A new energy driving ocean monitoring terminal energy distribution optimization method

Abstract

The present application provides a kind of new energy drive ocean monitoring terminal energy distribution optimization method, belong to marine equipment technical field, the real-time data of the present application is collected wave energy converter, photovoltaic array and battery, constructs space-time sequence streaming feature tensor, input to multidimensional feature adaptive sparse activation hybrid expert energy regulation and control model;The artificial intelligence model uses Markov chain state prediction gate routing weight adaptive updating algorithm, perceives energy deficiency trend in advance, adjusts charge-discharge rate adaptively according to electrochemical damage risk assessment function;Execute asymmetric parallel pipeline decoupling mechanism to reduce processor power consumption;When the state of charge is lower than the safety threshold, the non-core monitoring module is forced to close, and the core sensing link power supply is preferentially guaranteed, which solves the technical problems that the ocean monitoring terminal cannot real-time adaptive regulation and control of computing resource allocation and energy distribution strategy under the condition of dynamic energy state change, and the system fails due to high computing load when energy drops suddenly.

Description

Technical Field

[0001] This invention belongs to the field of marine equipment technology, and more specifically, relates to a method for optimizing energy distribution in a new energy-driven marine monitoring terminal. Background Technology

[0002] Marine monitoring terminals are typically deployed in extreme environments such as open seas or islands, relying on wave energy converters and photovoltaic arrays as their primary power sources, supplemented by batteries for energy buffering. Current technologies for marine monitoring terminals driven by new energy sources generally employ energy management strategies based on fixed rules or static thresholds, combined with traditional neural network models for predicting input power and controlling charging and discharging. In current marine monitoring terminal energy management systems, due to the highly non-stationary and multi-source coupling of wave energy and photovoltaic output power, existing fixed-structure neural network models maintain full computational load even under sudden energy drops, resulting in persistently high processor power consumption and further exacerbating energy consumption. Simultaneously, fixed-threshold gating strategies cannot perceive the temporal evolution of energy states, failing to reduce computational scale in advance during periods of extreme energy scarcity, causing system failure due to the depletion of both computational and energy resources. In other words, existing technologies suffer from the technical problem of marine monitoring terminals being unable to adaptively adjust computational resource allocation and energy allocation strategies in real time under dynamically changing energy states, leading to system failure due to excessive computational load during sudden energy drops. Summary of the Invention

[0003] In view of this, the present invention provides a new energy-driven marine monitoring terminal energy allocation optimization method, which can solve the technical problem in the prior art that the marine monitoring terminal cannot adaptively adjust the allocation of computing resources and energy allocation strategies in real time under dynamic changes in energy status, resulting in the system failing due to excessive computing load when energy drops sharply.

[0004] This invention is implemented as follows: This invention provides a method for optimizing energy distribution in a new energy-driven marine monitoring terminal, comprising the following steps:

[0005] Real-time output power of wave energy converter and photovoltaic array, state of charge of battery and marine environmental parameters are collected to construct spatiotemporal sequence streaming feature tensor, which is then input into multidimensional feature adaptive sparse activation hybrid expert energy regulation model;

[0006] The multidimensional feature adaptive sparse activation hybrid expert energy regulation model projects meteorological and ocean current features to independent subspaces through a spatial orthogonal decoupling operator. The prior probability of the next period of extreme energy scarcity is calculated by the Markov chain state prediction gating routing weight adaptive update algorithm. The sparse activation gating routing weight is dynamically adjusted, and the output power fluctuation smoothing command and charging and discharging control parameters are output.

[0007] Based on the transient nonlinear energy buffer mathematical model, a multi-level decoupled spatiotemporal attention operator is performed on wave energy and photovoltaic input power to adaptively smooth the input power fluctuation within a sub-second time window and output the smoothed charging and discharging current command.

[0008] Calculate the electrochemical damage risk assessment function value, and adaptively adjust the charge and discharge rate parameters according to the interval to which the function value belongs, so as to protect the microscopic physical characteristics of the battery.

[0009] An asymmetric parallel pipeline decoupling mechanism is implemented to separate the power prediction channel and the control instruction generation channel into heterogeneous multi-channel independent operation. A lightweight asynchronous lock mechanism is used to eliminate synchronization wait between threads and reduce the power consumption of the core processor under high load.

[0010] Based on the battery state of charge and the predicted input power, the energy distribution mode is dynamically switched. When the state of charge is lower than the safety threshold, the preset non-core monitoring modules are forcibly shut down to prioritize the power supply of the core sensing link.

[0011] The dimension of the spatiotemporal sequence streaming feature tensor is defined as a three-dimensional tensor composed of batch size, time step and feature dimension. The specific values ​​of batch size, time step and feature dimension are determined by deploying experiments multiple times in different marine environments and analyzing the model convergence speed and inference delay.

[0012] Among them, the spatial orthogonal decoupling operator separates multi-source environmental features into independent sub-feature spaces that do not interfere with each other through orthogonal projection transformation. The inner product between each sub-space is zero, thereby eliminating the mutual interference between meteorological features and ocean current features.

[0013] The multidimensional feature adaptive sparse activation hybrid expert energy regulation model includes an input layer, a sparse activation gated routing layer, and an output layer. When the battery state of charge is lower than the safety threshold, the gate function forcibly increases the sparsity of the routing probability vector through a step nonlinear transformation, activating only two sets of low-power expert networks, while the remaining expert networks enter a dormant state.

[0014] Within each expert network, a long-span residual dense jump matrix is ​​constructed to achieve dimensional alignment and direct summation of deep and shallow tensors. Dynamic feedforward network routing is introduced into the feedforward network, and conditional branch jumps are executed by the gating unit according to the nonlinear complexity of the input environment data, dynamically pruning inactive neuron blocks.

[0015] Among them, the Markov chain state prediction gating routing weight adaptive update algorithm constructs a multi-order state transition probability matrix by statistically analyzing the historical power generation and energy consumption states of the terminal. The calculated prior probabilities are directly used as independent variables to input sparse activation gating, and the activation thresholds of each expert network are adjusted instantaneously. The order of the multi-order state transition probability matrix is ​​determined by comparative experiments on the state prediction accuracy under different sea states.

[0016] The training dataset for the multidimensional feature adaptive sparse activation hybrid expert energy regulation model includes output power time-series data, battery state of charge data, marine meteorological data, and ocean current data continuously collected in the target sea area. The collection duration is not less than the collection duration threshold, and power mutation events are recorded at a sub-second sampling frequency.

[0017] Among them, the training of the multi-dimensional feature adaptive sparse activation hybrid expert energy regulation model adopts a multi-task loss function to jointly optimize the power smoothing accuracy and charge-discharge control accuracy, and uses the battery state of charge as an auxiliary supervision signal to train the sparse activation threshold of the gated routing layer.

[0018] Among them, the multi-level decoupled spatiotemporal attention operator decouples the time dimension attention and the spatial dimension attention in layers and applies them sequentially to the input power sequence; the time dimension attention captures the sub-second dynamic changes of power fluctuations, and the spatial dimension attention distinguishes the power components from wave energy and photovoltaic sources. After the two are decoupled and operated, they are fused to output a smooth result.

[0019] The electrochemical damage risk assessment function uses the current rate of change of charge / discharge current, the battery state of charge, and the ambient temperature as inputs to calculate the electrochemical damage risk value. The calculation formula is ,in , , The dimensionless coefficient was determined by regression analysis after multiple sets of charge-discharge cycle experiments and electrochemical impedance spectroscopy.

[0020] Among them, when When the risk level is below the low-risk threshold, the charge / discharge rate parameter is adjusted to the high-rate range; when When the risk level is between the low-risk and high-risk thresholds, the charge / discharge rate parameter remains in the medium-rate range; when When the risk level is not lower than the high-risk threshold, the charge / discharge rate parameter is forcibly compressed to the low-rate range.

[0021] In the asymmetric parallel pipeline decoupling mechanism, the power prediction channel is allocated more computing cores and memory resources, while the control instruction generation channel is allocated lightweight computing resources. Data synchronization between the two channels is achieved through a lightweight asynchronous lock mechanism.

[0022] Among them, the lightweight asynchronous lock mechanism replaces the traditional mutex lock with non-blocking atomic operations; after the power prediction channel completes the calculation, it updates the shared data area with atomic write operations, and the control instruction generation channel obtains the latest prediction result with atomic read operations, so the two do not need to wait for each other.

[0023] The safety threshold is determined by measuring the actual energy consumption of the monitoring terminal in the target sea area and analyzing the battery discharge curve, combined with the inflection point of the sudden drop in battery voltage during multiple discharge experiments.

[0024] The batch size of the spatiotemporal sequence streaming feature tensor ranges from 1 to 16, the time step from 10 to 60, and the feature dimension from 8 to 32. The number of expert networks ranges from 4 to 8, the cross-layer span of the long-span residual dense jump matrix ranges from 2 to 6 layers, and the pruning ratio of the dynamic feedforward network routing ranges from 20% to 60%. The order of the multi-order state transition probability matrix ranges from 2 to 5. The data acquisition duration threshold is 6 months. , , The initial range is 0.1 to 2.0, which is the low-risk threshold. The range is 0.2 to 0.5, which is the high-risk threshold. The range is 0.6 to 0.9; the safety threshold range is 15% to 30%.

[0025] This invention constructs a multi-dimensional feature-adaptive sparse activation hybrid expert energy regulation model. Utilizing a Markov chain state prediction gating route weight adaptive update algorithm, it anticipates scarcity trends at the moment of energy state evolution, driving the sparse activation gating route to dynamically shrink the number of activation expert networks, thus adaptively compressing the model's effective computational load with energy state changes. Simultaneously, an asymmetric parallel pipeline decoupling mechanism heterogeneously separates the prediction and control channels, eliminating thread synchronization waiting through a lightweight asynchronous lock mechanism, further reducing the core processor's power consumption under high load. An electrochemical damage risk assessment function and a multi-level decoupled spatiotemporal attention operator collaboratively ensure battery safety and power smoothing, enabling the terminal to maintain power supply to the core sensing link even under sudden energy drops. In summary, this invention solves the technical problem mentioned in the background art: marine monitoring terminals cannot adaptively adjust computational resource allocation and energy allocation strategies in real time under dynamically changing energy states, leading to system failure due to excessive computational load during sudden energy drops. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 This is a graph showing the relationship between the state of charge, the number of activated expert networks, and inference power consumption.

[0028] Figure 3A comparison curve of charging and discharging current fluctuations before and after multi-level decoupled spatiotemporal attention operator processing.

[0029] Figure 4 This is a graph showing the change in the state of charge of the battery during a 72-hour test cycle. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0031] like Figure 1 The diagram shown is a flowchart of a new energy-driven marine monitoring terminal energy allocation optimization method provided by the present invention. This method includes the following steps:

[0032] S01. Collect the real-time output power of the wave energy converter and photovoltaic array, the state of charge of the battery and marine environmental parameters, construct the spatiotemporal sequence streaming feature tensor, and input it into the multidimensional feature adaptive sparse activation hybrid expert energy regulation model.

[0033] S02. The multidimensional feature adaptive sparse activation hybrid expert energy regulation model projects meteorological and ocean current features to an independent subspace through a spatial orthogonal decoupling operator. The prior probability of the next period of extreme energy scarcity is calculated by the Markov chain state prediction gated routing weight adaptive update algorithm. The sparse activation gated routing weight is dynamically adjusted, and the output power fluctuation smoothing command and charging and discharging control parameters are output.

[0034] S03. Based on the transient nonlinear energy buffer mathematical model, perform multi-level decoupled spatiotemporal attention operator processing on wave energy and photovoltaic input power, adaptively smooth the input power fluctuation within the sub-second time window, and output the smoothed charging and discharging current command.

[0035] S04. Calculate the electrochemical damage risk assessment function value, and adaptively adjust the charge and discharge rate parameters according to the interval to which the function value belongs, so as to protect the microscopic physical characteristics of the battery.

[0036] S05. Implement an asymmetric parallel pipeline decoupling mechanism to separate the power prediction channel and the control instruction generation channel into heterogeneous multi-channel independent operation, and eliminate inter-thread synchronization waiting through a lightweight asynchronous lock mechanism to reduce the power consumption of the core processor under high load.

[0037] S06. Based on the battery state of charge and the predicted input power, dynamically switch the energy distribution mode. When the state of charge is lower than the safety threshold, forcibly shut down the preset non-core monitoring modules and prioritize power supply to the core sensing link.

[0038] The dimension of the spatiotemporal sequence streaming feature tensor is defined as a three-dimensional tensor composed of batch size, time step and feature dimension. The batch size is 1 to 16, the time step is 10 to 60, and the feature dimension is 8 to 32. The specific values ​​were determined by deploying experiments in different marine environments and analyzing the model convergence speed and inference delay.

[0039] Among them, the spatial orthogonal decoupling operator refers to a linear transformation operator that separates multi-source environmental features into independent sub-feature spaces that do not interfere with each other through orthogonal projection transformation. The inner product between each sub-space is zero, thereby eliminating the mutual interference between meteorological features and ocean current features, making subsequent gated route calculations more targeted.

[0040] Among them, the multidimensional feature adaptive sparse activation hybrid expert energy regulation model is a prediction and control model based on deep learning. Its specific structure is as follows: The input layer receives the spatiotemporal sequence streaming feature tensor, and projects the meteorological and ocean current features to independent subspaces through the spatial orthogonal decoupling operator to form an orthogonalized feature stream; the orthogonalized feature stream enters the sparse activation gated routing layer, which is driven by the Markov chain state prediction gated routing weight adaptive update algorithm to calculate the activation probability vector of each expert network in real time; when the battery state of charge is lower than the safety threshold, the gate function forcibly increases the sparsity of the routing probability vector through a step nonlinear transformation, activating only two groups of low-power expert networks, while the remaining expert networks enter a dormant state, realizing the dynamic allocation of neuron activation weights among different expert networks; each group of expert The network internally constructs a long-span residual dense jump matrix to achieve dimensional alignment and direct summation of deep and shallow tensors, forming a high-speed gradient path across layers. Dynamic feedforward routing is introduced into the feedforward network. Based on the nonlinear complexity of the input environment data, the gated unit executes conditional branch jumps, dynamically pruning inactive neuron blocks and directly skipping useless forward propagation. The memory allocation for each expert network is dynamically scaled according to the charge state. In a high charge state, all expert networks are activated and allocated full memory; in a low charge state, only low-power expert networks are allocated memory, and the rest are released. The cache allocation for inter-layer data loops is linearly compressed based on the number of currently active experts to reduce redundant data residency. The output layer fuses the outputs of each activated expert network to generate power fluctuation smoothing instructions and charging / discharging control parameter sequences. The steps for establishing the training dataset of the multidimensional feature adaptive sparse activation hybrid expert energy regulation model specifically include: continuously collecting output power time-series data, battery state-of-charge data, marine meteorological data, and ocean current data for no less than 6 months from wave energy converters and photovoltaic arrays in the target sea area; recording power mutation events at a sub-second sampling frequency; labeling the corresponding charging and discharging control actions and changes in battery health status; constructing labeled time-series sample pairs; and expanding the sample size through a sliding window method; simultaneously collecting historical operation records of the terminal under extremely low energy conditions as sparse activation positive samples to balance the ratio of positive to negative samples. The training steps of the multidimensional feature adaptive sparse activation hybrid expert energy regulation model specifically include: jointly optimizing power smoothing accuracy and charge / discharge control accuracy using a multi-task loss function; training the sparse activation threshold of the gated routing layer using the battery state of charge as an auxiliary supervision signal; and adjusting the number of expert networks, the cross-layer span of the long-span residual dense jump matrix, and the pruning ratio of the dynamic feedforward network routing on the training and validation sets through iterative experiments. The specific values ​​are determined by the hyperparameter combination corresponding to the lowest point of loss in the validation set in multiple rounds of comparative experiments. It is recommended that the number of expert networks be 4 to 8, the cross-layer span be 2 to 6, and the pruning ratio be 20% to 60%.

[0041] The Markov chain state prediction gating routing weight adaptive update algorithm constructs a multi-order state transition probability matrix by statistically analyzing the historical power generation and energy consumption states of the terminal. At the instant of streaming data input, it calculates the prior probability that the terminal will be in a state of extreme energy scarcity in the next time period. This prior probability is directly used as the independent variable input to the sparse activation gating, instantly adjusting the activation thresholds of each expert network. If the predicted probability of future energy inflow is low, the algorithm automatically increases the gating sparsity, forcibly shutting down heavily computational expert networks, causing the network computation flow to redirect to an ultra-lightweight control branch, thus achieving instantaneous adjustment and dynamic evolution of the expert network activation weights driven by the Markov state chain. This algorithm is organically nested within the gating mechanism of the hybrid expert model. The recommended order of the multi-order state transition probability matrix is ​​2 to 5, with the specific order determined through comparative experiments on the accuracy of state prediction under different sea states. The Markov chain state prediction gated routing weight adaptive update algorithm introduces historical prior information about energy states, so that gated routing no longer depends on a single instantaneous input, but integrates the temporal evolution law of energy states. This allows the expert network activation range to be narrowed in advance in scenarios of sudden energy drops, eliminating unnecessary forward propagation calculations before triggering, significantly reducing the real-time computational overhead of the system under extreme sea states, ensuring the real-time performance and robustness of energy allocation decisions, and enabling the model to have self-organizing and adjusting capabilities in response to different ocean energy states, avoiding the failure of fixed threshold strategies in non-stationary environments.

[0042] Among them, the multidimensional feature adaptive sparse activation hybrid expert energy regulation model dynamically selects a subset of activated expert networks under different energy states through a sparse activation gated routing layer, so that the effective computational load of the model adaptively shrinks with the system energy state, fundamentally solving the problem of high consumption of computing resources in marine terminals; the long-span residual dense jump matrix ensures high-speed gradient transmission between deep and shallow features, avoiding gradient vanishing due to increasing layer depth; the dynamic feedforward network routing achieves fine pruning at the neuron level through conditional branch jumps, further compressing the memory occupation of non-critical computational paths; the prior-driven Markov chain state prediction makes the gated sparsity forward-looking, enabling the entire model to spontaneously evolve into an ultra-lightweight computational form when energy is tight, and restore a high-precision fully activated form when energy is abundant, realizing the dynamic coupling of computational resource allocation and system energy state, and ensuring the real-time performance and system stability of energy allocation optimization from the principle level.

[0043] Among them, the transient nonlinear energy buffer mathematical model refers to the mathematical description model established for the high-frequency nonlinear abrupt change characteristics of wave energy and photovoltaic input power. It is used to quantify the transient impact of input power fluctuations on the electrochemical characteristics of the battery and to provide smooth target constraints for the multi-level decoupled spatiotemporal attention operator.

[0044] Among them, the multi-level decoupled spatiotemporal attention operator refers to an operator that decouples the time dimension attention and the spatial dimension attention in layers and applies them sequentially to the input power sequence. The time dimension attention captures the sub-second dynamic changes of power fluctuations, while the spatial dimension attention distinguishes between the power components from wave energy and photovoltaic sources. After the two are decoupled and operated, they are fused to output a smooth result, avoiding the interference introduced by spatiotemporal coupling calculation.

[0045] The electrochemical damage risk assessment function is used to adjust the battery charge / discharge rate parameters. This function takes the current charge / discharge current change rate, battery state of charge, and ambient temperature as inputs to calculate the electrochemical damage risk value. The calculation formula is ;in The rate of change of charging and discharging current. It is a nonlinear mapping function of the charged state. This is a temperature correction function. , , To ensure dimensionless coefficients, regression analysis was performed using multiple sets of charge-discharge cycle experiments and electrochemical impedance spectroscopy analysis. An initial range of 0.1–2.0 is recommended. < When the charge / discharge rate parameter is adjusted to the high rate range, the charge / discharge response is accelerated; when At this time, the charge / discharge rate parameter is maintained in the medium rate range to balance the charge / discharge speed and the risk of damage; when At this time, the charge / discharge rate parameters are forcibly compressed to a low rate range to suppress electrochemical damage. and The specific values ​​were determined through multiple sets of charge-discharge cycle aging experiments at different temperatures and under different charging states, using the risk value corresponding to the inflection point of battery capacity decay as a benchmark, combined with statistical analysis. Recommendations are provided. The range is 0.2 to 0.5. The range is 0.6 to 0.9.

[0046] The asymmetric parallel pipeline decoupling mechanism refers to designing the power prediction pipeline and the control instruction generation pipeline as two heterogeneous independent pipelines with asymmetric allocation of computing resources. The power prediction pipeline is allocated more computing cores and memory resources to ensure prediction accuracy, while the control instruction generation pipeline is allocated lightweight computing resources to ensure real-time instruction response. The two pipelines are synchronized through a lightweight asynchronous lock mechanism to eliminate inter-thread blocking and waiting.

[0047] The lightweight asynchronous lock mechanism refers to a thread synchronization mechanism that replaces the traditional mutex lock with non-blocking atomic operations. After the energy prediction channel completes the calculation, the shared data area is updated with an atomic write operation. The control instruction generation channel obtains the latest prediction result with an atomic read operation. The two do not need to wait for each other, thereby eliminating the risk of deadlock and reducing the computational latency introduced by thread synchronization.

[0048] Among them, the safety threshold refers to the minimum value of the battery charge state that ensures the battery does not suffer over-discharge damage and the core sensing link can be continuously powered. The safety threshold is determined by measuring the actual energy consumption of the target sea area monitoring terminal and analyzing the battery discharge curve, combined with the inflection point of the battery voltage drop in multiple discharge experiments. The recommended range is 15% to 30%.

[0049] Optionally, the present invention also provides an implementation method using a computer as the running carrier to form a new energy-driven marine monitoring terminal energy distribution optimization system. The system integrates hardware interfaces for wave energy converters, photovoltaic arrays, and batteries. The computer is equipped with a readable storage medium, which stores program instructions. When the program instructions are run on the computer, they execute all the steps of the above-mentioned energy distribution optimization method.

[0050] The specific implementation of step S01 is as follows: During terminal operation, the instantaneous output power of the wave energy converter, the instantaneous output power of the photovoltaic array, the state of charge of the battery, and marine meteorological parameters (including wind speed, wave height, and irradiance) and ocean current parameters (including current velocity, current direction, and water temperature) are collected in real time. The acquisition frequency must reach the sub-second level to capture high-frequency abrupt events in wave energy output. After acquisition, the data is organized into a spatiotemporal sequence streaming feature tensor according to three dimensions: batch size, time step, and feature dimension. The batch size ranges from 1 to 16, the time step ranges from 10 to 60, and the feature dimension ranges from 8 to 32. The specific values ​​are determined based on the comprehensive analysis results of the model convergence speed and inference delay in multiple deployment experiments in the target sea area. The constructed spatiotemporal sequence streaming feature tensor is then input into the multidimensional feature adaptive sparse activation hybrid expert energy regulation model as the input data basis for subsequent prediction and control.

[0051] The specific implementation of step S02 is as follows: The multidimensional feature adaptive sparse activation hybrid expert energy regulation model first receives the spatiotemporal sequence streaming feature tensor through the input layer. Then, it performs orthogonal projection transformation on the multi-source environmental features using a spatial orthogonal decoupling operator, separating meteorological and ocean current features into independent sub-feature spaces with zero inner product, forming an orthogonalized feature stream and eliminating mutual interference between the two types of features. The orthogonalized feature stream enters the sparse activation gated routing layer, which is driven by a Markov chain state prediction gated routing weight adaptive update algorithm. This algorithm statistically analyzes the historical power generation and energy consumption states of the terminal, constructs a 2-5 order state transition probability matrix, calculates the prior probability of the next period's extreme energy shortage state at each streaming data input, and directly uses this prior probability as the independent variable input to the sparse activation gate, instantaneously adjusting the activation threshold of each expert network. When the predicted energy inflow probability is low, the gate function forcibly increases the sparsity of the routing probability vector through a step nonlinear transformation, keeping only two low-power expert networks in an active state, while the remaining expert networks enter a dormant state, and the GPU memory and system memory allocation are simultaneously compressed. Within each expert network, a long-span residual dense jump matrix aligns the dimensions of deep and shallow tensors and enables direct summation, ensuring high-speed gradient propagation in sparse activation states. The dynamic feedforward network routing, based on the nonlinear complexity of the input data, executes conditional branch jumps via gating units, directly skipping useless forward propagation and further reducing computational load. During model training, a multi-task loss function is used to jointly optimize power smoothing accuracy and charge / discharge control accuracy, using the battery state of charge as an auxiliary supervisory signal to train the sparse activation threshold of the gating routing layer. The output layer fuses the outputs of each activated expert network to generate power fluctuation smoothing commands and charge / discharge control parameter sequences.

[0052] The specific implementation of step S03 is as follows: Taking the power fluctuation smoothing command output in step S02 as the target constraint, and combining the transient nonlinear energy buffer mathematical model to quantitatively describe the high-frequency nonlinear abrupt change characteristics of wave energy and photovoltaic input power, a multi-level decoupled spatiotemporal attention operator is performed on the input power sequence. The multi-level decoupled spatiotemporal attention operator decouples the time dimension attention and the spatial dimension attention in layers: the time dimension attention mechanism assigns differentiated weights to the sub-second dynamic changes of the power sequence to capture the temporal pattern of power abrupt changes; the spatial dimension attention mechanism distinguishes between the power components from wave energy and photovoltaic sources, and applies differentiated smoothing strengths to the power fluctuations from different sources. After the attention calculations of the two dimensions are completed, the outputs are fused to complete the adaptive smoothing of the input power fluctuations within the sub-second time window, and finally output the smoothed charge and discharge current command to avoid the transient impact of high-frequency abrupt current on the electrochemical characteristics of the battery.

[0053] The specific implementation of step S04 is as follows: using the current rate of change of charging and discharging current... Battery state of charge and ambient temperature As input, based on the electrochemical damage risk assessment function Calculate the electrochemical damage risk value ,in It is a nonlinear mapping function of the charged state. For temperature correction functions, the coefficients are... , , The initial suggested range was 0.1–2.0, determined through multiple charge-discharge cycle experiments and electrochemical impedance spectroscopy analysis. The charge / discharge rate parameter is adaptively adjusted within the specified range: when Below the low risk threshold When the recommended range is 0.2–0.5, adjust the charge / discharge rate parameter to the high-rate range to accelerate the charge / discharge response; when In With high risk threshold When the recommended range is between 0.6 and 0.9, the charge / discharge rate parameter should be maintained in the medium rate range; when Not less than At this time, the charge / discharge rate parameters are forcibly compressed to a low rate range to suppress electrochemical damage at the microscopic physical level and protect the health of the battery.

[0054] The specific implementation of step S05 is as follows: The entire energy allocation optimization process is split into two heterogeneous independent channels—an energy prediction channel and a control instruction generation channel—that run in parallel. The energy prediction channel is responsible for the inference calculation of the multi-dimensional feature adaptive sparse activation hybrid expert energy regulation model, allocating more computing cores and memory resources to ensure prediction accuracy. The control instruction generation channel is responsible for the real-time generation and issuance of charging and discharging control parameters, allocating lightweight computing resources to ensure real-time instruction response. Data synchronization between the two channels is achieved through a lightweight asynchronous lock mechanism: after completing inference, the energy prediction channel updates the shared data area with a non-blocking atomic write operation, and the control instruction generation channel obtains the latest prediction result with a non-blocking atomic read operation. The two channels do not need to wait for each other, eliminating the thread blocking and deadlock risks introduced by traditional mutex locks, and effectively compressing the overall power consumption of the core processor under high load.

[0055] The specific implementation of step S06 is as follows: continuously monitor the battery state of charge and the predicted input power, and dynamically switch the energy distribution mode based on the combination of the two. When the state of charge is lower than the safety threshold (recommended range 15% to 30%, the specific value is determined by the actual energy consumption measurement of the terminal in the target sea area and the analysis of the battery discharge curve combined with the voltage drop inflection point in multiple discharge experiments), the system forcibly shuts down the preset non-core monitoring modules, including the non-real-time video acquisition module, the low-priority data reporting module and the non-critical environmental auxiliary sensor module, and prioritizes the allocation of the released electrical energy to the core sensing link (including hydrological sensors, meteorological sensors and communication modules) to ensure the continuous acquisition and uploading of monitoring data and avoid the terminal from failing completely due to energy depletion.

[0056] It should be noted that the key technologies of this invention include: a Markov chain-driven sparse activation gating routing mechanism, which embeds historical prior information of energy states into neural network gating, enabling the computational load of the model to adaptively shrink with the energy state, and completing the compression of the computational scale before the energy drop is triggered, thus avoiding the problem of energy depletion caused by continuous full inference under extreme conditions in fixed-structure models; a multi-level decoupled spatiotemporal attention and electrochemical damage risk assessment collaborative mechanism, which unifies power smoothing and charge / discharge rate protection under the same control framework, suppressing the impact of power fluctuations on the microscopic physical characteristics of the battery within sub-seconds; and an asymmetric parallel pipeline decoupling and lightweight asynchronous locking mechanism, which eliminates thread synchronization waiting through heterogeneous flow channel separation and non-blocking atomic operations, further reducing processor power consumption under high load conditions. The three key technologies mentioned above work together to form a complete closed loop from state prediction, computational contraction, power smoothing, damage protection to pipeline optimization. This enables the energy allocation optimization scheme to have real-time performance, low power consumption, and battery safety under sudden energy drop conditions. Compared with fixed rule strategies and static neural network schemes, it solves the fundamental defect of lack of dynamic coupling between computational resource consumption and energy state at the principle level.

[0057] It should be noted that during the long-term operation of marine monitoring terminals, the output power of wave energy converters and photovoltaic arrays will exhibit long-period non-stationary fluctuations due to seasonal changes in sea conditions. After repeated deep charge and discharge cycles, the internal resistance characteristics of the battery undergo non-linear drift, causing the fixed-parameter charge / discharge rate control strategy to be unable to adapt to the actual electrochemical characteristics of the battery after aging. This can lead to sudden voltage drops or even over-discharge damage under conditions of sudden temperature drops or deep discharge. The reason for the above technical problem is that the electrochemical characteristics of the battery are highly non-linear with changes in aging, temperature, and state of charge. Fixed charge / discharge rate parameters are based on the initial factory calibration and cannot detect the characteristic drift caused by increased internal resistance and capacity decay after aging. When the ambient temperature is low or the state of charge is in the extreme range at both ends, the rate of change of charge / discharge current at the fixed rate exceeds the current tolerance range of the battery, causing intensified local polarization and accelerating irreversible capacity decay. The usual solution to the above technical problem is to use a fixed segmented charge / discharge curve or a preset temperature compensation coefficient to statically correct the rate parameters. However, the correction coefficients of the fixed segmentation strategy are determined based on statistical averaging characteristics, which cannot track the individual aging state of a single battery cell in real time. Furthermore, the temperature compensation coefficient is usually a linear approximation, and the error is significantly amplified in extreme low temperatures or nonlinear intervals at both ends of the state of charge. This results in instances where the charge / discharge rate is still too high during actual charging and discharging, failing to fundamentally prevent the accumulation of electrochemical damage. This invention effectively solves this technical problem with its electrochemical damage risk assessment function. The charging / discharging current change rate, the nonlinear mapping value of the state of charge, and the temperature correction value are fused in real time. and All are nonlinear functions, capable of accurately describing the nonlinear variation of battery characteristics in the extreme ranges at both ends of the state of charge and under low-temperature conditions. The coefficients... , , Electrochemical impedance spectroscopy analysis and regression testing confirmed its ability to fit the individualized characteristics of single-cell batteries. When Exceeding the high-risk threshold At that time, the charge and discharge rate parameters are forcibly compressed to the low rate range, thereby judging and avoiding high-risk charge and discharge behaviors in real time in each control cycle. This allows the battery to maintain operation within the safe rate range under combined adverse conditions such as aging, low temperature or deep discharge, avoiding the electrochemical damage accumulation problem caused by the fixed rate strategy under the above conditions.

[0058] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned technical problems lies in its embedding of the temporal evolution information of energy state into a neural network gating mechanism, thereby dynamically coupling the allocation of computing resources with the system's energy state. Specifically, the Markov chain state prediction gating routing weight adaptive update algorithm constructs a multi-order state transition probability matrix, statistically analyzes the transition patterns of historical power generation and energy consumption states at the terminal, and can calculate the prior probability of the next period's extreme energy scarcity state at the instant of streaming data input. This prior probability is then directly used as the independent variable to input sparse activation gating, instantly adjusting the activation thresholds of each expert network. The logical rationale for this mechanism lies in the fact that Markov chains naturally describe the conditional dependencies of state sequences, and the transition of energy states has a clear temporal dependency. Therefore, driving the gating sparsity with a multi-order transition probability matrix physically corresponds to the prior knowledge of the energy system's state evolution, making the gating judgment forward-looking rather than solely dependent on instantaneous input. When the predicted energy inflow probability is low, the gating function forcibly increases the sparsity of the routing probability vector through a step nonlinear transformation, activating only two sets of low-power expert networks while the remaining expert networks enter a dormant state. Memory allocation is simultaneously compressed, thus shrinking the computational scale before a sudden energy drop triggers, avoiding the problem of energy depletion exacerbated by continuous full inference under extreme conditions in fixed-structure models. A long-span residual dense jump matrix ensures high-speed transmission of feature gradients between shallow and deep layers in the sparse activation state, and dynamic feedforward network routing achieves fine-grained pruning at the neuron level. Both together ensure that the prediction accuracy of the sparse activation model does not significantly degrade due to the compression of the activation scale. A transient nonlinear energy buffering mathematical model, in conjunction with a multi-level decoupled spatiotemporal attention operator, adaptively smooths wave energy and photovoltaic power fluctuations within a sub-second time window, avoiding the transient impact of high-frequency sudden current on the battery's electrochemical characteristics. The electrochemical damage risk assessment function further dynamically adjusts the charge / discharge rate based on the rate of change of charge / discharge current, state of charge, and temperature, protecting battery health from a microscopic physical level. The asymmetric parallel pipeline decoupling mechanism designs the power prediction channel and the control command generation channel as heterogeneous independent channels. A lightweight asynchronous locking mechanism replaces traditional mutexes with non-blocking atomic operations, eliminating inter-thread blocking and waiting, further reducing the power consumption of the core processor under high load. When the state of charge falls below a safe threshold, the system forcibly shuts down non-core monitoring modules, prioritizing power supply to the core sensor link, forming a final safety net for energy allocation. Logically, these mechanisms form a complete closed loop from state prediction, computational contraction, power smoothing, damage protection to pipeline optimization. Each step is designed for specific failure modes under sudden energy drops, thus ensuring the real-time performance and system stability of energy allocation optimization from a fundamental perspective.

[0059] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0060] The specific implementation method of step S01 is as follows.

[0061] Real-time output power of wave energy converter and photovoltaic array, battery state of charge, and marine environmental parameters are collected to construct a spatiotemporal sequence streaming characteristic tensor. ,in This represents the batch size, with a value ranging from 1 to 16. This represents the time step, with a value ranging from 10 to 60. The feature dimension, with a value ranging from 8 to 32, was determined through multiple deployments in different marine environments, and after comprehensive analysis of the model's convergence speed and inference latency. Tensor The The first batch, the first The time step, the first The elements of each feature channel are denoted as Including the instantaneous output power of wave energy Instantaneous output power of photovoltaic Battery state of charge Multidimensional information including meteorological parameters and ocean current parameters will then be... Input to a multidimensional feature adaptive sparse activation hybrid expert energy regulation model.

[0062] The specific implementation method of step S02 is as follows.

[0063] The multidimensional feature adaptive sparse activation hybrid expert energy regulation model first projects the input features into independent subspaces using a spatial orthogonal decoupling operator. Let the meteorological feature projection matrix be... The ocean current feature projection matrix is The orthogonal constraint conditions are:

[0064] ;

[0065] in For the meteorological subspace dimension, For the ocean current subspace dimension, superscript Indicates transpose. Zero matrix. Meteorological orthogonalized characteristic flow. Orthogonalization of ocean currents The two are concatenated to obtain the orthogonalized feature flow:

[0066] ;

[0067] Construction of Markov Chain State Prediction Gated Routing Weight Adaptive Update Algorithm Order state transition probability matrix , its first element Indicates the system's energy state Transition to energy state The probability, The total number of energy states. The order is the Markov chain order, empirically ranging from 2 to 5. The specific order is determined through comparative experiments comparing the accuracy of state prediction under different sea states. In the... The prior probability of a state of extreme energy scarcity at a given moment. The calculation formula is as follows:

[0068] ;

[0069] in For the first The one-hot encoded column vector of the energy state of the system at any given time. This is the one-hot encoded column vector corresponding to a state of extreme energy scarcity. The state transition probability matrix Power of 1 This is a dimensionless probability value, with a range of values ​​of [value missing]. Sparse activation gating routing layer based on Dynamic adjustment of the first Activation threshold of an expert network The formula is as follows:

[0070] ;

[0071] in For the first The initial activation threshold of an expert network. The prior probability sensitivity coefficient has an empirical value of 0.5 to 2.0. Both are dimensionless scalars. This is a dimensionless activation threshold. When the battery's state of charge... Below the safety threshold At that time, the gate sparsity is forcibly increased through a step nonlinear transformation, thereby activating the probability vector. ( (the total number of expert networks) The components are calculated as follows:

[0072] ;

[0073] in for Activation function, maps input to interval, For the first The gating weight vector of an expert network, It is a collection of low-power expert networks. This is an indicator function that takes the value 1 when the condition within the parentheses is true, and 0 otherwise. The dimensionless activation probability. and All values ​​are dimensionless percentages of the state of charge. Each expert network internally constructs a long-span residual dense jump matrix. , For the internal feature dimension of the expert network, achieve the first Layer and First Dimension alignment and direct summation of layer outputs:

[0074] ;

[0075] in For the first The layer outputs a feature vector. For the first Layer weight matrix, For the first The layer outputs a feature vector. For the first The layer outputs a feature vector. For multi-layer span, empirical values ​​range from 2 to 6 layers. The dynamic feedforward network routing executes conditional branching based on the input's non-linear complexity, dynamically pruning inactive neuron blocks. The pruning ratio empirically ranges from 20% to 60%, with the specific value determined by the lowest loss point on the validation set during multiple rounds of comparative experiments. Memory allocation... Based on dynamic scaling according to the state of charge, the formula is as follows:

[0076] ;

[0077] in For the first The video memory requirements for a fully activated expert network. For the first The activation probability of an expert network. and All quantities are in bytes. The output layer integrates the outputs of each activated expert network to generate power fluctuation smoothing instructions and charge / discharge control parameter sequences.

[0078] The specific implementation method of step S03 is as follows.

[0079] Based on the transient nonlinear electrical energy buffer mathematical model, the wave energy input power... With photovoltaic input power Perform multi-level decoupled spatiotemporal attention operator processing. Temporal dimension attention operator. Used to capture sub-second dynamic changes in power fluctuations, its first The element calculations are as follows:

[0080] ;

[0081] in For the first The query vector at time step, For the first The key vector at time step, For time-based attention features, Scaling factor These are dimensionless attention weights. Spatial dimension attention operators. The system applies to two power sources: wave energy and photovoltaic energy, and distinguishes the contribution weights of the two power sources. It is also a dimensionless weight matrix. After decoupling and merging the two operations, the output is a sub-second smoothed charge / discharge current command. The unit is This enables effective suppression of high-frequency nonlinear power mutations.

[0082] The specific implementation method of step S04 is as follows.

[0083] Electrochemical damage risk assessment function based on the rate of change of charge / discharge current Battery state of charge and ambient temperature As input, calculate the electrochemical damage risk value. The formula is as follows:

[0084] ;

[0085] in The rate of change of charging and discharging current, in units of This is a nonlinear mapping function of the charged state, reflecting the nonlinear effect of the charged state on electrochemical damage, and is a dimensionless quantity. This is a temperature correction function, reflecting the effect of ambient temperature on the electrochemical reaction rate; it is a dimensionless quantity. , , Here is the dimensionless coefficient, where The unit is , and The unit is dimensionless, determined through regression analysis of multiple sets of charge-discharge cycle experiments and electrochemical impedance spectroscopy, with an initial range of 0.1–2.0. This is a dimensionless risk index. When... At that time, the charge / discharge rate parameter is adjusted to the high rate range; when At that time, maintain a medium multiplier; when At that time, it is forcibly compressed to a low magnification range, in which The empirical range is 0.2 to 0.5. The empirical range is 0.6 to 0.9. Both are determined through multiple sets of charge-discharge cycle aging experiments at different temperatures and under different charging states, with the risk value corresponding to the inflection point of battery capacity decay as the benchmark, combined with statistical analysis.

[0086] The specific implementation method of step S05 is as follows.

[0087] An asymmetric parallel pipeline decoupling mechanism is implemented, which separates the power prediction pipeline and the control instruction generation pipeline into heterogeneous multi-channel independent operation. The power prediction pipeline is allocated more computing cores and memory resources, while the control instruction generation pipeline is allocated lightweight computing resources. The two pipelines synchronize data through a lightweight asynchronous lock mechanism, replacing the traditional mutex lock with non-blocking atomic operations, eliminating the risk of thread blocking and deadlock, and reducing the power consumption of the core processor under high load.

[0088] The specific implementation method of step S06 is as follows.

[0089] According to the battery state of charge Based on the predicted input power, the energy distribution conditions are dynamically switched when... Below the safety threshold In certain situations, non-core monitoring modules are forcibly shut down to prioritize power supply to core sensor links. The empirical range is 15% to 30%, determined by measuring the actual energy consumption of the monitoring terminal in the target sea area and analyzing the battery discharge curve, combined with the inflection point of the sudden drop in battery voltage in multiple discharge experiments.

[0090] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0091] This invention's method is fully technically verified using a marine monitoring terminal powered by a hybrid wave energy and photovoltaic system in the open ocean. The terminal is equipped with a wave energy converter (rated power 200 W) and a photovoltaic array (rated power 150 W), with a capacity of 5... The lithium iron phosphate battery pack's core sensing links include hydrological sensors, meteorological sensors, and a satellite communication module. Pre-set non-core monitoring modules include a non-real-time underwater video acquisition module, a low-priority data reporting module, and non-critical environmental auxiliary sensor modules (including a water color sensor and an underwater optical sensor). The test scenario selected was a 72-hour continuous terminal operation under typical severe sea conditions, during which two strong wave events with wave heights exceeding 4 meters and one low-irradiance event caused by prolonged overcast skies resulting in near-zero photovoltaic output.

[0092] After the terminal is powered on, step S01 begins to continuously collect data on the output power of the wave energy converter, the output power of the photovoltaic array, the state of charge of the battery, wind speed, wave height, irradiance, and ocean current speed and direction at a sampling period of 10 ms. A spatiotemporal sequence streaming feature tensor with a batch size of 1, a time step of 30, and a feature dimension of 16 is constructed and input frame by frame into a multidimensional feature adaptive sparse activation hybrid expert energy control model. At the initial moment of the test, the battery state of charge was 82%, the wave energy output power was 143 W, the photovoltaic array output power was 97 W, and the total load power consumption of the terminal was 180 W, indicating that the system was in an energy surplus state.

[0093] In step S02, the spatial orthogonal decoupling operator separates meteorological features (wind speed, wave height, irradiance) and ocean current features (current velocity, direction) into independent sub-feature spaces with an inner product of zero, eliminating the mutual interference between the two types of features. The Markov chain state prediction gated routing weight adaptive update algorithm, based on the first three order state transition probability matrices, calculates the prior probability of entering an extremely energy-deficient state in the next period as approximately 3% under an energy surplus state. The gated routing layer activates all six expert networks, and the memory allocation is at full capacity. The model outputs power fluctuation smoothing instructions and charging / discharging control parameters, at which point the inference computation is at its highest level. At the 18th hour, continuous overcast skies caused photovoltaic output to drop to 8 W, and wave energy, affected by the wind field, also dropped to 61 W, resulting in a power deficit at the terminal. The battery began to discharge, and the state of charge decreased from 64% to 29% within 4 hours. The Markov chain state prediction algorithm is based on real-time calculation of the state transition probability matrix. When the state of charge drops to 40%, the prior probability of entering a state of extreme energy depletion has risen to 71%. The gating function improves the sparsity of the routing probability vector through a step nonlinear transformation, reducing the number of activated expert networks from 6 to 2 low-power expert networks, while the remaining 4 enter sleep mode. Approximately 67% of the GPU memory is released, resulting in a significant reduction in inference power consumption. Figure 2 As shown, with the continuous decrease in state of charge, the number of activated expert networks shrinks in a stepwise manner, and the power consumption curve of model inference also tends to flatten out, reflecting the adaptive compression mechanism of sparse activation gated routing for computing resources.

[0094] Step S03 continues in the above process, performing multi-level decoupled spatiotemporal attention operator processing on wave energy and photovoltaic input power. During the 18-hour strong wave event, the wave energy output power abruptly changed from 87 W to 214 W within 10 seconds. The temporal dimension attention mechanism assigns high weight to the moment of abrupt change, while the spatial dimension attention mechanism applies a significant smoothing force to the wave energy component. In the charging and discharging current command output after the fusion of the two-dimensional attention, the rate of change of current significantly narrows from the peak corresponding to the original power fluctuation. The battery charging current remains on a smooth curve within a sub-second time window, avoiding the transient impact of large current abrupt changes on the battery. Figure 3As shown, the comparison of the charge and discharge current curves before and after smoothing indicates that after processing by the multi-stage decoupled spatiotemporal attention operator, the peak amplitude of the charge and discharge current fluctuations is significantly reduced, and the current change rate tends to be smoother, reflecting the suppression effect of the spatiotemporal decoupling smoothing mechanism on the electrochemical shock of the battery.

[0095] Step S04: In each control cycle, based on the real-time data collected... , With ambient temperature Calculate the electrochemical damage risk value The ambient temperature during the test was approximately 12℃, and the coefficients were determined through prior experimental regression. , , Low risk threshold High risk threshold Under operating conditions where the battery's state of charge drops to 29% and the ambient temperature is low, The temperature correction value increased. The nonlinear mapping value also increases in the low charge state range, as calculated. ,Exceed The charge / discharge rate parameters are forcibly compressed to a low rate range, and the upper limit of the discharge current is reduced, thereby avoiding damage from over-speed discharge under combined low-temperature and low-charge conditions. Within each control cycle... The changes in values ​​and the corresponding adjustments to charge / discharge rates are shown in Table 1.

[0096] Table 1 Electrochemical Damage Risk Value and Charge / Discharge Rate Adjustment Record Table

[0097]

[0098] As can be seen from Table 1, with the combined effects of the continuous decrease in state of charge and low temperature, The value shows a monotonically increasing trend, with the charge / discharge rate range transitioning sequentially from high rate to medium rate and then to low rate. The rate adjustment boundary is... , The segmented structure corresponds one-to-one, reflecting the nonlinear segmented protection mechanism of the electrochemical damage risk assessment function for charge / discharge rates. After 36 hours, as illumination conditions gradually recovered, the battery's state of charge rebounded. The value then decreased, and the multiplier range returned to the medium multiplier, indicating that the mechanism has the ability to adapt to the energy state in both directions.

[0099] Step S05 continuously executes the asymmetric parallel pipeline decoupling mechanism throughout the entire test. The power prediction channel independently undertakes the inference calculation of the multi-dimensional feature adaptive sparse activation hybrid expert energy control model, while the control command generation channel uses non-blocking atomic read operations to obtain the latest prediction results in real time and generate charge and discharge control commands. There is no blocking or waiting between the two channels. Under the low-power state of activating expert networks to 2 groups, the inference latency of the power prediction channel is reduced from approximately 23 ms when all 6 groups are fully activated to approximately 9 ms, and the processor's overall load rate is reduced from a peak of approximately 78% to approximately 31%, effectively compressing the power consumption of the core processor under high load.

[0100] Step S06 triggers a safety threshold judgment (the safety threshold is set at 30%) when the state of charge drops to 29%. This forcibly shuts down the non-real-time underwater video acquisition module (power consumption approximately 35 W), the low-priority data reporting module (power consumption approximately 12 W), and the non-critical environmental auxiliary sensor module (power consumption approximately 8 W), releasing approximately 55 W of power. This reduces the total power consumption of the terminal from 180 W to approximately 125 W, while the core sensing links (hydrological sensor, meteorological sensor, and communication module, with a combined power consumption of approximately 83 W) continue to be powered. The terminal did not experience a complete failure under low-irradiance conditions for approximately 6 hours. The trend of the battery's state of charge during the 72-hour test cycle is as follows: Figure 4 As shown, the state of charge curve shows a downward trend in the interval from 18 h to 24 h. Under the dual protection of non-core module shutdown and rate compression, the slope of the curve tends to flatten after 24 h, and does not reach the over-discharge protection lower limit (set to 10%), which reflects the synergistic effect of steps S06 and S04 on suppressing deep discharge of the battery.

[0101] From a technical perspective, as the state of charge decreases, the prior probability calculated by the Markov chain state prediction algorithm increases, driving the sparse activation gating to shrink the number of activated expert networks, thus reducing the power consumption of model inference; the electrochemical damage risk assessment function, under the combined conditions of low state of charge and low temperature, ... When the charge level rises, it triggers rate compression protection; when the state of charge falls below the safety threshold, non-core modules are forcibly shut down to further release power consumption. These three mechanisms work synergistically from the three dimensions of computing resources, charge / discharge rate, and load management, enabling the system to maintain power supply to the core sensing link even under conditions of continuous insufficient energy inflow. This solves the technical defects of the background technology, where fixed computing structures and fixed strategies cannot adaptively adjust computing resources and energy allocation when energy drops sharply.

[0102] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing energy distribution in a new energy-driven marine monitoring terminal, characterized in that, Includes the following steps: Real-time output power of wave energy converter and photovoltaic array, state of charge of battery and marine environmental parameters are collected to construct spatiotemporal sequence streaming feature tensor, which is then input into multidimensional feature adaptive sparse activation hybrid expert energy regulation model; The multidimensional feature adaptive sparse activation hybrid expert energy regulation model projects meteorological and ocean current features to independent subspaces through a spatial orthogonal decoupling operator. The prior probability of the next period of extreme energy scarcity is calculated by the Markov chain state prediction gating routing weight adaptive update algorithm. The sparse activation gating routing weight is dynamically adjusted, and the output power fluctuation smoothing command and charging and discharging control parameters are output. Based on the transient nonlinear energy buffer mathematical model, a multi-level decoupled spatiotemporal attention operator is performed on wave energy and photovoltaic input power to adaptively smooth the input power fluctuation within a sub-second time window and output the smoothed charging and discharging current command. Calculate the electrochemical damage risk assessment function value, and adaptively adjust the charge and discharge rate parameters according to the interval to which the function value belongs, so as to protect the microscopic physical characteristics of the battery. An asymmetric parallel pipeline decoupling mechanism is implemented to separate the power prediction channel and the control instruction generation channel into heterogeneous multi-channel independent operation. A lightweight asynchronous lock mechanism is used to eliminate synchronization wait between threads and reduce the power consumption of the core processor under high load. Based on the battery state of charge and the predicted input power, the energy distribution mode is dynamically switched. When the state of charge is lower than the safety threshold, the preset non-core monitoring modules are forcibly shut down to prioritize the power supply of the core sensing link.

2. The energy allocation optimization method for a new energy-driven marine monitoring terminal according to claim 1, characterized in that, The dimension of the spatiotemporal sequence streaming feature tensor is defined as a three-dimensional tensor composed of batch size, time step, and feature dimension. The specific values ​​of batch size, time step, and feature dimension were determined by deploying experiments multiple times in different marine environments and analyzing the model convergence speed and inference latency.

3. The energy allocation optimization method for a new energy-driven marine monitoring terminal according to claim 2, characterized in that, The spatial orthogonal decoupling operator separates multi-source environmental features into independent sub-feature spaces that do not interfere with each other through orthogonal projection transformation. The inner product between each sub-space is zero, thereby eliminating the mutual interference between meteorological features and ocean current features.

4. The energy allocation optimization method for a new energy-driven marine monitoring terminal according to claim 3, characterized in that, The multidimensional feature adaptive sparse activation hybrid expert energy regulation model includes an input layer, a sparse activation gated routing layer, and an output layer. When the battery state of charge is lower than the safety threshold, the gate function forcibly increases the sparsity of the routing probability vector through a step nonlinear transformation, activating only two sets of low-power expert networks, while the remaining expert networks enter a dormant state.

5. The energy allocation optimization method for a new energy-driven marine monitoring terminal according to claim 4, characterized in that, Each expert network constructs a long-span residual dense jump matrix to achieve dimensional alignment and direct summation of deep tensors and shallow tensors; dynamic feedforward network routing is introduced in the feedforward network, and the gated unit executes conditional branch jumps according to the nonlinear complexity of the input environment data, dynamically pruning inactive neuron blocks.

6. The energy allocation optimization method for a new energy-driven marine monitoring terminal according to claim 5, characterized in that, The Markov chain state prediction gating routing weight adaptive update algorithm constructs a multi-order state transition probability matrix by statistically analyzing the historical power generation and energy consumption states of the terminal. The calculated prior probabilities are directly used as independent variables to input sparse activation gating, and the activation thresholds of each expert network are adjusted instantaneously. The order of the multi-order state transition probability matrix is ​​determined through comparative experiments on the state prediction accuracy under different sea states.

7. The energy allocation optimization method for a new energy-driven marine monitoring terminal according to claim 6, characterized in that, The training dataset for the multidimensional feature adaptive sparse activation hybrid expert energy regulation model includes output power time-series data, battery state of charge data, marine meteorological data, and ocean current data continuously collected in the target sea area. The collection duration is not less than the collection duration threshold, and power mutation events are recorded at a sub-second sampling frequency.

8. The energy allocation optimization method for a new energy-driven marine monitoring terminal according to claim 7, characterized in that, The training of the multidimensional feature adaptive sparse activation hybrid expert energy regulation model adopts a multi-task loss function to jointly optimize the power smoothing accuracy and charge / discharge control accuracy, and uses the battery state of charge as an auxiliary supervision signal to train the sparse activation threshold of the gated routing layer.

9. The energy allocation optimization method for a new energy-driven marine monitoring terminal according to claim 8, characterized in that, The multi-level decoupled spatiotemporal attention operator decouples the time-dimensional attention and spatial-dimensional attention in layers and applies them sequentially to the input power sequence. The time-dimensional attention captures the sub-second dynamic changes of power fluctuations, while the spatial-dimensional attention distinguishes between the power components from wave energy and photovoltaic sources. After the two are decoupled and processed, they are fused to output a smooth result.

10. The energy allocation optimization method for a new energy-driven marine monitoring terminal according to claim 9, characterized in that, In the asymmetric parallel pipeline decoupling mechanism, the power prediction pipeline is allocated more computing cores and memory resources, while the control instruction generation pipeline is allocated lightweight computing resources. Data synchronization between the two pipelines is achieved through a lightweight asynchronous lock mechanism.

Images