A transient power compensation control method and system for a ship DC networking

Abstract

The present application relates to the technical field, specifically to a kind of transient power compensation control method and system of ship dc networking, in the present application, difference comparison and stable interval positioning are carried out to ascending sequence and descending sequence by means of heuristic mining algorithm, threshold setting and trigger condition can be self-adaptive generated according to partition operating state, reduce false triggering and missing triggering caused by global fixed threshold, voltage upper limit, current lower limit, error amplitude and bus current difference are combined into feature group, rapid determination of node gain level is realized through multilayer feedforward neural network, ensure that different node compensation intensity is matched with its transient fluctuation amplitude, reduce overcompensation and undercompensation phenomenon, in gain execution link, compensation sequence formed is accurately issued through instruction coding, frame construction and address matching, in combination with the record of sending and confirming time, form traceable instruction link, improve abnormal response positioning efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of load distribution control, in particular to a transient power compensation control method and system for ship direct current networking. BACKGROUND

[0002] The technical field of load distribution control aims to realize the power coordination and optimized distribution among multiple power sources in a power supply system, to maintain the power balance and operational stability of the system, focusing on adjusting the output power of each power source according to the real-time changes of the system load, and collectively bearing the load demand according to the set strategy, to prevent problems such as single power source overload, system voltage or frequency fluctuation, etc.

[0003] A transient power compensation control method for ship direct current networking aims to solve the problem of transient voltage fluctuation in the direct current bus of the power system when the load of the ship changes. By compensating for transient power, it ensures that voltage fluctuations in the ship's direct current network will not affect the stability of the system and the safe operation of the equipment. The purpose is to optimize power distribution, monitor and adjust the output power of each power source in real time, ensure that transient load changes of the system are quickly and effectively compensated, and enable the system to maintain power balance when facing load fluctuations, avoiding voltage instability, power shortage or overload, and ultimately achieving stability and efficiency of the power system.

[0004] Traditional load distribution control lacks high-resolution sampling and multi-parameter joint screening in topology structure identification, has low partitioning accuracy, is prone to partition drift when node connections change, and causes power distribution to be inconsistent with the actual structure. In the voltage and current monitoring segment, it relies on global statistical values and single trend judgment, lacks sequence difference and stable interval identification, cannot accurately capture the trigger point during rapid fluctuations, often causes compensation delay, and uses fixed gear or linear proportional mode in compensation parameter setting, which fails to adjust the compensation of some nodes and other nodes based on the real-time running characteristics of the nodes, resulting in uneven power sharing, increasing the risk of bus voltage fluctuation and load impact. SUMMARY

[0005] The purpose of the present application is to solve the problems existing in the prior art and to provide a transient power compensation control method and system for ship direct current networking.

[0006] In order to achieve the above-mentioned purpose, the present application adopts the following technical scheme: a transient power compensation control method for ship direct current networking, comprising the following steps:

[0007] S1: Collect the resistance level and adjacency depth parameters between the bus nodes, read the resistance value and adjacency depth value at ten millisecond intervals through the power distribution bus terminal board, judge the valley value and screen the section, match the resistance threshold and connection depth critical value, complete the grouping of multiple nodes, and establish a topology partition set;

[0008] S2: Based on the topology partition set, the energy storage cabinet derivation module is used to obtain the voltage drop value and current impact value of each partition, a heuristic mining algorithm is used to arrange the drop value in ascending order and the impact value in descending order, a stable interval is located through bidirectional residual comparison, and a self-adaptive threshold matrix is generated accordingly;

[0009] S3: Based on the self-adaptive threshold matrix, the error signal amplitude and bus current difference value are obtained from the DC circuit breaker, a multi-layer feedforward neural network is used to select the gain gear by comparing the highest voltage threshold and the lowest current threshold, error adjustment is performed using the corresponding proportional coefficient to form a sliding mode compensation gain sequence;

[0010] S4: Based on the sliding mode compensation gain sequence, a control byte is written by combining the gain level code and the stage identification bit, a communication frame is constructed and frame checking and node address matching are performed, gain instruction sending is completed using the bus control channel and the response time is registered, and the control instruction log is output.

[0011] S5: Based on the control instruction log, the response timestamp and execution confirmation mark are extracted, the node response interval and compensation gain deviation are classified and summarized, the nodes with deviation exceeding twice the average value are screened, the gain correction level interval is divided and the update rule is calibrated, and the threshold fine-tuning parameter set is established.

[0012] As a further scheme of the application, the topology partition set includes DC bus partition number, node identification list in the partition, and node connection type label, the self-adaptive threshold matrix includes voltage trigger gear list of each partition, current impact gear list, and partition index mapping, the sliding mode compensation gain sequence includes response stage gain value list, convergence stage gain value list, and corresponding time step index, the control instruction log includes instruction sending timestamp list, gain level identification list, and node confirmation state list, and the threshold fine-tuning parameter set includes out-of-limit node number list, corresponding correction gear value list, and correction execution frequency statistics.

[0013] As a further scheme of the application, the specific steps for generating the topology partition set are as follows:

[0014] Based on the collected resistance level and adjacency depth parameters between bus nodes, the terminal board resistance values are continuously obtained by configuring the reader, the corresponding connection depth is extracted, the resistance and depth difference before and after is compared, and the fluctuation is filtered out to generate a node electrical parameter list.

[0015] Based on the node electrical parameter list, the resistance valley value and depth low value are screened through the threshold matching module, the node screening and classification operation is performed, the partition structure index is extracted and the node attributes are summarized, and the topology partition set is established.

[0016] As a further scheme of the present application, the specific steps for generating the adaptive threshold matrix are:

[0017] Based on the topological partition set, the energy storage cabinet derivation module is used to extract the voltage drop value and current impact value of each partition, construct a drop value sequence and an impact value sequence, extract the same period data according to the time index, arrange the voltage value in ascending order according to the amplitude size and mark the sorting index, arrange the current value in descending order according to the amplitude intensity and mark the corresponding index, and generate a voltage drop sequence.

[0018] Based on the voltage drop sequence, a heuristic mining algorithm is used to extract adjacent data pairs to calculate a drop difference sequence and adjacent impact data pairs to calculate an impact difference sequence, set a change tolerance range to filter data intervals that continuously satisfy a change amplitude less than a threshold value, superimpose the two types of difference intervals to extract an intersection section as a stability region, and generate a stable interval.

[0019] Based on the stable interval, the node index number contained therein is extracted and mapped to a partition structure table according to the node type, the drop value and impact value of all data sample points in the stable interval are extracted, the drop maximum value and impact minimum value are statistically merged according to the node number, a double-threshold corresponding table is constructed, and an adaptive threshold matrix is generated.

[0020] As a further scheme of the present application, the heuristic mining algorithm extracts the voltage drop sequence and current impact sequence in the same period according to the time index, arranges the voltage sequence in ascending order according to the amplitude and marks the sorting index, and arranges the current sequence in descending order according to the amplitude and marks the corresponding index; based on adjacent data pairs, the voltage drop difference sequence and the current impact difference sequence are calculated, the voltage change tolerance and the current change tolerance are set, the data intervals that continuously satisfy the change amplitude less than the tolerance are filtered, the voltage difference interval and the current difference interval that meet the conditions are intersected on the time axis to obtain a stability region, the samples in the stability region are mapped to a partition structure table according to the node index, the voltage drop maximum value and the current impact minimum value of each node are statistically merged, a double-threshold corresponding table is obtained, and an adaptive threshold matrix is generated.

[0021] As a further scheme of the present application, the specific steps for generating the sliding mode compensation gain sequence are:

[0022] Based on the adaptive threshold matrix, the current period error signal sample value of each node is read from the DC circuit breaker and an amplitude vector is constructed, the adjacent difference is extracted from the bus current sample sequence to construct a current difference vector, the voltage upper limit value and the current lower limit value corresponding to each node are extracted as a comparison standard, and an error signal set is generated.

[0023] Based on the error signal set, the error signal amplitude, the current difference, the voltage upper limit, the current lower limit and other parameters are extracted according to the node index to construct a feature vector, which is input into an offline trained multi-layer feedforward neural network model, and the gain level label corresponding to the node is output, a node gain level list is established, and a gain gear list is generated;

[0024] Based on the gain gear list, the current error signal original amplitude is read and the corresponding gain level multiplier value is found, the gain output value is calculated by performing node level multiplication operation, the multi-node gain output sequence is assembled in time sequence and combined into section identification, the power control action sequence is constructed, and the sliding mode compensation gain sequence is formed.

[0025] As a further scheme of the application, the multi-layer feedforward neural network constructs a feature vector according to the node and completes normalization according to the pre-stored mean value and scale parameter, and inputs the input layer, sequentially passes through multiple hidden layers, the hidden layer performs weight lookup table and multiplication and addition operation and applies ReLU activation, the weight and activation use 8-bit symmetric quantization to reduce operation delay, the gain level probability distribution is calculated in the output layer and the gain level label is output, the multiplier value is retrieved in the gain gear table according to the gain level label, the degradation strategy is triggered to use the conservative gain level when the probability is lower than the set threshold, the entire reasoning is completed in the control period and the gain label and confidence are written into the gain gear list.

[0026] As a further scheme of the application, the specific steps for generating the control instruction log are:

[0027] Based on the sliding mode compensation gain sequence, the gain level code is written into the gain level code through the mapping module, the identification bit is marked in the labeling stage, the binary structure combination is performed, the check field and synchronization bit are added, and the control byte sequence is generated;

[0028] Based on the control byte sequence, the communication frame is assembled through the frame constructor, the check code filling, address label matching and frame structure packaging are performed, the bus channel broadcast control byte is called, and the control instruction log is generated.

[0029] As a further scheme of the application, the specific steps for generating the threshold fine-tuning parameter set are:

[0030] Based on the control instruction log, the sending timestamp is extracted through the log parser, the confirmation mark is separated, the response interval calculation and node identification association are performed, the node response structure sequence is constructed, and the node response data set is generated;

[0031] Based on the node response data set, the response deviation over-limit node is identified through the deviation filter, the difference level is calculated by comparing the average value range, the gear is divided and the rule calibration is corrected, and the threshold fine-tuning parameter set is established.

[0032] A kind of transient power compensation control system of ship DC networking, the transient power compensation control system of ship DC networking is used to execute above-mentioned transient power compensation control method of ship DC networking, the system includes:

[0033] Bus topology module: based on power distribution bus terminal board acquisition ten millisecond interval node resistance value and adjacent depth value, compare adjacent sample point difference and mark valley value, filter section according to threshold value, group node according to DC bus segmentation number and generate partition structure table;

[0034] Threshold construction module: call heuristic mining algorithm, based on the partition structure table and energy storage cabinet data read voltage drop sequence and current shock sequence, align time index and sort, compare adjacent difference and tolerance filter stable segment, intersection merge segment and reflect node, count maximum drop value and minimum shock value, generate trigger gear set;

[0035] Gain generation module: based on the trigger gear set and DC circuit breaker data, adopt multilayer feedforward neural network, extract error signal amplitude and current difference, form feature group and compare threshold value to determine gain level, look up rate table and calculate output value, generate gain action sequence according to order combination;

[0036] Instruction issuing module: based on gain action sequence write gain level code and stage identification bit, form communication frame and broadcast in bus control channel, record sending and confirmation time, output control instruction log;

[0037] Threshold fine-tuning module: based on control instruction log extract response time and gain deviation, compare mean interval to filter out-of-limit node, update trigger gear and shock gear step according to level, establish fine-tuning parameter set.

[0038] Compared with prior art, the advantages and positive effects of the present application are that:

[0039] In the present application, the heuristic mining algorithm is used to compare the ascending and descending sequences and locate the stable interval, so that the threshold setting and trigger condition can be adaptively generated according to the partition operating state, reducing the false triggering and missed triggering caused by global fixed threshold value;

[0040] In the present application, the voltage upper limit, current lower limit, error amplitude and bus current difference are combined into a feature group, and the node gain level is quickly determined through multilayer feedforward neural network, ensuring that different nodes have compensation strength matching their transient fluctuation amplitude, reducing overcompensation and undercompensation phenomenon;

[0041] In the application, in the gain execution link, the compensation sequence formed is accurately issued through instruction coding, frame construction and address matching, combined with the record of sending and confirmation time, to form a traceable instruction link, improve the positioning efficiency of abnormal response, dynamically adjust the threshold step length by joint screening of node response interval and gain deviation, so that the threshold can be fine-tuned with the change of node performance and running state, and the bus operation stability and power distribution balance are maintained in the scene of load mutation, topology switching or multi-power cooperative power supply. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 It is a workflow diagram of the application;

[0043] Figure 2 It is a system flowchart of the application. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and not to limit the application. EMBODIMENT

[0045] Please refer to Figure 1 The application provides a technical scheme: a transient power compensation control method for ship DC networking, comprising the following steps:

[0046] S1: Collecting the resistance level and adjacent depth parameters between bus nodes, reading the resistance value and adjacent depth value at ten millisecond intervals through the power distribution bus terminal board, judging the valley value and screening the section, matching the resistance threshold and connection depth critical value, completing the grouping of multiple nodes, and establishing a topological partition set;

[0047] S2: Based on the topological partition set, using the energy storage cabinet export module to obtain the voltage drop value and current impact value of each partition, using a heuristic mining algorithm to respectively arrange the drop value in ascending order and the impact value in descending order, positioning the stable interval through bidirectional residual error comparison, and generating an adaptive threshold matrix accordingly;

[0048] S3: Based on the adaptive threshold matrix, obtaining the error signal amplitude and bus current difference value from the DC circuit breaker, using a multi-layer feedforward neural network, comparing the voltage highest threshold and the current lowest threshold to select the gain position in stages, applying the corresponding proportional coefficient to adjust the error, and forming a sliding mode compensation gain sequence;

[0049] S4: Based on the sliding mode compensation gain sequence, writing gain level code and stage identification bit to construct control bytes, constructing communication frames and performing frame checking and node address matching, using the bus control channel to complete gain instruction sending and register response time, and outputting control instruction log;

[0050] S5: Based on the control instruction log, extract the response timestamp and execution confirmation mark, classify and summarize the node response interval and compensation gain deviation, screen the deviation exceeding twice the average value node, divide the gain correction level interval and mark the update rule, and establish the threshold fine-tuning parameter set.

[0051] The topology partition set includes DC bus partition number, partition node identification list, node connection type label, adaptive threshold matrix including partition voltage trigger gear list, current impact gear list, partition index mapping, sliding mode compensation gain sequence including response stage gain value list, convergence stage gain value list, corresponding time step index, control instruction log including instruction sending timestamp list, gain level identification list, node confirmation state list, threshold fine-tuning parameter set including out-of-limit node number list, corresponding correction gear value list, correction execution frequency statistics.

[0052] The specific steps for generating the topology partition set are as follows:

[0053] Based on the collected resistance level and adjacent depth parameters between bus nodes, the terminal board resistance values are continuously obtained by configuring the reader, the corresponding connection depth is extracted, the resistance and depth difference is compared, and the fluctuation is filtered to generate the node electrical parameter list;

[0054] Based on the node electrical parameter list, the resistance valley value and depth low value are screened by the threshold matching module, the node screening and classification operation is performed, the partition structure index is extracted and the node attribute is summarized, and the topology partition set is established;

[0055] Based on the collected resistance level and adjacent depth parameters between bus nodes, Hampel filter window is used to perform outlier rejection, window width is 9, threshold is three times the standard deviation, center statistics is median, then Savitzky-Golay filter is used to perform fluctuation filtering, window length is 11, fitting order is 3, end point processing method is mirror expansion, terminal board resistance values and connection depth are continuously obtained by the READ_RES reading instruction in channel A with a sampling period of ten milliseconds, first-order difference is used to generate a sequence of before and after difference values and determine the candidate valley point by sign flip, PELT variable point detection is called to set the penalty weight to five times the standard deviation and the minimum segment length to 50 to locate the start and end index of the section, resistance difference and depth difference are executed threshold clipping, threshold upper limit is 1 ohm, lower limit is 0 ohm, denoising sequence is generated, adjacent repeated value is removed according to the first item, and the fields Rmean, Rmin, Rmax and Dcnt are aggregated to generate the node electrical parameter list;

[0056] Based on the node electrical parameter list, the threshold matching process is used to perform interval judgment on Rmin and connection depth, the resistance valley threshold interval is 0.05 to 0.5 ohms, the depth low value threshold set is 1, 2, 3, the local minimum value search command PEAKS is used to set the minimum significance to 0.02 ohms and the minimum spacing to 5 to mark the valley point index, the union set is called and the parent pointer is initialized to itself, the FIND and UNION operations are performed, the node set is merged according to the condition that the adjacent depth difference is less than 1 and the resistance valley point is in the same segment, the partition structure index is generated by multiplying the BUS number by 100 and adding the set number, the node ID, connection type label and Rmean, Rmin, Rmax and Dcnt fields are summarized for each set, the partition structure index is sorted in ascending order and written into the index mapping table, the exclusion rule of set size less than 2 is called, the index mapping table and node attribute list are written into the partition entry sequence, and the topological partition set is established.

[0057] The specific steps of generating the adaptive threshold matrix are:

[0058] Based on the topological partition set, the energy storage cabinet export module is used to extract the voltage drop value and current impact value of each partition, the drop value sequence and impact value sequence are constructed, the same period data is extracted according to the time index, the voltage value is sorted in ascending order according to the amplitude size and marked with the sorting index, the current value is sorted in descending order according to the amplitude intensity and marked with the corresponding index, and the voltage drop sequence is generated;

[0059] Based on the voltage drop sequence, the heuristic mining algorithm is used to extract the adjacent data pair to calculate the drop difference value sequence and the adjacent impact data pair to calculate the impact difference value sequence, the change tolerance range is set to filter the data interval that continuously satisfies the change amplitude less than the threshold value, the intersection section is extracted as the stability region after superimposing the two types of difference interval, and the stable interval is generated;

[0060] Based on the stable interval, the node index number contained is extracted and mapped to the partition structure table according to the node type, the drop value and impact value of all data sample points in the stable interval are extracted, the drop maximum value and impact minimum value are merged and counted according to the node number, the double threshold value corresponding table is constructed, and the adaptive threshold matrix is generated;

[0061] Based on the topological partition set, the energy storage cabinet export module calls the EXPORT interface channel V and channel I to read the partition voltage drop value and current impact value, sets the sampling period to ten milliseconds and the cycle length to two hundred milliseconds, uses ALIGN_CYCLE to align the same cycle data by start and end time and fills the empty position with the missing measurement zero strategy, uses SORT_ASC to sort the voltage amplitude in ascending order and writes it into the fields rank_v and pos_v, uses SORT_DESC to sort the current amplitude in descending order and writes it into the fields rank_i and pos_i, calls TAG_INDEX to write the original time index and partition identification into the record header, concatenates the record stream according to the partition and time order and writes the file block serial number and checksum, and generates the voltage drop sequence;

[0062] Based on the voltage drop sequence, a heuristic mining algorithm is used to execute the DIFF1 difference command on adjacent samples to generate a drop difference value sequence, and the DIFF1 difference command is executed on the current impact sequence to generate an impact difference value sequence. The tolerance threshold tau_v is set to 0.3% of the rated voltage and tau_i is set to 0.5% of the rated current. The SLIDE window length is set to five and the step length is set to one. The two types of difference value sequences are compared window by window and the continuous fragments that meet the threshold condition are marked. The INTERSECT is executed to superimpose the two types of fragments according to the time index, and the MERGE is called to merge the fragments with an interval of not more than thirty milliseconds. The start and end times of the fragments, the number of samples contained, the partition number and the list of participating nodes are written, and the stable interval is output.

[0063] Based on the stable interval, the node index in the fragment is remapped according to the node type in the partition structure table, and the corresponding voltage drop sample and current impact sample are pulled. The GROUP_BY command is used to merge the records according to the node number and calculate the maximum drop value and the minimum impact value. The v_drop_max and i_surge_min fields are written, as well as the node number and partition number. A double-threshold correspondence table is established. The CLASSIFY command is used to map v_drop_max to the trigger gear boundary sequence 2, 4, 6 and 8, and i_surge_min to the impact gear boundary sequence 5, 10, 15 and 20. A mapping relationship table of partition index, gear index and node index is generated, and is summarized as an adaptive threshold matrix.

[0064] The heuristic mining algorithm extracts voltage drop and current surge sequences within the same period by time index. The voltage sequences are sorted in ascending order of amplitude and labeled with sorting index, while the current sequences are sorted in descending order of amplitude and labeled with corresponding index. Based on adjacent data pairs, the voltage drop difference sequence and current surge difference sequence are calculated. Voltage change tolerance and current change tolerance are set, and data intervals that continuously meet the condition of change amplitude being less than the tolerance are selected. The intersection of the voltage difference interval and current difference interval that meet the condition is found on the time axis to obtain the stability region. The samples in the stability region are mapped to the partition structure table by node index. The maximum voltage drop value and the minimum current surge value of each node are counted to obtain a dual threshold correspondence table and generate an adaptive threshold matrix.

[0065] Heuristic mining algorithms, according to the formula:

[0066] ;

[0067] in: Indicates the first Voltage sag of the DC bus within each control cycle Indicates the first The instantaneous value of the bus voltage obtained by sampling in each control cycle Indicates the first The instantaneous value of the bus voltage obtained by sampling in each control cycle This represents the bus voltage disturbance sensitivity adjustment coefficient. This represents the factor for estimating the rate of change of bus voltage. This indicates the set voltage fluctuation tolerance threshold. This represents the transient fluctuation suppression penalty coefficient. Indicates the priority weight of node voltage control;

[0068] Execution process: First, the bus voltage value of the ship's DC network in each control cycle is collected. and The voltage difference is calculated to assess transient voltage changes. Then, to more accurately reflect the impact of voltage fluctuations on system stability, several adjustment factors are introduced into the formula. This determines the strength of the system's response to voltage changes. Used to reflect the rate of voltage change. Used to determine whether the voltage variation exceeds the acceptable range, by The product of these factors allows for dynamic adjustment of the tolerance to voltage variations, enabling more precise power compensation. To further reduce the impact of voltage fluctuations on the system, a penalty term is introduced. ,in It is a punishment factor. The voltage control priority weight coefficient is used to suppress excessive or rapid voltage fluctuation, prevent the voltage fluctuation from causing a large transient impact on the system, ensure the stability and reliability of the ship DC networking in transient power fluctuation, and ultimately ensure the voltage stability of the DC bus and optimize the power scheduling through the combination of the two adjustment and correction terms.

[0069] The specific steps for generating the sliding mode compensation gain sequence are as follows:

[0070] Based on the adaptive threshold matrix, the error signal sampling values of each node in the current period are read from the DC circuit breaker and an amplitude vector is constructed, the adjacent difference is extracted from the bus current sampling sequence to construct a current difference vector, the voltage upper limit value and the current lower limit value corresponding to each node are extracted as comparison standards to generate an error signal set;

[0071] Based on the error signal set, the error signal amplitude, current difference, voltage upper limit and current lower limit parameters are extracted according to the node index to construct a feature vector, which is input into the offline trained multi-layer feedforward neural network model to output the gain grade label of the corresponding node, a node gain grade list is established, and a gain gear list is generated;

[0072] Based on the gain gear list, the original amplitude of the current error signal is read and the corresponding gain grade multiplier value is found, the node-level multiplication operation is performed to calculate the gain output value, the multi-node gain output sequence is assembled in time sequence and combined into a section identifier, a power control action sequence is constructed, and a sliding mode compensation gain sequence is formed;

[0073] Based on the adaptive threshold matrix, the READ_ERR instruction is executed on the DC circuit breaker channel E to read the error signal sampling values of each node with a 10 millisecond sampling period and write them into the array err_raw, the ABS command is used to calculate the amplitude of err_raw and generate the vector err_amp, the READ_CUR instruction is executed on the BUS channel I to read the bus current sampling sequence cur_raw and the DIFF1 command is called to extract the adjacent difference to generate the vector cur_diff, the CLIP command is called to perform upper and lower limit clipping on err_amp and cur_diff with upper and lower limits of zero and unit one respectively, the voltage upper limit V_up and the current lower limit I_low are extracted from the adaptive threshold matrix according to the node index and combined with err_amp and cur_diff to form a record item node_id and err_amp and cur_diff and V_up and I_low, the record items are merged according to the time index and written into a list to generate an error signal set;

[0074] Based on the error signal set, the four parameters of err_amp, cur_diff, V_up and I_low are extracted according to the node index to construct a length four feature vector, MODEL.LOAD is called to load the inference instance MFFN_001, the input quantization bit width is set to eight bits, the activation function is set to ReLU, the output is set to Softmax, and the threshold judgment p_thr is set to zero point six, MODEL.RUN is executed with an input batch size of one and an output gain level label L, and the confidence p is recorded, WRITE_LIST is called to write node_id, L and p in chronological order, TABLE_MAP is called to generate a corresponding rate index for L and write it into the entry, and each node entry is merged into a row sequence to generate a gain position list.

[0075] Based on the gain position list, the original amplitude err_now of the current period err_raw is read according to the node index, and GAIN_TABLE is called to find the corresponding rate value K for L, MUL command is executed to calculate the node level gain output value gain_out as the product of err_now and K, and each node gain_out is spliced in chronological order and inserted into the partition boundary identifier and time step index, SEGMENT command is called to generate paragraph numbers with a window of twenty milliseconds and write them into the paragraph head, and combined into a power control action sequence, written into the sequence buffer and node address mapping table in control cycle order, forming a sliding mode compensation gain sequence.

[0076] A multi-layer feedforward neural network constructs a feature vector according to the node and completes normalization according to the pre-stored mean and scale parameters, and inputs the input layer, and then passes through multiple hidden layers in turn, the hidden layer performs weight lookup table and multiplication and addition operation and applies ReLU activation, the weight and activation are 8-bit symmetric quantization to reduce operation delay, the gain level probability distribution is calculated in the output layer and the gain level label is output, the rate value is retrieved in the gain position table according to the gain level label, and when the probability is lower than the set threshold, the conservative gain level is triggered to trigger the degradation strategy, the whole inference is completed in the control period and the gain label and confidence are written into the gain position list;

[0077] The multi-layer feedforward neural network is according to the formula:

[0078] ;

[0079] Among them: L(n) represents the gain level label of the n-th node in the ship DC networking, L(n) represents the gain level label of the n-th node in the ship DC networking, L(n) represents the gain level label of the n-th node in the ship DC networking, L(n) represents the gain level label of the n-th node in the ship DC networking, L(n) represents the gain level label of the n-th node in the ship DC networking, L(n) represents the gain level label of the n-th node in the ship DC networking, L(n) represents the gain level label of the n-th node in the ship DC networking, L(n) represents the gain level label of the n-th node in the ship DC networking, a bias term representing an output offset adjustment in the neural network, a weighting coefficient representing an error signal amplitude, an error signal amplitude of the first node, a weighting coefficient representing a current difference, a current difference between the first node and its adjacent node, a suppression coefficient representing a voltage upper limit, a voltage upper limit setting value of the first node, a compensation coefficient representing a current lower limit, a current lower limit setting value of the first node;

[0080] The execution process is as follows: first, for all monitored nodes in the ship DC power distribution system, the running state data at the current time is collected to construct a feature vector group, each node corresponds to a group of feature parameters, including power load, voltage fluctuation frequency, transient current change in the original feature vector, and additional parameters such as error signal amplitude , current difference , voltage upper limit and current lower limit , which respectively reflect the state offset, local load fluctuation, operation upper limit constraint and low power support capability of the current node, the importance of which is adjusted by introducing weight coefficients , , , , then all parameters are input into the multi-layer feedforward neural network which has been trained offline, the neural network structure provides basic feature combination capability through weight matrix and bias term , and then superimposes various electrical feature modulation terms with engineering physical significance to comprehensively reflect the demand degree of the node gain under transient environment, and then through a nonlinear activation function mapping to gain level label , the final output result is used to establish a node gain level list, and the gain gear list of the whole system is generated according to the grade division, which is used to control the power distribution behavior of each energy router or compensation device, so as to realize the rapid compensation and voltage stability control of the whole ship DC networking under transient disturbance conditions.

[0081] The specific steps of generating the control instruction log are as follows:

[0082] Based on the sliding mode compensation gain sequence, write the gain level code through the mapping module, identify the stage in the marking phase, execute the binary structure combination, add the check field and the synchronization bit, and generate the control byte sequence;

[0083] Based on the control byte sequence, assemble the communication frame through the frame constructor, execute the check code filling, address label matching and frame structure packaging, call the bus channel broadcast control byte, and generate the control instruction log;

[0084] Based on the sliding mode compensation gain sequence, map the gain level L to the two-bit gain level code table 00 01 10 11 using the mapping module, identify the bit number 5 in the marking phase, call the bit write command BIT_SET to write the gain level code and the stage identification bit in the target byte, write the field identification code one byte, the time step index two bytes, the node address one byte and the gain rate index one byte in sequence using the byte packing rule PACK8, calculate the check value two bytes of the foregoing byte sequence after executing the binary structure combination, write the synchronization byte sequence 55 AA as the synchronization bit at the frame head and the frame middle separation position, and perform the escape coding command ESC to insert the escape byte 7D and the exclusive or value 20 for the same value in the data byte and the synchronization byte, merge the fields to form a continuous byte stream and write it into the buffer area, and generate the control byte sequence;

[0085] Based on the control byte sequence, use the frame constructor to execute the HDLC frame packaging to set the frame start byte 7E and the frame end byte 7E, write one byte of the target address in the address table according to the node address and the partition address priority, write the payload length one byte using the length calculation command LEN_CALC, write the check value two bytes at the frame tail position using the check code filling, perform the SLIP escape for the 7E and 7D and 55 and AA bytes in the frame, set the baud rate 115200, the data bit 8, the check bit none, the stop bit 1 and the timeout 50 milliseconds using the bus channel sending command BUS_TX to broadcast the control byte, set the window one hundred milliseconds to receive the confirmation frame using the receiving command BUS_RX, and record the sending timestamp and the confirmation timestamp, the node address, the frame number and the check result, write them in the log recording area in time sequence and persist them to the file segment, and generate the control instruction log.

[0086] The specific steps of generating the threshold fine-tuning parameter set are:

[0087] Based on the control instruction log, extract the sending timestamp through the log parser, separate the confirmation mark, execute the response interval calculation and node identification association, construct the node response structure sequence, and generate the node response data set;

[0088] Based on the node response data set, the response deviation exceeds the limit node is identified by the deviation filter, the difference level is calculated by comparing with the average value range, the level is divided and the rule calibration is corrected, and the threshold fine-tuning parameter set is established;

[0089] Based on the control instruction log, the log parser LOG_PARSE is called to execute the LOAD_FILE command to load the log file into the memory buffer and split it into lines according to the record separator 0x0A, the SPLIT command is used to extract the sending timestamp field and the confirmation marker field according to the field separator 0x2C and convert them into UNIX time format with second-level precision, the MATCH_ID command is called to associate the confirmation marker with the node identification field, the SUB command is executed to subtract the sending timestamp from the confirmation timestamp to obtain the response interval value and write it into the field resp_gap, the GROUP_BY command is called to group by node identification to generate an ordered sequence containing node ID, resp_gap and confirmation marker, the node response structure sequence is generated, the node response structure sequence is written into the memory object and serialized as a binary file, and the node response data set is outputted;

[0090] Based on the node response data set, the deviation filter DEVIATION_SCAN is called to execute the MEAN_STD command to calculate the mean value mean_gap and the standard deviation std_gap of each node response interval, the GT command is used to filter the node ID marked as exceeding the limit node when resp_gap is greater than mean_gap plus twice std_gap, the LOOKUP command is called to find the response interval value and gain execution rate of the corresponding node in the node response data set, the SUB command is executed to calculate the difference between the gain execution rate and the reference rate to obtain the difference level value diff_lvl, the RANGE_CLASSIFY command is called to map diff_lvl to the level interval table and generate the level number, the RULE_SET command is called to write the correction rule parameter fields step_val and update_count according to the level number, the MAP command is executed to generate the mapping table of node ID and step_val and update_count, the parameter set file is written and the threshold fine-tuning parameter set is outputted.

[0091] Please refer to Figure 2 A transient power compensation control system for ship DC networking, the transient power compensation control system for ship DC networking is used to execute the above-mentioned transient power compensation control method for ship DC networking, the system comprises:

[0092] Bus topology module: based on the power distribution bus terminal plate, the resistance value and the adjacent depth value between nodes are collected at ten millisecond intervals, the difference value between adjacent sampling points is compared and the valley value is marked, the section is screened according to the threshold, the nodes are grouped according to the DC bus section number and the partition structure table is generated;

[0093] Threshold construction module: call heuristic mining algorithm, based on partition structure table and energy storage cabinet data reading voltage drop sequence and current shock sequence, align time index and sort, compare adjacent difference and tolerance filter stable fragments, intersection and fragment mapping nodes, statistics maximum drop value and minimum shock value, trigger gear set generation;

[0094] Gain generation module: based on trigger gear set and DC circuit breaker data, using multi-layer feedforward neural network, extract error signal amplitude and current difference, form feature group and compare threshold to determine gain level, check multiplier table and calculate output value, generate gain effect sequence in order combination;

[0095] Instruction issuing module: based on gain effect sequence write gain level code and stage identification bit, form communication frame and broadcast in bus control channel, record sending and confirmation time, output control instruction log;

[0096] Threshold fine-tuning module: based on control instruction log to extract response time and gain deviation, compare mean interval to filter out-of-limit nodes, update trigger gear and shock gear step by level, establish fine-tuning parameter set.

[0097] The above is only the preferred embodiment of the present application, not other forms of the present application is limited, any skilled in the art of the technical personnel may use the disclosed technical content to change or modify as equivalent changes equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification of the above embodiments without departing from the technical scheme content of the present application, according to the technical essence of the present application, still belongs to the protection scope of the technical scheme of the present application.

Claims

1. A transient power compensation control method for a ship's DC network, characterized in that, Includes the following steps: S1: Collect the resistance level and adjacency depth parameters between bus nodes, read the resistance value and adjacency depth value at ten-millisecond intervals through the distribution bus terminal board, determine the valley value and filter the segment, match the resistance threshold and connection depth critical value, complete the grouping of multiple nodes, and establish a topology partition set. S2: Based on the topological partition set, the voltage drop value and current surge value of each partition are obtained by the energy storage cabinet export module. The heuristic mining algorithm is used to sort the drop value in ascending order and the surge value in descending order respectively. The stable interval is located by bidirectional residual comparison, and an adaptive threshold matrix is ​​generated accordingly. S3: Based on the adaptive threshold matrix, obtain the error signal amplitude and the bus current difference value from the DC circuit breaker, use a multi-layer feedforward neural network, select the gain level according to the highest voltage threshold and the lowest current threshold, apply the corresponding proportional coefficient to adjust the error, and form a sliding mode compensation gain sequence. S4: Based on the sliding mode compensation gain sequence, write the gain level code and stage identifier bit to construct control bytes, construct a communication frame and perform frame verification and node address matching, use the bus control channel to complete the gain command transmission and register the response time, and output the control command log; S5: Based on the control command log, extract the response timestamp and execution confirmation mark, classify and summarize the node response interval and compensation gain deviation, filter nodes with deviations exceeding twice the average value, divide the gain correction level range and calibrate the update rules, and establish a threshold fine-tuning parameter set. The specific steps for generating the sliding mode compensation gain sequence are as follows: Based on the adaptive threshold matrix, the error signal sampling values ​​of each node in the current cycle are read from the DC circuit breaker and an amplitude vector is constructed. Adjacent differences are extracted from the bus current sampling sequence and a current difference vector is constructed. The upper voltage limit and lower current limit corresponding to each node are extracted as comparison standards to generate an error signal set. Based on the error signal set, parameters such as error signal amplitude, current difference, upper voltage limit, and lower current limit are extracted according to the node index to construct a feature vector, which is then input into the offline-trained multilayer feedforward neural network model. The corresponding node gain level label is output, a node gain level list is established, and a gain level list is generated. Based on the gain level list, the corresponding gain ratio information is determined for each node. The node gain result is generated according to the mapping relationship between the error signal and the ratio. The gain results of all nodes are arranged in order of sampling time and labeled with their respective segment numbers to obtain the gain action sequence for power control. The gain action sequence is the sliding mode compensation gain sequence.

2. The transient power compensation control method for ship DC grids according to claim 1, characterized in that, The topology partition set includes a DC bus partition number, a list of node identifiers within the partition, and node connection type labels. The adaptive threshold matrix includes a list of voltage trigger levels, a list of current impact levels, and a partition index mapping for each partition. The sliding mode compensation gain sequence includes a list of gain values ​​during the response phase, a list of gain values ​​during the convergence phase, and a corresponding time step index. The control command log includes a list of command sending timestamps, a list of gain level identifiers, and a list of node confirmation statuses. The threshold fine-tuning parameter set includes a list of out-of-limit node numbers, a list of corresponding correction levels, and a count of correction execution times.

3. The transient power compensation control method for ship DC grids according to claim 1, characterized in that, The specific steps for generating the topology partition set are as follows: Based on the collected bus node resistance level and adjacency depth parameters, the terminal block resistance value is continuously acquired by the configured reader, the corresponding connection depth is extracted, the difference between the resistance and depth is compared before and after, and fluctuations are filtered out to generate a list of node electrical parameters. Based on the list of node electrical parameters, the threshold matching module filters out resistance valleys and low depth values, performs node filtering and classification operations, extracts the partition structure index and summarizes node attributes, and establishes a topology partition set.

4. The transient power compensation control method for ship DC grids according to claim 1, characterized in that, The specific steps for generating the adaptive threshold matrix are as follows: Based on the topology partition set, the voltage drop value and current surge value of each partition are extracted by the energy storage cabinet export module. The drop value sequence and the surge value sequence are constructed. Data of the same period are extracted by time index. The voltage values ​​are sorted in ascending order by amplitude and the sorting index is marked. The current values ​​are sorted in descending order by amplitude intensity and the corresponding index is marked to generate the voltage drop sequence. Based on the voltage drop sequence, a heuristic mining algorithm is used to extract adjacent data pairs to calculate the drop difference sequence and extract adjacent impact data pairs to calculate the impact difference sequence. A change tolerance range is set to filter data intervals that continuously meet the change amplitude less than the threshold. After superimposing the two types of difference intervals, the intersection segment is extracted as the stability region to generate a stable interval. Based on the stable interval, the index numbers of the included nodes are extracted and mapped to the partition structure table according to the node type. The drop value and impact value of all data sample points in the stable interval are extracted. The maximum drop value and minimum impact value are merged and counted according to the node number. A dual threshold correspondence table is constructed to generate an adaptive threshold matrix.

5. The transient power compensation control method for ship DC grids according to claim 4, characterized in that, The heuristic mining algorithm extracts voltage drop sequences and current surge sequences within the same period by time index. It sorts the voltage sequences in ascending order of amplitude and marks them with sorting indices, and sorts the current sequences in descending order of amplitude and marks them with corresponding indices. Based on adjacent data pairs, it calculates voltage drop difference sequences and current surge difference sequences, sets voltage change tolerances and current change tolerances, and filters data intervals that continuously satisfy the condition of a change amplitude less than the tolerances. It then finds the intersection of the voltage difference intervals and current difference intervals that meet the conditions on the time axis to obtain a stability region. The samples within the stability region are mapped to a partitioned structure table by node index. The maximum voltage drop value and the minimum current surge value of each node are statistically analyzed to obtain a dual-threshold correspondence table, and an adaptive threshold matrix is ​​generated.

6. The transient power compensation control method for ship DC grids according to claim 1, characterized in that, The multilayer feedforward neural network constructs feature vectors for each node and normalizes them according to pre-stored mean and scale parameters, then inputs them into the input layer. The vectors then pass through multiple hidden layers, where weight lookup and multiply-accumulate operations are performed, and ReLU activation is applied. The weights and activations use 8-bit symmetric quantization to reduce computational latency. The output layer calculates the probability distribution of gain levels and outputs gain level labels. Based on the gain level labels, the gain value is retrieved from the gain level table. When the probability is lower than a set threshold, a downgrade strategy is triggered to adopt a conservative gain level. The entire inference is completed within the control cycle, and the gain labels and confidence scores are written into the gain level list.

7. The transient power compensation control method for ship DC grids according to claim 1, characterized in that, The specific steps for generating the control command log are as follows: Based on the sliding mode compensation gain sequence, the gain level code is written through the mapping module, the stage identification bit is marked, the binary structure combination is executed, the check field and the synchronization bit are added, and the control byte sequence is generated. Based on the control byte sequence, a communication frame is assembled by the frame constructor, and checksum filling, address tag matching and frame structure encapsulation are performed. The bus channel is then called to broadcast control bytes and generate a control command log.

8. The transient power compensation control method for ship DC grids according to claim 1, characterized in that, The specific steps for generating the threshold fine-tuning parameter set are as follows: Based on the control command log, the sending timestamp is extracted by the log parser, the confirmation flag is separated, the response interval is calculated and associated with the node identifier, a node response structure sequence is constructed, and a node response dataset is generated. Based on the node response dataset, nodes with excessive response deviations are identified by a deviation filter, the difference level is calculated by comparing with the average range, the levels are divided and the rules are corrected, and a threshold fine-tuning parameter set is established.

9. A transient power compensation control system for a ship's DC network, characterized in that, The transient power compensation control method for ship DC grids according to any one of claims 1-8, wherein the system comprises: Bus topology module: Based on the power distribution bus terminal board, it collects the resistance value and adjacency depth value between nodes at ten-millisecond intervals, compares the difference between adjacent sampling points and marks the valley value, filters the segments according to the threshold, groups the nodes according to the DC bus segment number and generates a partition structure table; Threshold construction module: Calls a heuristic mining algorithm to read voltage drop sequences and current surge sequences based on the partition structure table and energy storage cabinet data, aligns and sorts the time index, compares adjacent differences and tolerances to filter stable segments, merges segments and maps them back to the mapping node, counts the maximum drop value and minimum surge value, and generates a set of trigger gears; Gain generation module: Based on the trigger position set and DC circuit breaker data, a multi-layer feedforward neural network is used to extract the error signal amplitude and current difference, form a feature group and compare the threshold to determine the gain level, look up the multiplier table and calculate the output value, and generate the gain action sequence in sequence. Command issuance module: Writes gain level code and stage identifier bit based on gain action sequence, forms communication frame and broadcasts it on bus control channel, records sending and confirmation time, and outputs control command log; Threshold fine-tuning module: Extract response time and gain deviation based on control command log, compare mean range to filter out over-limit nodes, update trigger gear and impact gear step size according to level, and establish fine-tuning parameter set.

Images