A carbon emission prediction driven hybrid ship power distribution optimization system

Abstract

The application discloses a carbon emission prediction driven hybrid ship power distribution optimization system, and particularly relates to the technical field of hybrid ship energy management and adaptive control, and comprises a ship-mounted edge computing node, engine power code and propulsion power code are generated according to shaft end torque and same moment speed, fuel mass flow is read according to moment index and state of charge code is generated by integrating direct current side current, and power state sequence is output; power code is tried and calculated according to engine control resolution increment, engine power is deducted by using a hybrid power simulator, propulsion power is supplemented by a storage battery, and equipment registration interval is verified along a power chain; engine load reduction and recovery response are formed by replacing engine power with the storage battery during the period when the input power of the propulsion motor remains unchanged, a target supercharged intake state is identified according to a measured recovery sequence, and power distribution instructions are generated by using a corrected hybrid power simulator and a reliability domain strategy optimization algorithm.

Description

Technical Field

[0001] This invention relates to the field of energy management and adaptive control technology for hybrid-powered ships, and more specifically, to a carbon emission prediction-driven power allocation optimization system for hybrid-powered ships. Background Technology

[0002] Existing hybrid ships mostly rely on engine speed, output power, coolant temperature, battery state of charge and fuel consumption records to establish carbon emission prediction models, and use the prediction results to adjust the power sharing between the engine and the battery. When sailing at a constant speed in windy and wavey areas, the propeller will be repeatedly unloaded and loaded by the waves. The propulsion system still needs to maintain shaft power and cannot stop to identify the engine status. After the engine experiences a short period of unloading, its speed, output power and coolant temperature can recover to the values ​​before unloading, but the turbocharger speed and intake manifold pressure have not recovered synchronously. Therefore, when the engine power is increased again, there will be a difference between the fuel flow increase and the power recovery time. Existing prediction models merge operational records based on externally measurable parameters, which can easily treat records with different internal states as the same state, causing carbon emission prediction errors and causing repeated adjustments to the power distribution between the engine and the battery. The technical problem this application aims to solve is: to distinguish the operating states of engines with the same external measurable parameters but different boost and intake states while maintaining ship propulsion power, thereby improving the accuracy of carbon emission prediction and power distribution. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a carbon emission prediction-driven power allocation optimization system for hybrid ships. This system utilizes a battery to take over engine power while the propulsion motor input power remains constant, thereby generating an engine load reduction and recovery response. Based on the measured recovery sequence, the system identifies the target boosted air intake state and then uses a modified hybrid simulator and a trust region strategy optimization algorithm to generate power allocation commands, thus solving the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a carbon emission prediction-driven power distribution optimization system for hybrid-powered ships, comprising: The edge state acquisition module is set on the shipborne edge computing node. It generates engine power code and propulsion power code based on shaft end torque and synchronous speed, reads fuel mass flow rate by time index and integrates DC side current to generate charge state code, and outputs power state sequence. Replace the solver module, increment the power code trial calculation according to the engine control resolution, use the hybrid power simulator to deduct the engine power and supplement the propulsion power by the battery, verify the equipment registration interval along the power chain, and read the previous power code trial calculation when the boundary is crossed for the first time to generate a replacement instruction; The centralized execution module reduces engine power and rewrites converter current according to bus missing power by the shipborne edge computing node, keeps propulsion motor input power unchanged, restores engine power and withdraws converter current, and outputs measured recovery sequence. The state identification module performs an ancestor sampling particle Gibbs algorithm on the measured recovery sequence, generates boosted intake particles according to the state transition probability, uses a hybrid power simulator to deduce fuel flow, generates particle weights based on fuel flow difference and power recovery difference, extracts ancestors according to particle weights and resamples, and outputs the target boosted intake state with deviation count. The power allocation optimization module corrects the hybrid power simulator based on the target boosted intake state. It generates candidate engine power through reinforcement learning and the battery takes over the remaining power. It simulates the carbon emission return after the state of charge is restored. It generates the trust region radius by dividing one by the sum of the particle number and the deviation count. It executes the trust region policy optimization algorithm, obtains the update direction through the conjugate gradient, and halves the update step size successively until the average relative entropy does not exceed the trust region radius and the carbon emission return is not lower than the value before the update. Finally, it outputs the power allocation command.

[0005] In a preferred embodiment, the edge state acquisition module includes: Using the time when the replacement command is issued as the zero index, the engine shaft end torque value and engine speed value are read according to the sampling number and multiplied to generate the engine power code. The propulsion shaft end torque value and propulsion shaft speed value are read according to the same sampling number and multiplied to generate the propulsion power code. The engine power code with zero index is used as the power code before replacement. The first index bit after zero index that is different from the power code before replacement is used as the first change index. Candidate offset bits are generated bit by bit from zero index to first change index. The engine simulator deduces the fuel mass flow rate according to the engine power code. The measured fuel mass flow rate is shifted according to the candidate offset bits and the difference between the measured fuel mass flow rate and the deduce fuel mass flow rate is calculated bit by bit and converted into a flow difference count. The first bit of the candidate offset bits is taken in ascending order of flow difference count and ascending order to generate the fuel index offset bit. The engine simulator refers to the computing program that runs within the shipborne edge computing node. It reads the engine fuel mapping table according to the engine power code, engine speed value, the turbocharger speed code of the previous index, and the intake pressure code of the previous index, and generates the injection quality index by index. The simulated fuel mass flow rate refers to the mass flow rate value obtained by dividing the fuel injection mass generated by the engine simulator at each index position by the sampling period. The fuel mass flow rate is shifted according to the fuel index offset position. The engine power code is moved from the index position of the power code before replacement to the index position of the power code before replacement. The product of the DC side current and the sampling period is accumulated to generate the accumulated charge. The accumulated charge is used to correct the zero-position index charge status code and generate the charge status code. The engine power code, propulsion power code, fuel mass flow rate and charge status code are written to the same index position, and the power status sequence is output.

[0006] In a preferred embodiment, the replacement solver module includes: Based on the last index of the power state sequence, read the current engine power code, current charge state code and bus voltage, and read the steady-state lower limit power code, engine control resolution and power decrease rate registered by the engine controller. Subtract the steady-state lower limit power code from the current engine power code to obtain the load reduction margin, and generate the trial power code by dividing the load reduction margin by the engine control resolution. The bus voltage refers to the DC bus voltage sample value written by the last index of the power state sequence, which is used to convert the bus power difference into a compensation current code. The engine controller is an electronic control unit that receives the engine power code issued by the shipborne edge computing node and outputs control quantities to the engine fuel injection actuator. The steady-state lower limit power code refers to the lower limit of the power code registered by the engine controller that allows the engine to maintain steady-state combustion control; Engine control resolution refers to the power difference between two adjacent writable power codes in the engine controller's power command register; For each calculated power code, the calculated engine power code is generated by subtracting the calculated power code from the current engine power code. The power reduction rate is multiplied by the simulation cycle duration and quantized according to the engine control resolution to generate a single-cycle load reduction code. The hybrid simulator subtracts the single-cycle load reduction code from the current engine power code cycle by cycle, and generates a load reduction power trajectory when the calculated engine power code is reached. The hybrid power simulator refers to a computing program that runs within a shipboard edge computing node, which infers fuel mass flow rate and engine power response by indexing the engine power code, boosted intake particles, and replacement instructions. Read the engine to DC bus efficiency table according to the load reduction power trajectory, multiply the adjacent engine power codes by the corresponding efficiency values ​​and subtract them to generate the bus power difference, and divide the bus power difference by the bus voltage of the same period to generate the compensation current code. The DC bus efficiency table is a registration table that uses the engine power code as an index and outputs the efficiency value after the engine shaft power is transferred to the DC bus.

[0007] In a preferred embodiment, the replacement solver module further includes: If the calculated engine power code is lower than the steady-state lower limit power code, the compensation current code is higher than the inverter's registered discharge current upper limit, and the current charge status code, after deducting the compensation current code's integral charge and the ratio of the battery management system's registered rated capacity, is lower than the discharge termination charge status, write these values ​​into the engine, inverter, and battery over-limit positions, respectively. The battery management system refers to the battery control unit that collects battery current, voltage and capacity data and outputs state of charge and discharge limit data to the shipboard edge computing node. The state of charge at the end of discharge refers to the lower limit of the state of charge registered by the battery management system as the point at which discharge will cease. Read the engine over-limit position, inverter over-limit position, and battery over-limit position according to the sequence number. When the over-limit position is read for the first time, take the previous sequence number as the target sequence number. When the target sequence number is zero, generate a zero code replacement instruction. When the target sequence number is not zero, read the load reduction power trajectory and compensation current code of the target sequence number. Write the forward adjacent difference of the load reduction power trajectory as the engine load reduction code, write the reverse adjacent difference of the load reduction power trajectory as the engine recovery code, and write the compensation current code in reverse order as the withdrawal current code, and generate a replacement instruction.

[0008] In a preferred embodiment, the centralized execution module includes: Starting from the point of execution of the replacement instruction, the bus voltage and the propulsion motor input current are read and multiplied to generate the reference input power code. Then, the engine load reduction code and compensation current code are read in sequence according to the replacement instruction. The propulsion motor input current refers to the sampled value of the current flowing from the DC bus into the DC input terminal of the propulsion motor controller, which is used to multiply with the bus voltage to generate the propulsion motor input power. When load reduction is performed, the shipborne edge computing node writes the engine load reduction code and compensation current code on the same local clock edge. The next local clock generates the measured input power code by multiplying the bus voltage and the propulsion motor input current. The reference input power code is subtracted from the measured input power code and divided by the bus voltage to generate the current correction code. The current correction code is then added to the converter current code.

[0009] In a preferred embodiment, the centralized execution module further includes: When performing recovery, the engine recovery code and withdrawn current code are read in reverse order according to the replacement instruction. The engine recovery code is written on the same local clock edge and the withdrawn current code is deducted from the converter current code. The next local clock generates a current correction code based on the difference between the reference input power code and the measured input power code and adds it to the converter current code until the engine power code and converter current code return to their values ​​before replacement. During the load reduction and recovery phases, the engine power code is paired with the same code and recorded. The fuel recovery difference is generated by subtracting the fuel mass flow rate of the load reduction phase from the fuel mass flow rate of the recovery phase. The internal response difference is generated by subtracting the product of the current correction code and the bus voltage of the recovery phase from the product of the current correction code and the bus voltage of the load reduction phase. The measured recovery sequence is written according to the replacement instruction order.

[0010] In a preferred embodiment, the state identification module includes: Read the fuel recovery difference and internal response difference in the measured recovery sequence according to the replacement instruction order, divide the absolute value of the fuel recovery difference by the fuel flow detection resolution to generate the fuel difference count, divide the absolute value of the internal response difference by the engine control resolution to generate the power difference count, and write it into the sequential observation record; The fuel flow detection resolution refers to the mass flow difference between two adjacent output digital codes in the fuel mass flow sensor calibration table. The number of particles is generated by adding one to the number of sequential observation records. The turbocharger speed range registered by the engine controller is evenly divided into turbocharger speed particles according to the number of particles. The intake pressure particles are calculated by subtracting the cylinder intake mass flow rate from the compressor intake mass flow rate. The turbocharger speed particles and intake pressure particles with the same number are combined into turbocharged intake particles. The turbocharger speed range refers to the speed range between the lower limit of the turbocharger speed registered by the engine controller and the upper limit of the turbocharger speed. Compressor intake mass flow rate refers to the mass flow rate of air entering the intake manifold, which is read from the compressor flow meter by the hybrid power simulator using the turbocharger speed particle and the previous indexed intake pressure particle as indices. Cylinder intake mass flow rate refers to the mass flow rate of air entering the cylinder, which is read from the engine intake flow meter by the hybrid simulator using the engine power code, engine speed value, and the previous indexed intake pressure particle as indices. The process of calculating the intake pressure particle is as follows: subtract the cylinder intake mass flow rate from the compressor intake mass flow rate, multiply the difference by the sampling period, intake temperature and gas constant, divide by the intake manifold volume, and add it to the previous index intake pressure particle to generate the current index intake pressure particle.

[0011] In a preferred embodiment, the state identification module further includes: The hybrid simulator extrapolates the fuel difference and power difference of each boosted intake particle according to the replacement instruction. It converts the difference between the extrapolation result and the sequential observation record into a residual count. The negative value of the residual count is used as the natural constant exponent to generate the particle weight. The ancestor extraction probability is generated by the previous sequential particle weight, the state transition probability and the tail residual count. The ancestor sampling particle Gibbs resampling is performed according to the ancestor extraction probability. The process of deriving the fuel difference and power difference for each boosted intake particle is as follows: The hybrid simulator executes the replacement command with the same boosted intake particle, and records the fuel mass flow rate and actual engine power code when the engine power code first equals the same command code in the unloading phase and the recovery phase, respectively. The fuel difference is generated by subtracting the fuel mass flow rate in the unloading phase from the fuel mass flow rate in the recovery phase, and the power difference is generated by subtracting the actual engine power code in the unloading phase from the actual engine power code in the recovery phase. The state transition probability refers to the probability value generated by subtracting the predicted turbocharger speed from the current indexed turbocharger speed particle, dividing the absolute value by the particle interval to generate a transition difference count, and then generating the probability value with the natural constant as the base and the negative value of the transition difference count as the exponent. The predicted turbocharger speed is obtained by adding the difference between the turbine torque and the compressor torque to the previous indexed turbocharger speed particle, multiplying it by the sampling period, and then dividing it by the rotor moment of inertia. After each round of resampling, the target boosted air intake state is generated according to the particle weight. The difference between each boosted air intake particle and the target boosted air intake state is converted into a particle deviation count. Resampling stops when the XOR result of the target boosted air intake state and the ancestor index sequence relative to the previous round is zero. When the number of resampling rounds reaches the number of sequential observation records, the first result is taken according to the residual count and ascending order, and the target boosted air intake state with particle deviation count is output. The process of generating the target boosted intake state is as follows: each boosted intake particle is multiplied by its corresponding normalized particle weight and then summed one by one to obtain the target booster speed and target intake pressure. The target boosted intake state is composed of the target booster speed and target intake pressure.

[0012] In a preferred embodiment, the power allocation optimization module includes: Write the target boost intake state into the hybrid simulator. Generate engine candidate power bit by bit from the lower power limit registered by the engine controller to the upper power limit according to the engine control resolution. Subtract the engine candidate power from the propulsion demand power to generate the battery candidate power. Divide the battery candidate power by the bus voltage to generate the battery candidate current. Write the records where the battery candidate current does not exceed the converter's registered current limit as candidate power records. The propulsion power demand refers to the current period's propulsion power demand value generated by the shipborne edge computing node based on the speed control command, propulsion shaft speed, and propulsion shaft torque. For each candidate power record, the hybrid power simulator calculates the engine candidate power and battery candidate power according to the number of records in the measured recovery sequence. Then, it generates a recovery current based on the difference between the battery's state of charge and the starting point. The engine then bears the remaining propulsion power after the power corresponding to the recovery current until the state of charge returns to the starting point. After the state of charge is restored, the hybrid simulator reads the fuel map table according to the target boost intake state and engine candidate power, reads the fuel mass flow rate cycle by cycle and multiplies it by the simulation cycle duration to generate cycle fuel mass, accumulates the cycle fuel mass and multiplies it by the fuel carbon content coefficient, and takes the negative value of the product to generate carbon emission reward. The fuel mapping table refers to an engine registration table that uses engine candidate power, turbocharger speed and intake pressure as indexes and fuel mass flow rate as output. The fuel carbon content coefficient refers to the carbon mass fraction of fuel read from the fuel refueling batch record corresponding to the current fuel tank by the ship's edge computing node, and then multiplied by the carbon mass fraction of fuel by the ratio of the molecular weight of carbon dioxide of 44 to the atomic weight of carbon of 12 to obtain the carbon dioxide mass conversion value corresponding to the unit fuel mass. When a fuel tank switch occurs during the same voyage, the ship's edge computing node reads the fuel refueling batch record corresponding to the fuel tank after the switch according to the fuel tank switch time and updates the fuel carbon content coefficient.

[0013] In a preferred embodiment, the power allocation optimization module further includes: When executing the trust region policy optimization algorithm, the engine candidate power index is used to reinforce the registered value of the policy parameters. The registered value of the policy parameters is exponentially taken and divided by the sum of all exponents to generate the current action probability. The current action probability is multiplied by the carbon emission reward and summed to generate the current policy reward. The current policy reward is generated by subtracting the current policy reward from the carbon emission reward. The updated policy parameter register value is generated by adding the parameter increment to the current reinforcement learning policy parameter register value. The updated policy parameter register value is exponentially divided by the sum of all exponents to generate the updated action probability. The updated action probability is divided by the current action probability, multiplied by the advantage value, and summed according to the current action probability to generate the policy objective function. The average relative entropy is generated by multiplying the current action probability by the natural logarithm of the updated action probability and summing the results. The trust region radius is generated by dividing one by the sum of the number of particles and the particle deviation count. The first derivative of the policy objective function is used to generate the gradient vector, and the second derivative of the average relative entropy is used to generate the Fisher matrix. The update direction is obtained through the conjugate gradient. The update direction is scaled by the trust region radius and then halved successively to generate parameter increments. The first parameter increment that satisfies the condition that the average relative entropy does not exceed the trust region radius and the carbon emission reward is not lower than the current policy reward is accumulated and written into the corresponding reinforcement learning policy parameter register value. Then, the engine candidate power is sorted in descending order of update action probability. If the probabilities are the same, the first engine candidate power is selected in ascending order to generate the engine target power. The target power of the battery is generated by subtracting the propulsion demand power from the engine target power. Finally, the power allocation command is output.

[0014] The technical effects and advantages of this invention are as follows: 1. This solution replaces engine power with a battery while maintaining constant propulsion input, allowing the recovery response to expose the difference in boost intake, thus relatively reducing the carbon emission prediction bias caused by the combination of external conditions. 2. Use the hybrid power simulator to calculate replacement power step by step and verify the registration limitations of the engine, converter and battery to ensure that the replacement instructions are executable before execution; 3. Set a fuel index offset for the fuel mass flow rate and correct the state of charge by integrating the DC side current, so that the power state sequence is relatively reduced by the impact of timing mismatch. 4. In the ancestor sampling particle Gibbs algorithm, the fuel recovery difference and internal response difference are reused to preserve the historical trajectory source for the boost intake state identification, thereby improving the state confusion problem; 5. The trust region radius is generated by particle deviation counting, so that the update amplitude of the reinforcement learning policy is adjusted according to the dispersion of state recognition, and the power allocation of the suppression is repeatedly corrected. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the system modules of the present invention; Figure 2 This is a system diagram showing the internal components of the edge state acquisition module of the present invention; Figure 3 This is a system diagram illustrating the operating principle of the edge state acquisition module of the present invention; Figure 4 This is a system diagram showing the internal components of the replacement solver module of the present invention; Figure 5 This is a system diagram illustrating the operating principle of the replacement solver module of the present invention; Figure 6 This is a system diagram showing the internal components of the centralized execution module of the present invention; Figure 7 This is a system diagram illustrating the operating principle of the centralized execution module of the present invention; Figure 8 This is a system diagram showing the internal components of the state identification module of the present invention; Figure 9 This is a system diagram illustrating the operating principle of the state identification module of the present invention; Figure 10 This is a system diagram showing the internal components of the power distribution optimization module of the present invention; Figure 11 This is a system diagram illustrating the operating principle of the power allocation optimization module of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Refer to the instruction manual appendix Figures 1 to 11 The present invention provides a carbon emission prediction-driven power allocation optimization system for hybrid-powered ships, comprising: The edge state acquisition module is set on the shipborne edge computing node. It generates engine power code and propulsion power code based on shaft end torque and synchronous speed, reads fuel mass flow rate by time index and integrates DC side current to generate charge state code, and outputs power state sequence. In this embodiment, the edge state acquisition module is used to write the engine power, propulsion power, fuel mass flow rate, and state of charge after the replacement command is issued into the same-time index for the replacement solution module to read; the power code is generated from the torque value and speed value under the same sampling number, the fuel mass flow rate is aligned with the measured value through the engine simulator, and the state of charge code is obtained by integrating the DC side current; the implementation process includes the following steps: When generating engine power code and propulsion power code, the shipborne edge computing node uses the time of replacement instruction issuance as the zero index, reads the sampling number in the engine shaft torque frame, retrieves the engine speed value under the same sampling number, multiplies the engine shaft torque value and the engine speed value, and converts it into engine power code according to the ratio of engine power register. The shipborne edge computing node then reads the propulsion shaft end torque value and propulsion shaft speed value according to the same sampling number. After multiplying the propulsion shaft end torque value and propulsion shaft speed value, it converts them into propulsion power code according to the proportion of the propulsion power register. If the same sampling number lacks a torque value or speed value, the shipborne edge computing node does not write the corresponding index bit and reads the next sampling number to recalculate. When determining the fuel index offset, the shipborne edge computing node uses the zero-bit index engine power code as the power code before replacement, and writes the first index bit after the zero-bit index that is different from the power code before replacement as the first change index. When the first change index exists, the shipborne edge computing node generates candidate offset bits bit by bit from the zero index to the first change index. The engine simulator reads the engine fuel mapping table according to the engine power code, engine speed value, the turbocharger speed code of the previous index and the intake pressure code of the previous index, generates the injection mass index by index and divides it by the sampling period to obtain the inferred fuel mass flow rate. The shipborne edge computing node shifts the measured fuel mass flow rate backward according to the candidate offset position. After the shift, the index position without a measured value is not included in the calculation. Then, the difference between the inferred fuel mass flow rate and the measured fuel mass flow rate after the shift is calculated bit by bit. The absolute value of the difference is divided by the mass flow rate difference corresponding to the adjacent digital code of the fuel mass flow sensor to generate a flow difference count. The first position of the candidate offset position is taken in ascending order of the flow difference count to generate the fuel index offset position. If there is no first index change, the fuel index offset position is written to zero. When generating the charge status code, the shipboard edge computing node shifts the fuel mass flow rate according to the fuel index offset bit, and reads the DC side current starting from the index bit of the engine power code that leaves the previous power code. The product of the DC side current and the sampling period is accumulated bit by bit to the index bit of the engine power code that returns to the previous power code to generate the accumulated charge. When the DC current flows out of the battery, the accumulated charge is positive. The shipborne edge computing node deducts the ratio of the accumulated charge to the rated capacity of the battery from the zero-bit index charge status code. When the DC current flows into the battery, the accumulated charge is negative. The shipborne edge computing node adds the zero-bit index charge status code according to the same operation, and then converts it into a charge status code according to the proportion of the charge status register. The shipborne edge computing node writes the engine power code, propulsion power code, translated fuel mass flow rate and charge status code into the same index bit, and outputs the power status sequence. In practical applications: After receiving the replacement instruction, the shipborne edge computing node writes the time of issuance as the zero-bit index and saves the engine power code of the zero-bit index; after the engine power code changes in subsequent index bits, the shipborne edge computing node uses the engine simulator to generate the inferred fuel mass flow rate and selects the candidate offset bit that ranks first in the flow difference count as the fuel index offset bit. Subsequently, the shipborne edge computing node integrates the DC-side current during the replacement period to obtain the battery charge change during the replacement period, and writes the engine power code, propulsion power code, fuel mass flow rate and charge status code into the power state sequence for the replacement solution module to continue reading.

[0018] Replace the solver module, increment the power code trial calculation according to the engine control resolution, use the hybrid power simulator to deduct the engine power and supplement the propulsion power by the battery, verify the equipment registration interval along the power chain, and read the previous power code trial calculation when the boundary is crossed for the first time to generate a replacement instruction; In this embodiment, the replacement solver module is used to calculate the amount of power that the engine can relinquish without changing the propulsion requirements, and writes the amount of current that the battery needs to compensate as a replacement instruction. The replacement solver module first generates a trial power code according to the engine control resolution, then the hybrid power simulator calculates the engine load reduction, bus power gap, and battery discharge results cycle by cycle, and finally uses equipment registration data to determine whether the trial power code can be executed. This implementation process includes the following steps: When determining the trial power code, the shipborne edge computing node reads the current engine power code, the current charge status code, and the bus voltage from the last index of the power state sequence. The bus voltage is the DC bus voltage sample value written to the last index of the power state sequence. The engine controller is an electronic control unit that receives the engine power code and outputs control quantities to the fuel injection actuator. The shipborne edge computing node reads the steady-state lower limit power code, engine control resolution and power drop rate from the engine controller. The steady-state lower limit power code is the lower limit of the power code for maintaining steady-state combustion control of the engine. The engine control resolution is the power difference between two adjacent writable power codes in the engine power command register. The shipborne edge computing node obtains the load reduction margin by subtracting the steady-state lower bound power code from the current engine power code. The load reduction margin is divided by the engine control resolution and the integer quotient is taken. The sequence number is incremented from one to the integer quotient plus one. The sequence number is multiplied by the engine control resolution to generate the trial power code. When the integer quotient is zero, the trial power code corresponding to sequence number one is still generated. When generating the load reduction power trajectory, the hybrid power simulator reads a trial power code and subtracts the trial power code from the current engine power code to generate a trial engine power code. The hybrid power simulator is a computing program that runs within a shipboard edge computing node, including an engine simulator, engine-to-DC bus efficiency calculation, battery capacity calculation, and converter current calculation. The shipborne edge computing node multiplies the power reduction rate by the simulation cycle duration, then quantizes it according to the engine control resolution to generate a single-cycle load reduction code. When the quantization result is zero, it is written to an engine control resolution. The hybrid simulator deducts the single-cycle load reduction code from the current engine power code cycle by cycle. When the remaining deduction amount is lower than the single-cycle load reduction code, it is directly written into the trial engine power code, and the cycle-by-cycle engine power code is written as the load reduction power trajectory. When calculating the compensation current code, the hybrid power simulator reads the engine-to-DC bus efficiency table cycle by cycle according to the load reduction power trajectory. The engine-to-DC bus efficiency table is a registration table with the engine power code as the index and the efficiency value after the engine shaft power is transferred to the DC bus as the output. For two adjacent engine power codes, the hybrid simulator will multiply the power code of the engine before deduction by the corresponding efficiency value to obtain the bus power before deduction, and multiply the power code of the engine after deduction by the corresponding efficiency value to obtain the bus power after deduction. Then, the bus power difference is obtained by subtracting the bus power after deduction from the bus power before deduction. The bus voltage of the same period in the replacement solution stage is taken from the bus voltage written by the last index of the power state sequence. The hybrid power simulator divides the bus power difference by the bus voltage to generate a compensation current code. The compensation current code takes the battery discharging to the DC bus as a positive value. When generating an out-of-bounds bit, the shipborne edge computing node compares the trial engine power code with the steady-state lower bound power code. If the trial engine power code is lower than the steady-state lower bound power code, it writes the engine out-of-bounds bit. The shipborne edge computing node compares the compensation current code with the upper limit of the discharge current registered by the bidirectional converter of the battery. If the compensation current code is higher than the upper limit of the discharge current, it is written into the converter's out-of-bounds bit. The battery management system is a battery control unit that collects battery current, voltage and capacity data and outputs state of charge and discharge limit data to the shipboard edge computing node. The shipboard edge computing node multiplies the compensation current code with the simulation cycle duration cycle by cycle and accumulates them to generate compensation power. Then, it subtracts the ratio of compensation power to the rated capacity registered by the battery management system from the current state of charge code to generate a trial state of charge. When the calculated state of charge is lower than the discharge termination state of charge registered by the battery management system, the shipborne edge computing node writes the battery out-of-bounds bit. When generating replacement instructions, the shipborne edge computing node reads the engine over-limit position, converter over-limit position and battery over-limit position according to the sequence number. When any over-limit position is read for the first time, the previous sequence number is taken as the target sequence number. If no out-of-bounds position is read in any of the sequences, the shipborne edge computing node takes the last sequence as the target sequence; if the target sequence is zero, the shipborne edge computing node generates a zero code replacement instruction. When the target order is not zero, the shipborne edge computing node reads the load reduction power trajectory and compensation current code corresponding to the target order, writes the forward adjacent difference of the load reduction power trajectory as the engine load reduction code, writes the reverse adjacent difference of the load reduction power trajectory as the engine recovery code, writes the compensation current code in reverse order as the withdrawal current code, and writes the engine load reduction code, compensation current code, engine recovery code and withdrawal current code into the replacement instruction according to the execution order. Through the above processing, the replacement solution module finds the amount of power that the engine can give up by trial calculation at each level. At the same time, it limits the replacement command with the battery discharge capacity, the inverter current capacity and the engine steady-state combustion capacity to avoid the centralized execution module receiving an unexecutable load reduction. In practical applications: When a ship is sailing at a constant speed, if the current engine power code is higher than the steady-state lower limit power code, the replacement solution module starts the trial calculation from one engine control resolution. The hybrid power simulator deducts the engine power and calculates the battery compensation current cycle by cycle. When the trial calculation order makes the compensation current code higher than the upper limit of the discharge current registered by the converter, the shipborne edge calculation node takes the previous order as the target order and writes the corresponding engine load reduction code, compensation current code, engine recovery code and withdrawal current code into the replacement instruction.

[0019] The centralized execution module reduces engine power and rewrites converter current according to bus missing power by the shipborne edge computing node, keeps propulsion motor input power unchanged, restores engine power and withdraws converter current, and outputs measured recovery sequence. In this implementation, the centralized execution module writes the engine load reduction code, compensation current code, engine recovery code, and withdrawal current code from the replacement instruction into the corresponding control registers, and corrects the converter current by the difference in the propulsion motor input power, so that the engine load reduction and recovery process does not change the propulsion motor input power; the shipborne edge computing node saves the engine power code and converter current code at the starting point of the replacement instruction execution, and drives subsequent writing with the local clock; this implementation process includes the following steps: When establishing the execution baseline, the shipborne edge computing node reads the bus voltage and propulsion motor input current at the starting point of the replacement instruction execution. The propulsion motor input current is the sampled value of the current flowing from the DC bus into the DC input terminal of the propulsion motor controller. The shipborne edge computing node multiplies the bus voltage with the propulsion motor input current and converts it into a reference input power code according to the power register ratio. At the same time, it saves the engine power code and converter current code at the execution start point, and then reads the engine load reduction code and compensation current code in sequence according to the replacement instruction. When the starting point of execution lacks bus voltage or propulsion motor input current, the shipborne edge computing node does not initiate a replacement instruction and rereads it on the next local clock. When performing load reduction, the shipborne edge computing node reads the current engine power register value and the current converter current register value on the same local clock edge, subtracts the engine load reduction code from the current engine power register value and writes it back to the engine controller, and adds the compensation current code to the current converter current register value and writes it back to the battery bidirectional converter. At the next local clock, the shipborne edge computing node resamples the bus voltage and propulsion motor input current, quantizes their product into a measured input power code, subtracts the measured input power code from the reference input power code, divides it by the bus voltage to generate a current correction code, and accumulates the current correction code into the converter current code; the current correction code is only written to the converter current register, without rewriting the engine load reduction code and compensation current code; During the recovery process, the shipboard edge computing node reads the engine recovery code and the withdrawal current code in reverse order according to the replacement instruction. It reads the current engine power register value and the current converter current register value on the same local clock edge. After adding the engine recovery code to the current engine power register value, it writes it back to the engine controller. After subtracting the withdrawal current code from the current converter current register value, it writes it back to the battery bidirectional converter. At the next local clock, the shipborne edge computing node continues to generate a current correction code based on the difference between the reference input power code and the measured input power code, and accumulates the current correction code into the converter current code; When the engine power register value returns the engine power code saved at the execution start point and the converter current register value returns the converter current code saved at the execution start point, the shipborne edge computing node ends the replacement instruction, and the last write directly writes back the two register values ​​saved at the execution start point. When generating the measured recovery sequence, the shipborne edge computing node reads the records where the engine power code is first equal to the same instruction code in the load reduction stage and the recovery stage respectively. The fuel recovery difference is generated by subtracting the fuel mass flow rate in the load reduction stage from the fuel mass flow rate in the recovery stage. Then, the product of the current correction code in the recovery stage and the corresponding bus voltage is subtracted from the product of the current correction code in the load reduction stage and the corresponding bus voltage to generate the internal response difference. The shipborne edge computing node writes the same instruction code, fuel recovery difference, and internal response difference into the measured recovery sequence according to the replacement instruction order, for the state identification module to read; When the same instruction code lacks a corresponding record during the unloading or recovery phase, the shipboard edge computing node deletes the corresponding sequence number and does not generate fuel recovery difference or internal response difference. Through the above processing, the centralized execution module writes to the engine power register and the converter current register through the same clock edge, so that the battery compensation follows the engine load reduction; the current correction code only acts on the converter current register, so that the propulsion motor input power is kept near the reference input power code, while retaining the engine's real recovery response for the state identification module to use. In practical applications: After receiving the replacement command, the shipborne edge computing node first saves the engine power code, converter current code, and reference input power code at the execution start point; then, each local clock reduces the engine power and increases the battery discharge current according to the replacement command. When the measured input power of the propulsion motor is lower than the reference input power code, the shipborne edge computing node only increases the converter current code; during the recovery phase, the engine power is restored and the battery current is withdrawn in the reverse order. Finally, the difference in fuel mass flow rate and current correction difference between the same engine power code during the load reduction phase and the recovery phase are written into the measured recovery sequence.

[0020] The state identification module performs an ancestor sampling particle Gibbs algorithm on the measured recovery sequence, generates boosted intake particles according to the state transition probability, uses a hybrid power simulator to deduce fuel flow, generates particle weights based on fuel flow difference and power recovery difference, extracts ancestors according to particle weights and resamples, and outputs the target boosted intake state with deviation count. In this embodiment, the state identification module is used to deduce the boosted intake state of the engine during load reduction and recovery based on the measured recovery sequence, so as to provide the power distribution optimization module with a way to correct the hybrid power simulator. The state identification module first converts the measured recovery sequence into a sequential observation record, then generates boosted intake particles, and then updates the particle trajectory through Gibbs resampling of ancestor sampled particles, and outputs the target boosted intake state and particle deviation count. This implementation process includes the following steps: When generating the sequential observation record, the state identification module reads the fuel recovery difference and internal response difference in the measured recovery sequence according to the replacement instruction sequence. It divides the absolute value of the fuel recovery difference by the fuel flow detection resolution to generate the fuel difference count, and divides the absolute value of the internal response difference by the engine control resolution to generate the power difference count. Then, it writes the replacement instruction sequence, fuel difference count, and power difference count into the sequential observation record. The fuel flow detection resolution is taken from the mass flow difference between two adjacent output digital codes in the fuel mass flow sensor calibration table. The order of missing fuel recovery difference or internal response difference in the measured recovery sequence is not written into the order observation record. When the number of order observation records is zero, the previous round of target boost intake state is used and the particle deviation count is written to zero. When generating boosted intake particles, the number of generated particles is increased by one from the number of sequential observation records, and the lower limit of the booster speed and the upper limit of the booster speed are read from the engine controller. The particle number starts from zero and increments. The particle interval is the upper limit of the turbocharger speed minus the lower limit of the turbocharger speed, divided by the number of particles minus one. The turbocharger speed particle is the lower limit of the turbocharger speed plus the particle number multiplied by the particle interval. The intake pressure particle of the previous index of the first index is taken from the target intake pressure of the previous round. If there is no target intake pressure of the previous round, the steady-state intake pressure calculated by the engine simulator according to the power code before replacement is taken. The hybrid simulator reads the compressor flow meter to obtain the compressor intake mass flow rate using the turbocharger speed particle and the previous index intake pressure particle. It reads the engine intake flow meter to obtain the cylinder intake mass flow rate using the engine power code, engine speed value, and the previous index intake pressure particle. The difference between the compressor intake mass flow rate and the cylinder intake mass flow rate is then multiplied by the sampling period, intake temperature, and gas constant, divided by the intake manifold volume, and accumulated to the previous index intake pressure particle to generate the current index intake pressure particle. The turbocharger speed particle and the intake pressure particle with the same number are combined to form a boosted intake particle. When updating the particle trajectory, the hybrid simulator executes the replacement command with the same boosted intake particle, and records the fuel mass flow rate and the actual engine power code when the engine power code first equals the same command code in the unloading phase and the recovery phase, respectively. The fuel difference is generated by subtracting the fuel mass flow rate in the unloading phase from the fuel mass flow rate in the recovery phase, and the power difference is generated by subtracting the actual engine power code in the unloading phase from the actual engine power code in the recovery phase. The absolute value of the difference between the fuel difference and the fuel recovery difference in the sequential observation record is divided by the fuel flow detection resolution to generate the fuel residual count. The absolute value of the difference between the power difference and the internal response difference in the sequential observation record is divided by the engine control resolution to generate the power residual count. The fuel residual count and the power residual count are added together to generate the residual count. The state transition probability is generated by subtracting the predicted turbocharger speed from the current indexed turbocharger speed particle and dividing by the particle interval to generate a transition difference count. The result is obtained with the natural constant as the base and the negative value of the transition difference count as the exponent. The predicted turbocharger speed is obtained by adding the difference between the turbine torque and the compressor torque to the previous indexed turbocharger speed particle, multiplying by the sampling period, and then dividing by the rotor moment of inertia. The negative residual count is used as the natural constant exponent to generate particle weights. The sum of the residual counts from the current order to the last order of observation records is used as the tail residual count. The ancestor product value is generated by multiplying the previous order particle weight, the state transition probability and the negative tail residual count exponent. The ancestor product value is divided by the sum of all ancestor product values ​​to obtain the ancestor extraction probability. Then, the ancestor extraction probabilities are accumulated according to the particle number to form a probability interval. The particle number falling into the probability interval is read by the equidistant extraction points to complete the Gibbs resampling of ancestor sampling particles. When outputting the target boosted intake state, the particle weight is divided by the sum of all particle weights to obtain the normalized particle weight. When the sum of all particle weights is quantized to zero according to the register resolution, the first particle weight is taken in ascending order of residual count and written as one, while the other particle weights are written as zero. The target turbocharger speed is obtained by multiplying the turbocharger speed particle of each turbocharger intake particle by the normalized particle weight and summing the results. The target intake pressure is obtained by multiplying the intake pressure particle of each turbocharger intake particle by the normalized particle weight and summing the results. The target turbocharger speed and the target intake pressure together form the target turbocharger intake state. Then, the absolute value of the difference between each turbocharger speed particle and the target turbocharger speed is divided by the particle interval and then weighted and summed. The absolute value of the difference between each intake pressure particle and the target intake pressure is divided by the intake pressure particle interval and then weighted and summed. The two sums are then added together to generate the particle deviation count. When the target boosting air intake state in this round is XORed bit by bit with the target boosting air intake state in the previous round and the ancestor index sequence in this round is XORed bit by bit with the ancestor index sequence in the previous round, resampling stops. If the number of resampling rounds reaches the number of sequential observation records and still does not stop, the first resampling result is taken according to the residual count and ascending order, and the target boosting air intake state with particle deviation count is output. Through the above processing, the state identification module uses the fuel change and power response in the measured recovery sequence for particle weight calculation at the same time, and retains the boost intake trajectory that can explain the measured recovery sequence before and after engine unloading through ancestor sampling particle Gibbs resampling, avoiding classifying different boost intake states into the same state based solely on the current engine power code. In practical applications: After the engine experiences surge unloading, different fuel mass flow rates occur during the unloading and recovery phases under the same engine power code. The state identification module writes the difference into the sequential observation record, and then generates boosted intake particles from the turbocharger speed range. The hybrid power simulator replays the replacement instruction particle by particle, and particles with low residual counts are retained in the resampling. Finally, the target boosted intake state and particle deviation count are output for the power distribution optimization module to calculate the trust region radius.

[0021] The power allocation optimization module corrects the hybrid simulator based on the target boosted intake state, generates candidate engine power through reinforcement learning and uses the battery to take over the remaining power, simulates the carbon emission return after the state of charge is restored, generates the trust region radius by dividing one by the sum of the particle number and the deviation count, executes the trust region policy optimization algorithm, obtains the update direction through the conjugate gradient and halves the update step size successively until the average relative entropy does not exceed the trust region radius and the carbon emission return is not lower than the value before the update, and outputs the power allocation command. In this embodiment, the power allocation optimization module is used to recalculate the power sharing between the engine and the battery after the target boosted intake state has been determined, and to update the reinforcement learning policy within the trust region. The power allocation optimization module first enumerates candidate engine power, then restores the battery state of charge to the simulation starting point and calculates the carbon emission reward, and then uses the trust region policy optimization algorithm to write back the reinforcement learning policy parameters. This implementation process includes the following steps: When generating candidate power records, the shipborne edge computing node writes the target boosted intake state into the hybrid power simulator and generates engine candidate power bit by bit from the lower power bound to the upper power bound registered by the engine controller according to the engine control resolution. The propulsion power requirement is calculated from the speed control command, propulsion shaft speed and propulsion shaft torque. The shipborne edge computing node generates the battery candidate power by subtracting the engine candidate power from the propulsion power requirement, and then generates the battery candidate current by dividing the battery candidate power by the bus voltage. When the candidate power of the battery is positive, it is judged according to the upper limit of the discharge current registered by the inverter; when the candidate power of the battery is negative, it is judged according to the upper limit of the charging current registered by the inverter. Records that do not exceed the corresponding upper limit of the current are written as candidate power records. When restoring the state of charge, the hybrid power simulator calculates the candidate power of the engine and the candidate power of the battery in the candidate power record according to the number of records in the measured recovery sequence, and updates the state of charge according to the candidate current of the battery in each simulation cycle. After the calculation is completed, the recovery difference is generated by subtracting the current state of charge from the starting state of charge. The recovery difference is multiplied by the rated capacity registered by the battery management system to generate the recovery charge. The recovery charge is then divided by the product of the bus voltage and the simulation cycle duration to generate the recovery current. When the recovery current exceeds the converter's registered current limit, the corresponding current limit is written. In the last cycle, the recovery current is generated according to the remaining recovery charge and the state of charge is written back to the starting state of charge. The recovery engine power is obtained by subtracting the product of the recovery current and the bus voltage from the propulsion demand power. When generating carbon emission returns, the hybrid simulator reads the fuel map table according to the target boosted intake state. The fuel map table is indexed by engine power, turbocharger speed and intake pressure and outputs fuel mass flow rate. The fuel mass flow rate is read according to the candidate power of the engine during the candidate power cycle and the fuel mass flow rate is read according to the restored engine power during the recovery cycle. The fuel mass flow rate of each cycle is multiplied by the simulation cycle duration to obtain the cycle fuel mass. The shipboard edge computing node accumulates the periodic fuel mass, multiplies it by the fuel carbon content coefficient, and then takes a negative value to generate carbon emission reports. The fuel carbon content coefficient is obtained by multiplying the fuel carbon mass fraction in the fuel refueling batch record corresponding to the current fuel tank by the ratio of the molecular weight of carbon dioxide of 44 to the atomic weight of carbon of 12. When switching fuel tanks, the fuel refueling batch record of the fuel tank after the switch is updated. When generating the advantage value, the reinforcement learning policy parameter register value is indexed by the engine candidate power. The index of each reinforcement learning policy parameter register value is taken and then divided by the sum of all the indexes to obtain the current action probability of the corresponding engine candidate power. The current action probability is multiplied by the corresponding carbon emission reward and summed to generate the current strategy reward. Then, the carbon emission reward of each candidate power record is subtracted from the current strategy reward to generate the advantage value of the corresponding engine candidate power. When the candidate power record is empty, the power allocation optimization module does not update the reinforcement learning strategy parameters and uses the power allocation instruction of the previous cycle. When calculating the trust region update, the updated policy parameter register value is generated by adding the parameter increment to the current reinforcement learning policy parameter register value. The updated policy parameter register value is then exponentially and normalized to generate the update action probability. The objective function of the strategy is the sum of the updated action probability corresponding to each candidate engine power divided by the current action probability, multiplied by the advantage value, and multiplied by the current action probability. The average relative entropy is the sum of the natural logarithm of the ratio of the current action probability multiplied by the current action probability divided by the updated action probability for each engine candidate power. The trust region radius is obtained by dividing one by the sum of the particle count and the particle deviation count, which are output by the state identification module. When writing back the power allocation instruction, the first derivative of the policy objective function is calculated to generate the gradient vector, the second derivative of the average relative entropy is calculated to generate the Fisher matrix, and the update direction is initialized with the zero vector and the residual vector and search vector are initialized with the gradient vector. In each round of conjugate gradient calculation, the conjugate step size is generated by the inner product of the residual and the inner product after the transformation of the search vector's Fischer matrix. The update direction, residual vector, and search vector are updated. The residual vector is quantized to all zeros according to the resolution of the strategy parameter register, or the calculation stops when the number of rounds reaches the number of candidate power records. The power allocation optimization module generates initial parameter increments according to the update direction, Fisher matrix, and trust region radius. It divides the parameter increments by two in each round and recalculates the update action probability, average relative entropy, and update policy reward in each round. The first parameter increment that satisfies the condition that the average relative entropy does not exceed the trust region radius and the update policy reward is not lower than the current policy reward is accumulated and written into the reinforcement learning policy parameter register value. When no parameter increment is written, the engine candidate power is selected in descending order of the current action probability. After the parameter increment is written, the engine candidate power is selected in descending order of the updated action probability. When the probabilities are the same, the first engine candidate power is selected in ascending order to generate the engine target power. Then, the target power of the battery is generated by subtracting the target power of the engine from the propulsion demand power, and the power allocation command is output. Through the above processing, before selecting the target engine power, the power distribution optimization module first incorporates the battery state of charge recovery process into the carbon emission return calculation, avoiding taking the short-term low fuel consumption brought by discharge as the low carbon result directly; the trust region radius is generated by the number of particles and particle deviation count output by the state identification module, so that the strategy update amplitude shrinks when the target boost intake state is dispersed. In practical applications: After a ship enters a windy and wave-prone area, the state identification module outputs the target turbocharger speed, target intake pressure, and particle deviation count. Based on this, the power allocation optimization module enumerates candidate engine power and calculates the fuel mass flow rate required for battery compensation and recovery for each candidate. If a candidate engine power reduces the fuel mass flow rate in the current cycle but requires more engine power compensation when recovering the state of charge, the carbon emission reward will deduct this result. The reinforcement learning policy parameters are only written when the average relative entropy satisfies the trust region radius and the updated policy reward is not lower than the current policy reward. Finally, the target engine power and target battery power are generated.

[0022] Working Principle: This scheme first collects data on the engine, propulsion shaft, fuel, and battery from the shipborne edge computing node, generating a power state sequence under the same index. Then, the hybrid power simulator calculates the engine power that can be reduced and the current that the battery needs to compensate for, selecting replacement instructions that will not cause the engine, converter, and battery to exceed their limits. When the shipborne edge computing node executes the replacement instructions, it reduces the engine power and simultaneously increases the battery output to keep the propulsion motor input power constant. Subsequently, it restores the engine power and withdraws the battery compensation, obtaining the measured recovery sequence. The state identification module uses the ancestor sampling particle Gibbs algorithm to deduce the engine's target boost intake state based on the fuel recovery difference and power response difference. The power allocation optimization module then uses the target boost intake state to correct the hybrid power simulator, updates the reinforcement learning strategy through the trust region policy optimization algorithm, and finally outputs the engine target power and battery target power. In practical applications, when ships are sailing at a constant speed in windy and wave areas, the propeller load will fluctuate briefly. Although the engine's surface power and speed will return to their original values, the internal state of the turbocharger and intake manifold may not be the same. Directly allocating power according to conventional data will cause carbon emission prediction errors. This solution allows the battery to temporarily take over part of the engine power while keeping the propulsion power constant, and then observes the changes in fuel and power when the engine recovers, thereby identifying the hidden turbocharged intake state. Subsequently, the system compares the candidate power of different engines in the simulator and includes the recharging process after the battery discharges in the carbon emission reward, so that reinforcement learning will not misjudge short-term discharge as a low-carbon strategy, thus completing a more reliable power allocation under real navigation conditions.

[0023] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A carbon emission prediction-driven power distribution optimization system for hybrid-powered ships, characterized in that, include: The edge state acquisition module is set on the shipborne edge computing node. It generates engine power code and propulsion power code based on shaft end torque and synchronous speed, reads fuel mass flow rate by time index and integrates DC side current to generate charge state code, and outputs power state sequence. Replace the solver module, increment the power code trial calculation according to the engine control resolution, use the hybrid power simulator to deduct the engine power and supplement the propulsion power by the battery, verify the equipment registration interval along the power chain, and read the previous power code trial calculation when the boundary is crossed for the first time to generate a replacement instruction; The centralized execution module reduces engine power and rewrites converter current according to bus missing power by the shipborne edge computing node, keeps propulsion motor input power unchanged, restores engine power and withdraws converter current, and outputs measured recovery sequence. The state identification module performs an ancestor sampling particle Gibbs algorithm on the measured recovery sequence, generates boosted intake particles according to the state transition probability, uses a hybrid power simulator to deduce fuel flow, generates particle weights based on fuel flow difference and power recovery difference, extracts ancestors according to particle weights and resamples, and outputs the target boosted intake state with deviation count. The power allocation optimization module corrects the hybrid power simulator based on the target boosted intake state. It generates candidate engine power through reinforcement learning and the battery takes over the remaining power. It simulates the carbon emission return after the state of charge is restored. It generates the trust region radius by dividing one by the sum of the particle number and the deviation count. It executes the trust region policy optimization algorithm, obtains the update direction through the conjugate gradient, and halves the update step size successively until the average relative entropy does not exceed the trust region radius and the carbon emission return is not lower than the value before the update. Finally, it outputs the power allocation command.

2. The carbon emission prediction-driven power distribution optimization system for hybrid-powered ships according to claim 1, characterized in that: The edge state acquisition module includes: Using the time when the replacement command is issued as the zero index, the engine shaft end torque value and engine speed value are read according to the sampling number and multiplied to generate the engine power code. The propulsion shaft end torque value and propulsion shaft speed value are read according to the same sampling number and multiplied to generate the propulsion power code. The engine power code with zero index is used as the power code before replacement. The first index bit after zero index that is different from the power code before replacement is used as the first change index. Candidate offset bits are generated bit by bit from zero index to first change index. The engine simulator deduces the fuel mass flow rate according to the engine power code. The measured fuel mass flow rate is shifted according to the candidate offset bits and the difference between the measured fuel mass flow rate and the deduce fuel mass flow rate is calculated bit by bit and converted into a flow difference count. The first bit of the candidate offset bits is taken in ascending order of flow difference count and ascending order to generate the fuel index offset bit. The fuel mass flow rate is shifted according to the fuel index offset position. The engine power code is moved from the index position of the power code before replacement to the index position of the power code before replacement. The product of the DC side current and the sampling period is accumulated to generate the accumulated charge. The accumulated charge is used to correct the zero-position index charge status code and generate the charge status code. The engine power code, propulsion power code, fuel mass flow rate and charge status code are written to the same index position, and the power status sequence is output.

3. The carbon emission prediction-driven power distribution optimization system for hybrid-powered ships according to claim 2, characterized in that: The replacement solution module includes: Based on the last index of the power state sequence, read the current engine power code, current charge state code and bus voltage, and read the steady-state lower limit power code, engine control resolution and power decrease rate registered by the engine controller. Subtract the steady-state lower limit power code from the current engine power code to obtain the load reduction margin, and generate the trial power code by dividing the load reduction margin by the engine control resolution. For each calculated power code, the calculated engine power code is generated by subtracting the calculated power code from the current engine power code. The power reduction rate is multiplied by the simulation cycle duration and quantized according to the engine control resolution to generate a single-cycle load reduction code. The hybrid simulator subtracts the single-cycle load reduction code from the current engine power code cycle by cycle, and generates a load reduction power trajectory when the calculated engine power code is reached. Read the engine to DC bus efficiency table according to the load reduction power trajectory, multiply the adjacent engine power codes by the corresponding efficiency values ​​and subtract them to generate the bus power difference, and divide the bus power difference by the bus voltage of the same period to generate the compensation current code.

4. The carbon emission prediction-driven power distribution optimization system for hybrid-powered ships according to claim 3, characterized in that: The replacement solution module further includes: If the calculated engine power code is lower than the steady-state lower limit power code, the compensation current code is higher than the inverter's registered discharge current upper limit, and the current charge status code, after deducting the compensation current code's integral charge and the ratio of the battery management system's registered rated capacity, is lower than the discharge termination charge status, write these values ​​into the engine, inverter, and battery over-limit positions, respectively. Read the engine over-limit position, inverter over-limit position, and battery over-limit position according to the sequence number. When the over-limit position is read for the first time, take the previous sequence number as the target sequence number. When the target sequence number is zero, generate a zero code replacement instruction. When the target sequence number is not zero, read the load reduction power trajectory and compensation current code of the target sequence number. Write the forward adjacent difference of the load reduction power trajectory as the engine load reduction code, write the reverse adjacent difference of the load reduction power trajectory as the engine recovery code, and write the compensation current code in reverse order as the withdrawal current code, and generate a replacement instruction.

5. A carbon emission prediction-driven power distribution optimization system for hybrid-powered ships according to claim 4, characterized in that: The centralized execution module includes: Starting from the point of execution of the replacement instruction, the bus voltage and the propulsion motor input current are read and multiplied to generate the reference input power code. Then, the engine load reduction code and compensation current code are read in sequence according to the replacement instruction. When load reduction is performed, the shipborne edge computing node writes the engine load reduction code and compensation current code on the same local clock edge. The next local clock generates the measured input power code by multiplying the bus voltage and the propulsion motor input current. The reference input power code is subtracted from the measured input power code and divided by the bus voltage to generate the current correction code. The current correction code is then added to the converter current code.

6. The carbon emission prediction-driven power distribution optimization system for hybrid-powered ships according to claim 5, characterized in that: The centralized execution module also includes: When performing recovery, the engine recovery code and withdrawn current code are read in reverse order according to the replacement instruction. The engine recovery code is written on the same local clock edge and the withdrawn current code is deducted from the converter current code. The next local clock generates a current correction code based on the difference between the reference input power code and the measured input power code and adds it to the converter current code until the engine power code and converter current code return to their values ​​before replacement. During the load reduction and recovery phases, the engine power code is paired with the same code and recorded. The fuel recovery difference is generated by subtracting the fuel mass flow rate of the load reduction phase from the fuel mass flow rate of the recovery phase. The internal response difference is generated by subtracting the product of the current correction code and the bus voltage of the recovery phase from the product of the current correction code and the bus voltage of the load reduction phase. The measured recovery sequence is written according to the replacement instruction order.

7. A carbon emission prediction-driven power distribution optimization system for hybrid-powered ships according to claim 6, characterized in that: The state identification module includes: Read the fuel recovery difference and internal response difference in the measured recovery sequence according to the replacement instruction order, divide the absolute value of the fuel recovery difference by the fuel flow detection resolution to generate the fuel difference count, divide the absolute value of the internal response difference by the engine control resolution to generate the power difference count, and write it into the sequential observation record; The number of particles is generated by adding one to the number of sequential observation records. The turbocharger speed range registered by the engine controller is evenly divided into turbocharger speed particles according to the number of particles. The intake pressure particles are calculated by subtracting the cylinder intake mass flow rate from the compressor intake mass flow rate. The turbocharger speed particles and intake pressure particles with the same number are combined into turbocharged intake particles.

8. A carbon emission prediction-driven power distribution optimization system for hybrid-powered ships according to claim 7, characterized in that: The state identification module further includes: The hybrid simulator extrapolates the fuel difference and power difference of each boosted intake particle according to the replacement instruction. It converts the difference between the extrapolation result and the sequential observation record into a residual count. The negative value of the residual count is used as the natural constant exponent to generate the particle weight. The ancestor extraction probability is generated by the previous sequential particle weight, the state transition probability and the tail residual count. The ancestor sampling particle Gibbs resampling is performed according to the ancestor extraction probability. After each round of resampling, the target boosted air intake state is generated according to the particle weight. The difference between each boosted air intake particle and the target boosted air intake state is converted into a particle deviation count. Resampling stops when the XOR result of the target boosted air intake state and the ancestor index sequence relative to the previous round is zero. When the number of resampling rounds reaches the number of sequential observation records, the first result is taken according to the residual count and ascending order, and the target boosted air intake state with particle deviation count is output.

9. A carbon emission prediction-driven power distribution optimization system for hybrid-powered ships according to claim 8, characterized in that: The power allocation optimization module includes: Write the target boost intake state into the hybrid simulator. Generate engine candidate power bit by bit from the lower power limit registered by the engine controller to the upper power limit according to the engine control resolution. Subtract the engine candidate power from the propulsion demand power to generate the battery candidate power. Divide the battery candidate power by the bus voltage to generate the battery candidate current. Write the records where the battery candidate current does not exceed the converter's registered current limit as candidate power records. For each candidate power record, the hybrid power simulator calculates the engine candidate power and battery candidate power according to the number of records in the measured recovery sequence. Then, it generates a recovery current based on the difference between the battery's state of charge and the starting point. The engine then bears the remaining propulsion power after the power corresponding to the recovery current until the state of charge returns to the starting point. After the state of charge is restored, the hybrid simulator reads the fuel map table according to the target boost intake state and engine candidate power, reads the fuel mass flow rate cycle by cycle and multiplies it by the simulation cycle duration to generate cycle fuel mass, accumulates the cycle fuel mass and multiplies it by the fuel carbon content coefficient, and takes the negative value of the product to generate carbon emission reward.

10. A carbon emission prediction-driven power distribution optimization system for hybrid-powered ships according to claim 9, characterized in that: The power allocation optimization module also includes: When executing the trust region policy optimization algorithm, the engine candidate power index is used to reinforce the registered value of the policy parameters. The registered value of the policy parameters is exponentially taken and divided by the sum of all exponents to generate the current action probability. The current action probability is multiplied by the carbon emission reward and summed to generate the current policy reward. The current policy reward is generated by subtracting the current policy reward from the carbon emission reward. The updated policy parameter register value is generated by adding the parameter increment to the current reinforcement learning policy parameter register value. The updated policy parameter register value is exponentially divided by the sum of all exponents to generate the updated action probability. The updated action probability is divided by the current action probability, multiplied by the advantage value, and summed according to the current action probability to generate the policy objective function. The average relative entropy is generated by multiplying the current action probability by the natural logarithm of the updated action probability and summing the results. The trust region radius is generated by dividing one by the sum of the number of particles and the particle deviation count. The first derivative of the policy objective function is used to generate the gradient vector, and the second derivative of the average relative entropy is used to generate the Fisher matrix. The update direction is obtained through the conjugate gradient. The update direction is scaled by the trust region radius and then halved successively to generate parameter increments. The first parameter increment that satisfies the condition that the average relative entropy does not exceed the trust region radius and the carbon emission reward is not lower than the current policy reward is accumulated and written into the corresponding reinforcement learning policy parameter register value. Then, the engine candidate power is sorted in descending order of update action probability. If the probabilities are the same, the first engine candidate power is selected in ascending order to generate the engine target power. The target power of the battery is generated by subtracting the propulsion demand power from the engine target power. Finally, the power allocation command is output.

Images