CN114638152B - Energy management method for deep-sea Argo profiling float based on HGP-MPC - Google Patents

Abstract

The present invention relates to a deep-sea Argo profile buoy energy management method based on HGP-MPC. A probability distribution model is established based on HGP to estimate the energy consumption of a single profile operation process of a buoy under uncertain sea conditions. After accurately predicting the power demand of the Argo profile buoy, the power prediction result is input into an energy management strategy based on MPC. Combined with the remaining battery power of the buoy, the energy management of the buoy is rolled optimized within the prediction time domain. This scheme adopts a heteroscedastic Gaussian process regression model to predict the power demand of the buoy load, conducts energy management research based on uncertainty quantification results, improves the environmental adaptability of the energy management strategy, adopts a model predictive control method to distribute buoy power, sets a prediction time domain, and performs rolling optimization to achieve real-time decision-making; the power prediction algorithm is combined with a real-time energy management strategy to improve the environmental adaptability of the energy management strategy and ensure real-time performance, and has high practical application and promotion value.

Description

Energy management method for deep-sea Argo profiling float based on HGP-MPC

Technical Field

The invention relates to the field of deep-sea Argo profiling buoy energy management, and in particular to a deep-sea Argo profiling buoy energy management method based on HGP-MPC.

Background technique

Energy is one of the common problems encountered by underwater vehicles. Deep-sea exploration conditions are complex and changeable. For buoys, the density and temperature of seawater at different depths and the speed of buoy movement have a significant impact on the energy consumption of the buoy to complete a single profile movement. The Argo profiling buoy occupies an important position in the field of deep-sea exploration with its excellent real-time, continuity and high efficiency.

There are many factors that affect the energy consumption of buoys. The flight time and service life of Argo profiling buoys are greatly limited by the capacity of the power system and the energy management strategy. The buoy cannot actively adjust its horizontal movement and mainly drifts with the ocean currents. As a result, even if multiple dive tests are conducted in the same sea area, there is a large uncertainty in the energy consumption of the buoy during the entire profiling process. In addition, the buoy working status pre-set by the staff and the real-time ocean current parameters in the operating sea area will lead to uncertainty in the energy consumption required by the buoy. The above factors will affect the buoy's estimation accuracy for the energy consumption of a single profile, thereby reducing the reliability of the number of profile cycles set by the buoy during its life cycle.

In order to better manage the energy of the buoy, after accurately predicting the required power of the Argo profiling buoy, an effective energy management strategy needs to be developed in order to reasonably distribute the power of the buoy during the real-time single profile movement. Among the existing energy management strategies, on the one hand, the rule-based energy management strategy is difficult to adapt to the changing sea conditions; on the other hand, the energy management strategy based on the optimization algorithm faces the problem of fixed working conditions and large amount of calculation, which makes it difficult to meet the needs of real-time applications. Therefore, the development of an optimized energy management strategy with real-time adaptability is an urgent problem to be solved.

Summary of the invention

In order to solve the defects of existing energy management strategies such as difficulty in adapting to changing sea conditions and difficulty in meeting real-time application needs, the present invention proposes a deep-sea Argo profiling buoy energy management method based on HGP-MPC, which can not only improve the environmental adaptability of the energy management strategy, but also ensure the real-time performance of the prediction algorithm.

The present invention is implemented by adopting the following technical scheme: a management method of a deep-sea Argo profiling buoy energy management system based on HGP-MPC, wherein the energy management system comprises a historical operating condition data collection module, a real-time data acquisition module, a terminal processing module, an execution module, a data transmission module and a server processing module; wherein the server processing module comprises a load power prediction unit based on HGP and an energy management unit based on MPC, and is characterized in that the method comprises the following steps:

Step A: Load power prediction based on HGP:

Step A1: Based on the historical operating condition data collection module, the historical operating condition data obtained is transmitted to the server processing module via the terminal processing module;

Step A2: Perform power prediction based on the load power prediction unit of HGP;

(1) Feature selection and data preprocessing: The input and output parameters of the HGP model are determined based on the longitudinal dynamic equation of the Argo profiling float, and the data are cleaned, integrated, transformed, and reduced to improve data quality. After data preprocessing, the data set is randomly divided into training and test sets.

(2) Determine the input characteristics according to the longitudinal dynamics model of the buoy, build a power demand prediction model, select the Gaussian process regression kernel function, and determine the number of model hyperparameters;

The power demand prediction model is as follows:

_yi =f( _xi )+ _εi , _{xi∈R ,} ^yi∈Rd , _εi ~N(0, _σi )

Among them, _xi is the input of the ith sample; _yi is the output of the ith sample, that is, power, both from the training set f(.) is the regression function; ε _i is the output noise, which obeys the Gaussian distribution with mean 0 and variance σ _i ; the heteroscedasticity problem is described as σ _i = r( _xi );

(3) Power demand prediction model training: After determining the number of model hyperparameters, an optimization function is established to optimize the hyperparameters;

(4) Verify the prediction performance of the power demand model, and verify the trained Argo profiling float power demand prediction model through the divided test set to make accurate predictions;

Step B: MPC-based energy management optimization

Step B1, based on the real-time data acquisition module, the real-time working condition data obtained is transmitted to the server processing module via the terminal processing module;

Step B2: Optimize energy management based on MPC energy management unit:

(1) According to the longitudinal dynamic equation of the Argo profiling float, the speed of the driving motor is selected as the control variable;

(2) Set constraints based on the power system equipment specifications and buoy profile motion conditions;

(3) Optimize the objective function to obtain the control variable signal in the prediction time domain;

The power prediction results are input into the MPC-based energy management unit, and the remaining battery power collected by the internal electronic control unit of the buoy is used to optimize the energy management of the buoy within the prediction time domain.

Step C: The driving motor speed instruction in the predicted time domain obtained in step B is sent to the terminal processing module through the data transmission module, and then sent to the buoy execution module by the terminal processing module to realize real-time rolling optimization.

Furthermore, in step B2, during the energy management optimization process:

Firstly, the collected seawater environment parameters, hydraulic system status parameters, and battery status parameters are combined; the load power prediction unit based on HGP is used to predict the power demand in the prediction time domain;

Then, the minimum equivalent energy consumption in the prediction time domain is taken as the objective function, and the control action command in the prediction target time domain, that is, the driving motor speed command, is optimized.

Finally, the command is sent to the actuator, and a single real-time optimization is completed; when the motor completes the action within the specified step time, the above process is repeated, and the rolling optimization obtains the real-time driving motor speed command until the buoy completes the single-profile operation task.

Furthermore, the real-time operating condition data type is consistent with the historical operating condition data type. The historical operating condition data includes environmental parameters collected by the Argo profiling buoy during multiple complete diving and surfacing profiling movements and parameters characterizing the state of the power system. The environmental parameters include seawater temperature, salinity, and pressure. The parameters characterizing the state of the power system include the pressure load of the hydraulic system, the speed of the drive motor, and the battery voltage, current, and remaining power.

Furthermore, in step A2, the input parameters include initial input and augmented input, the initial input includes seawater temperature, salinity and pressure, and the augmented input includes velocity, acceleration and velocity square calculated based on the initial input; the output parameter includes power.

Furthermore, in step A2, the following methods are mainly used when performing data preprocessing:

First, the Laida criterion is used to eliminate outliers, noise and missing values outside the 3σ range of the data;

Secondly, the exponential moving average method is used to smooth the data to reduce the error of the original data caused by measurement error and sensor hysteresis;

Finally, the z-score normalization method is used to standardize the original working condition data to eliminate the dimension and order of magnitude differences between data of different dimensions.

Furthermore, in step A2, when training the power demand model, based on the maximum likelihood estimation method, the negative log marginal likelihood function is established using the training set data as the optimization function, and the non-dominated sorting genetic algorithm NSGA-II with an elite strategy is used to search for the optimal solution of the hyperparameters.

Compared with the prior art, the advantages and positive effects of the present invention are:

This scheme uses the environmental parameters, hydraulic system power parameters, and battery status parameters obtained in the historical sea trials of the Argo profiling buoy as the original data, and establishes a load power prediction model based on the heteroscedastic Gaussian process regression model; then, based on the prediction results, the energy management of the buoy is optimized within the prediction time domain, and the drive motor speed command with the minimum energy consumption is output;

The heteroscedastic Gaussian process regression model can provide the prediction confidence interval of the required power during the buoy operation process, more accurately describe the uncertainty caused by external environmental changes and internal system dynamic performance, and improve the confidence of the prediction results; and through the heteroscedastic Gaussian process regression, the uncertain sea conditions and environmental factors are quantitatively modeled, and energy management research is conducted based on the uncertainty quantification results, which can improve the environmental adaptability of the energy management strategy;

The model predictive control method is used to distribute the buoy power, set the prediction time domain, and perform rolling optimization to achieve real-time decision-making. Combining the power prediction algorithm with the real-time energy management strategy can not only improve the environmental adaptability of the energy management strategy, but also ensure the real-time performance of the prediction algorithm.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of the process principle of the buoy energy management method according to an embodiment of the present invention;

FIG2 is a flow chart of heteroscedastic Gaussian process regression model training according to an embodiment of the present invention.

Detailed ways

In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described below in conjunction with the accompanying drawings and embodiments. In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention can also be implemented in other ways different from those described herein, and therefore, the present invention is not limited to the specific embodiments disclosed below.

With the development and application of artificial intelligence algorithms such as machine learning and deep learning, when the input, output and environmental data of the Argo profile buoy power system under historical operating conditions are obtained, they can be used as training data to train the model to predict the power demand of the buoy power system under the operating conditions. After the model is reasonably selected and the parameters are adjusted, the power demand of the buoy power system under the operating conditions can be predicted. In order to accurately predict the power demand of the buoy power system and characterize the uncertainty of the prediction results, the present invention adopts the heteroscedastic Gaussian Process Regression (HGP) model to establish a power demand prediction model for the buoy under different operating conditions. The Gaussian Process Regression (GPR) model is a regression model based on the Bayesian probability framework. This method quantifies the uncertainty of the predicted variables through probabilistic reasoning, provides a prediction mean and variance to characterize the uncertainty measurement, and can establish a probability model of the buoy power demand according to the characteristics of seawater density, pressure, buoy speed, acceleration, etc. At the same time, considering that the degree of influence of environmental factors on the required power under different working conditions varies, the heteroscedastic model is used for evaluation to more accurately describe the influence of the external environment and internal system state on the buoy, and the differences in power demand at different sea depths.

The energy management strategy configured by the buoy must meet the requirements of real-time, rapidity and effectiveness. Different from the huge amount of calculation of the dynamic programming algorithm, the model predictive control method performs energy management through real-time online rolling optimization, with low real-time calculation amount and strong anti-disturbance and dynamic correction capabilities. The present invention adopts model predictive control (MPC) to build a real-time energy management strategy for the buoy, uses the load power prediction unit based on HGP to obtain the required load power in the predicted time domain, establishes an optimization function with the minimum equivalent energy consumption, and performs rolling optimization with the drive motor speed as the control target variable.

Specifically, this embodiment proposes a management method for a deep-sea Argo profiling buoy energy management system based on HGP-MPC, wherein the energy management system includes a historical operating data collection module, a real-time data collection module, a terminal processing module, and a server processing module, wherein the historical operating data collection module and the real-time data collection module are both connected to the terminal processing module, the terminal processing module is connected to the Argo buoy through an execution module, and the terminal processing module is connected to the server processing module through a data transmission module, wherein the server processing module includes a load power prediction unit based on HGP and an energy management unit based on MPC, and the specific implementation steps of the energy management method are shown in FIG1, including:

Step A: Load power prediction based on HGP:

Step A1: Collect complete profile data in the Argo buoy sea trial based on the historical operating condition data collection module, and transmit the obtained historical operating condition data to the terminal processing module; in the terminal processing module, send the above historical operating condition data to the server processing module through the data transmission module, and perform data processing and model calculation based on the load power prediction unit of HGP;

The historical working condition data collection module mainly uses the sensor equipment of the buoy equipment to dynamically obtain environmental parameters, uses the sensors carried by the hydraulic system to collect the hydraulic system state parameters, and uses the internal electronic control unit of the buoy equipment to collect the battery state parameters. The above parameters are all complete profile data obtained during the Argo buoy sea trial. Among them, the buoy equipment refers to the deep-sea self-sustaining intelligent Argo profiling buoy system; the sensor equipment includes a temperature-salinity-depth profiler, a pressure sensor, and an electronic control system parameter acquisition system. The working condition refers to the combination of the operating state and environmental data of the buoy power system collected when the buoy is performing the profile movement, including the two stages of diving and floating. Correspondingly, the historical working condition data refers to the environmental parameters such as seawater temperature, salinity, pressure, etc. collected by the Argo profiling buoy during the completion of multiple complete diving and floating profile movements, the pressure load of the hydraulic system, the speed of the drive motor, the battery voltage, current, and the remaining power, etc., which characterize the power system state. In this embodiment, the working condition data mainly comes from the performance test experiments of the deep-sea Argo buoy organized by my country in the Mariana Trench.

Step A2: Perform power prediction based on the load power prediction unit of HGP;

(1) Feature selection based on professional knowledge and machine learning model characteristics, and preprocessing of historical data;

Feature selection refers to the determination of the input and output of the HGP model according to the longitudinal dynamic equation of the Argo profiling buoy when training the HGP power prediction model. The initial input is the seawater temperature, salinity, and pressure, and the seawater depth information is converted based on the initial input. The velocity, acceleration, and velocity square are calculated by first-order and second-order differences as augmented inputs. The model input consists of two parts: the initial input and the augmented input. The output is the power calculated based on the battery voltage and current, as shown in Appendix 1.

Table 1 Input and output selection of heteroskedastic Gaussian process regression model

Data preprocessing refers to data cleaning, data integration, data changes and data reduction to improve data quality. The main process is: first, use the Laida criterion to eliminate outliers, noise and missing values outside the 3σ range of the data; second, use the exponential moving average method to smooth the data to reduce the error of the original data due to measurement error and sensor hysteresis; finally, use the z-score standardization method to standardize the original working condition data, eliminate the dimension and order of magnitude differences between data of different dimensions, and organize the data format.

After data preprocessing, the data set is randomly divided into training set and test set. 80% of the obtained data set is randomly selected as the training set for model training; the remaining 20% is used as the test set to verify the model prediction ability.

(2) Based on the model input dimension, a power demand model is constructed and the Gaussian process regression kernel function is selected;

When constructing the power demand model (model initialization), considering that different inputs correspond to different output distributions, the noise variance used to fit the output distribution is also different. This embodiment adopts a heteroscedastic Gaussian process regression model (HGP), and the expression is shown in Appendix 2.

Table 2 Mathematical expression of heteroskedastic Gaussian process regression model

Kernel function selection refers to using multiple kernel functions (including but not limited to the kernel functions shown in Appendix 3) to perform regression fitting on the training set data, and selecting the kernel function with the best fitting effect as the kernel function used in the Gaussian process regression model training unit.

Table 3 Kernel function types for regression fitting

This embodiment uses a square exponential kernel as the kernel function used in the Gaussian process regression model training unit.

(3) Gaussian process regression model training: After determining the number of model hyperparameters in step (2), an optimization function is established and an intelligent optimization algorithm is used to optimize the hyperparameters;

Gaussian process regression model training refers to initializing the prediction model form, determining the number of hyperparameters that need to be optimized in the kernel function and the noise model, and deriving the hyperparameter optimization function based on the maximum likelihood estimation method; using the conditional probability of the training set samples to establish the negative logarithmic maximum marginal likelihood function about the hyperparameters, and using this as the objective function; with the purpose of minimizing the aforementioned objective function, using the non-dominated sorting genetic algorithm with elite strategy (NSGA-II) to optimize the hyperparameters and obtain the optimal solution for the hyperparameters; the training process is shown in Figure 2.

(4) Use the test set obtained in the model preprocessing stage to verify the model prediction performance and achieve accurate prediction of the required power of the Argo profiling float;

Test set validation refers to the completion of HGP model training and the verification of the model’s prediction performance on the test set. The indicators used include the accuracy indicators RMSE and ^R2 and the interval estimation indicators CP and MWP, see Appendix 4.

Table 4 Mathematical expressions of precision evaluation index and interval estimation evaluation index

The model prediction step refers to obtaining the predicted input before the real-time rolling optimization, inputting it into the trained heteroscedastic Gaussian process regression model, and performing output calculation according to the joint probability distribution formula in Table 5 to obtain the load power set by the present invention.

Table 5 Mathematical expressions for the prediction output calculation of heteroskedastic Gaussian process regression model

After the Argo profiling float's required power is accurately predicted, the power prediction result is input into the MPC-based energy management unit. Together with the remaining battery power collected by the float's internal electronic control unit, the float's energy management is optimized within the prediction time domain, the float's remaining power is reasonably allocated, and the drive motor speed command with the minimum energy consumption is output. Specifically:

Step B: MPC-based energy management optimization:

Step B1, based on the real-time data acquisition module, collect seawater environmental parameters, hydraulic system status parameters, battery status parameters and other data in real time, and transmit the obtained real-time working condition data to the terminal processing module; in the terminal processing module, send the above real-time collected data to the server processing module through the data transmission module, and perform data processing and model calculation based on the MPC energy management unit. The real-time data is consistent with the data type of the historical working condition, referring to the environmental parameters such as seawater temperature, salinity, pressure, hydraulic system load pressure, battery voltage, current, remaining power, and drive motor speed collected in real time when the buoy is undergoing profile testing.

Step B2: Optimize energy management based on MPC energy management unit:

(1) According to the longitudinal dynamic equation of the Argo profiling float, the speed of the driving motor is selected as the control variable;

(2) Set constraints based on the power system equipment specifications and buoy profile motion conditions;

Constraint setting refers to ensuring that the optimization results meet the actual buoy operation restrictions and objective conditions when performing MPC energy management optimization, setting the maximum operating current of the battery, the effective state of charge range of the battery, the maximum operating speed of the buoy, the maximum operating acceleration of the buoy, and the maximum speed of the drive motor. Among them: considering the battery discharge safety, the maximum operating current of the battery shall not exceed the maximum operating safety current of the lithium battery pack I _s ; considering the theoretical safe discharge range of the battery, the effective state of charge of the battery is limited to the charge range [SOC _min , SOC _max ]; considering the buoy operation stability and data collection continuity, the maximum operating speed of the buoy shall not exceed 150% of the internationally accepted ideal buoy operation speed (0.1m/s), and the maximum operating acceleration of the buoy shall not exceed 0.05m/s ² ; considering the operation stability of the motor and the buoy, the maximum speed of the drive motor shall not exceed 150% of the rated speed of the motor.

(3) The objective function set by the optimization is to obtain the control variable signal in the prediction time domain and send the control instruction to the data transmission module;

The objective function refers to the minimum equivalent energy consumption of the Argo profiling float in the prediction time domain; the objective function optimization refers to optimizing the objective function to obtain the control variable signal in the prediction time domain; the control variable refers to the speed command of the drive motor.

The energy management optimization process refers to first transmitting data such as seawater environment parameters, hydraulic system status parameters, and battery status parameters to the energy optimization unit; then using the HGP-based load power prediction unit to predict the power demand in the prediction time domain; then taking the minimum equivalent energy consumption in the prediction time domain as the objective function, optimizing the control action instructions in the prediction target time domain, that is, the drive motor speed instructions; sending the instructions to the actuator, and completing a single real-time optimization; when the motor completes the action within the specified step time, repeating the above process, rolling optimization to obtain the real-time drive motor speed instructions until the buoy completes the single-profile operation task.

Step C: The driving motor speed command in the future prediction time domain calculated by the service processing module is sent to the terminal processing module through the data transmission module, and then sent to the buoy execution module by the terminal processing module to realize real-time rolling optimization.

In order to ensure the real-time adaptability of the energy management strategy, first of all, it is necessary to estimate the required voltage and current of the buoy propulsion process for uncertain sea conditions. The complex changes in sea conditions make it difficult to estimate voltage and current. Conventional dynamic models are difficult to meet the prediction requirements, so a probability distribution model is established based on HGP to achieve this. Secondly, to achieve the goal of real-time performance, it is necessary to design an optimization algorithm that can meet the requirements of fast calculations and output the power distribution results in real time according to the torque requirements. This solution introduces the model predictive control method, sets the prediction time domain, and performs rolling optimization to achieve real-time decision-making and output the control variable signal-the speed of the drive motor, so as to achieve timely drive and optimization of the buoy.

In addition, it should be emphasized that the energy management optimization solution described in the present invention, that is, the proposed power prediction algorithm and real-time energy management strategy are also applicable to other autonomous underwater vehicles.

The above description is only a preferred embodiment of the present invention and does not limit the present invention in other forms. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims (5)

1. A management method for a deep-sea Argo profiling buoy energy management system based on HGP-MPC, wherein the energy management system comprises a historical operating condition data collection module, a real-time data acquisition module, a terminal processing module, an execution module, a data transmission module and a server processing module; wherein the server processing module comprises a load power prediction unit based on HGP and an energy management unit based on MPC, characterized in that the method comprises the following steps: Step A: Load power prediction based on HGP: Step A1: Based on the historical operating condition data collection module, the historical operating condition data obtained is transmitted to the server processing module via the terminal processing module; Step A2: Perform power prediction based on the load power prediction unit of HGP; (1) Feature selection and data preprocessing: The input and output parameters of the HGP model are determined based on the longitudinal dynamic equation of the Argo profiling float, and the data are cleaned, integrated, transformed, and reduced to improve data quality. After data preprocessing, the data set is randomly divided into training and test sets. (2) Determine the input characteristics according to the longitudinal dynamics model of the buoy, build a power demand prediction model, select the Gaussian process regression kernel function, and determine the number of model hyperparameters; The power demand prediction model is as follows: _yi =f( _xi )+ _εi , _{xi∈R ,} ^yi∈Rd , _εi ~N(0, _σi ) Among them, _xi is the input of the ith sample; _yi is the output of the ith sample, that is, power, both from the training set f(.) is the regression function; ε _i is the output noise, which obeys the Gaussian distribution with mean 0 and variance σ _i ; the heteroscedasticity problem is described as σ _i = r( _xi ); (3) Power demand prediction model training: After determining the number of model hyperparameters, an optimization function is established to optimize the hyperparameters; (4) Verify the prediction performance of the power demand model, and verify the trained Argo profiling float power demand prediction model through the divided test set to make accurate predictions; Step B: MPC-based energy management optimization Step B1, based on the real-time data acquisition module, the real-time working condition data obtained is transmitted to the server processing module via the terminal processing module; Step B2: Optimize energy management based on MPC energy management unit: (1) According to the longitudinal dynamic equation of the Argo profiling float, the speed of the driving motor is selected as the control variable; (2) Set constraints based on the power system equipment specifications and buoy profile motion conditions; (3) Optimize the objective function to obtain the control variable signal in the prediction time domain; The power prediction results are input into the MPC-based energy management unit, and the remaining battery power collected by the internal electronic control unit of the buoy is used to optimize the energy management of the buoy within the prediction time domain. Step C: The driving motor speed instruction in the predicted time domain obtained in step B is sent to the terminal processing module through the data transmission module, and then sent to the buoy execution module by the terminal processing module to realize real-time rolling optimization. 2. The management method of the deep-sea Argo profiling buoy energy management system based on HGP-MPC according to claim 1 is characterized in that: the real-time operating condition data type is consistent with the historical operating condition data type, and the historical operating condition data includes environmental parameters collected by the Argo profiling buoy when completing multiple complete diving and floating profiling movements and parameters characterizing the state of the power system, the environmental parameters include seawater temperature, salinity, and pressure, and the parameters characterizing the state of the power system include the pressure load of the hydraulic system, the speed of the drive motor, the battery voltage, current, and remaining power. 3. The management method of the deep-sea Argo profiling buoy energy management system based on HGP-MPC according to claim 2 is characterized in that: in the step A2, the input parameters include initial input and augmented input, the initial input includes seawater temperature, salinity and pressure, and the augmented input includes velocity, acceleration and velocity square calculated based on the initial input; the output parameter includes power. 4. The management method of the deep-sea Argo profiling buoy energy management system based on HGP-MPC according to claim 1 is characterized in that: in the step A2, when performing data preprocessing, the following methods are mainly adopted: First, the Laida criterion is used to eliminate outliers, noise and missing values outside the 3σ range of the data; Secondly, the exponential moving average method is used to smooth the data to reduce the error of the original data caused by measurement error and sensor hysteresis; Finally, the z-score normalization method is used to standardize the original working condition data to eliminate the dimension and order of magnitude differences between data of different dimensions. 5. The management method of the deep-sea Argo profiling buoy energy management system based on HGP-MPC according to claim 1 is characterized in that: in the step A2, when training the power demand model, based on the maximum likelihood estimation method, a negative logarithmic marginal likelihood function is established using the training set data as the optimization function, and a non-dominated sorting genetic algorithm NSGA-II with an elite strategy is used to search for the optimal solution of the hyperparameters.