load as objective functions, and combines the energy prediction trajectory output by DNN-LSTM to solve for the future time-domain continuous optimal control sequence.

[0038] Simultaneously, MPC reads the predicted solar irradiance time series and compares it with the preset threshold. The comparison process determines the operation mode as follows: When irradiance is above the threshold for most of the forecast window, the combined wind and solar propulsion mode is activated, allowing auxiliary electric propulsion to operate and utilizing surplus photovoltaic power to enhance navigation maneuverability, with excess power charging the lithium battery. When irradiance is below the threshold for most of the forecast window, the pure wind-powered self-sustaining navigation mode is activated, with a hard constraint of zero electric propulsion output power, cutting off auxiliary propulsion, and prioritizing the operation of the wind turbine, core observation sensors, and satellite communication modules with stored energy. Simultaneously, the controller continuously optimizes the wing sail angle of attack commands based on forecast wind direction to maximize wind energy capture efficiency.

[0039] In addition, the aircraft continuously records data on changes in real-time photovoltaic power generation, wind power utilization, and remaining battery charge during each control cycle, and writes this data, along with real-time environmental observation data, into a local time-series database. The main control unit extracts the latest time-series samples through periodic network fine-tuning tasks, calculates network prediction errors, and updates the DNN-LSTM network weights using a gradient descent algorithm, continuously optimizing the model's adaptability to the sea conditions of the operating area. The entire control process is executed periodically.

[0040] The above description is merely a preferred embodiment of the present invention. It should be noted that for those skilled in the art, other parts not specifically described are existing technology or common knowledge. Several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-source energy collaborative management method applied to an unmanned sailing vehicle, wherein the unmanned sailing vehicle is equipped with a photovoltaic array, an adjustable wing sail mechanism, an electric auxiliary propulsion system, and an energy storage battery, characterized in that, The method includes: Receive the set of future forecast parameters sent by the cloud-based weather forecasting system; The data in the forecast parameter set are preprocessed to obtain standardized forecast time series data; Acquire real-time status data and historical operating condition data of unmanned sailing vehicles; The standardized forecast time series data, real-time status data, and historical operating condition data are input into the pre-trained energy consumption and range prediction model to obtain multivariate time series prediction results in the prediction time domain. A set of time-domain inequality constraints is dynamically constructed, and the standardized forecast time series data, time-domain inequality constraints, real-time state data, and multivariate time series prediction results are input to the MPC controller. The MPC controller uses the objective function of stable state of charge of energy storage battery, minimum total power consumption of system, suppression of frequent fluctuations in propulsion power, and ensuring stable power supply to load as the objective function to solve for the future continuous optimal control sequence in the time domain.

2. The multi-source energy collaborative management method applied to unmanned sailing vehicles according to claim 1, characterized in that, The set of forecast parameters is as follows: ； in, for The set of forecast parameters sent at each time point, To forecast wind speed timing, To predict wind direction and timing, To forecast the timing of solar irradiance over the sea surface, To predict the timing of tidal currents; The specific real-time status data is as follows: ； in, for Real-time status data at any given moment. For real-time wind speed, Solar irradiance, The remaining power of the energy storage battery, The actual angle of attack of the adjustable wing sail mechanism. For real-time load power, This represents the real-time photovoltaic power generation of the photovoltaic array. This is to enable the adjustable wing sail mechanism to capture the equivalent mechanical power of wind energy in real time.

3. The multi-source energy collaborative management method applied to unmanned sailing vehicles according to claim 2, characterized in that, The energy consumption and range prediction model includes: The input layer is used to set the length of the time-series sliding window and to concatenate the standardized forecast time-series data, real-time status data and historical operating condition data to obtain a multi-dimensional input matrix. The feature fusion layer is used to perform nonlinear mapping and fusion on the multidimensional input matrix output by the input layer to obtain fused features; The LSTM temporal feature extraction layer is used to predict the dynamic evolution trend of future wind and solar power based on the fused features and output the hidden features at the current moment. The output layer is used to output the multivariate time series prediction results in the prediction time domain based on the hidden features at the current time.

4. The multi-source energy collaborative management method for unmanned sailing vehicles according to claim 3, characterized in that, The multidimensional input matrix is ​​represented as follows: ； in, The multidimensional input matrix obtained by splicing, For splicing operations, For standardized forecast time series data, Historical operating data; The feature fusion layer performs nonlinear mapping fusion in the following way: ； in, As a feature of fusion, For activation function, This is the fusion weight matrix of the feature fusion layer. This is the bias vector.

5. A multi-source energy collaborative management method for unmanned sailing vehicles according to claim 4, characterized in that, The hidden features at the current time point output by the LSTM temporal feature extraction layer are: ； ； ； ； ； ； in, for Hidden features of time, for The output vector of the output gate at each time step. For Hadama accumulation, The hyperbolic tangent activation function is used. for Cellular state at all times It is the Sigmoid activation function. This is the weight matrix of the output gate. for The fusion characteristics of moments for Hidden features of time, This is a vector concatenation symbol. This is the bias vector for the output gate. This is the output vector of the forget gate. for Cellular state at all times The output vector of the input gate. for The state of candidate cells at any given time. Here is the weight matrix for the candidate states. This is the bias vector for the candidate state. Here is the weight matrix of the input gate. This is the bias vector for the input gate. Here is the weight matrix for the forget gate. This is the bias vector for the forget gate.

6. The multi-source energy collaborative management method applied to an unmanned sailing vehicle according to claim 2, characterized in that, The multivariate time series prediction results are expressed as follows: ； in, For multivariate time series prediction results, for The predicted power generation of the photovoltaic array at any given time. for The time-adjustable wing sail mechanism captures wind energy equivalent to predicted mechanical power. for Predicted power consumption for various loads at all times. for The remaining self-continuous flight time at that moment, , This is the prediction time domain for the energy consumption and range prediction model.

7. A multi-source energy collaborative management method for unmanned sailing vehicles according to claim 6, characterized in that, The loss function used during training of the energy consumption and battery life prediction model for: ； ； ； ； in, The overall prediction mean square error, This is the absolute error term for photovoltaic power prediction. This is the absolute error term for wind power prediction. For adjustable weight hyperparameters, for Observations of a multivariate time series at time 10:

00. for Photovoltaic power generation at any given time for Real-time wind energy capture measured power. It is the absolute value symbol.

8. A multi-source energy collaborative management method for unmanned sailing vehicles according to claim 2, characterized in that, The time-domain inequality constraints are as follows: ； in, for The power generation of the photovoltaic array at any time for The maximum power output of the photovoltaic array at any given time. for Forecast of sea surface solar irradiance at specific times. for The adjustable wing sail mechanism captures the equivalent mechanical power of wind energy in real time. for The adjustable wing sail mechanism captures the maximum value of the equivalent mechanical power corresponding to wind energy in real time. for Forecast wind speed and direction at specific times. for The output power of the electric auxiliary thruster at any given time, , These are the minimum and maximum output power of the electric auxiliary thruster, respectively. for The remaining power of the energy storage battery at any given time. , These are the minimum and maximum remaining capacity of the energy storage battery, respectively. for Load power at any given time , These are the minimum and maximum values ​​of the load power, respectively. , This represents the prediction time domain for the MPC controller.

9. A multi-source energy collaborative management method for unmanned sailing vehicles according to claim 8, characterized in that, The objective function is expressed as: ； in, For control sequence Find the objective function The minimum value, for Total system power consumption at any given time. The target value for the remaining capacity of the energy storage battery. for The remaining power of the energy storage battery at any given time. for The change in the electric auxiliary propulsion load power at any given time. for The change in load power at any given time. , , , All are weighting coefficients.

10. A multi-source energy collaborative management method for unmanned sailing vehicles according to claim 8, characterized in that, The MPC controller also adjusts the forecast of sea surface solar irradiance time sequence. And solar irradiance determination threshold The pattern determination is performed as follows: When using an electric-assisted propulsion system in conjunction with an adjustable wing sail mechanism, if the predicted time period threshold exceeds [a certain threshold] within the prediction window... and If the conditions are met, the electric auxiliary thruster will continue to operate, and excess electricity generated by the photovoltaic array will be used to charge the energy storage battery; otherwise, the electric auxiliary thruster will not operate.

Images