Dynamic optimization method for ship intelligent power distribution system based on dynamic perception of ship load

Abstract

The present application relates to the technical field of data processing, and more particularly to a ship intelligent power distribution system dynamic optimization method based on ship load dynamic sensing; a sliding analysis window is obtained according to the spectrum data of the load influence data sequence; state characteristic values and state characteristic reference values are obtained according to the data change characteristics in the sliding analysis window; a deviation degree is obtained according to the state characteristic values and the state characteristic reference values; clustering is performed according to the deviation degree, and deviation boundary values and normal distribution proportions are obtained according to the deviation degree in the window cluster; a new normal distribution proportion is obtained according to the state characteristic values of a new sliding analysis window in the future; and a difference coefficient is obtained according to the new normal distribution proportion and the normal distribution proportion. The present application adjusts the initial learning rate step in the LSTM prediction model according to the difference coefficient, obtains an adaptive LSTM prediction model, and predicts the ship load state, so as to optimize power distribution according to the ship load state, and improve the load prediction accuracy and the power distribution optimization accuracy.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a dynamic optimization method for a ship intelligent power distribution system based on dynamic perception of ship load. Background Technology

[0002] During ship navigation, the electrical system has become the core power system ensuring navigation and operations, and its stability and reliability directly affect the safety and efficiency of the ship. Existing ship power distribution systems are mostly designed based on static load models and fixed operating conditions, making it difficult to adapt to the rapid fluctuations in load demand caused by changes in ship speed and other factors during actual operation. When load demand exceeds system supply, it can easily lead to problems such as decreased system stability, affecting the safe operation of the ship. Conversely, when load demand is low, it can easily result in wasted excess power supply. Especially in the complex and ever-changing marine environment, the load demand of ships exhibits significant dynamism and uncertainty.

[0003] During actual ship navigation, electrical load is affected by multiple factors such as changes in navigation speed, propeller operating conditions, hull attitude fluctuations, and sea state changes, resulting in rapidly changing and highly sudden load demands. Existing load forecasting methods based on BP neural networks or, more preferably, LSTM neural networks mainly rely on historical load data or modeling of single electrical quantities. Furthermore, the parameters are fixed after model training, making it difficult to fully consider real-time operating conditions and changes in the external environment. When newly added data differs significantly from historical patterns, using fixed model parameters for prediction can easily lead to increased deviations in results, affecting the accuracy of load forecasting. Consequently, intelligent power distribution systems cannot quickly and rationally perform load scheduling, reducing the safety and operational efficiency of ship operations.

[0004] Therefore, improving the accuracy of electrical load prediction during ship operation has become an urgent problem to be solved. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a dynamic optimization method for a ship intelligent power distribution system based on dynamic perception of ship load, in order to solve the problem of how to improve the accuracy of power load prediction during ship operation; the technical solution adopted is as follows:

[0006] Obtain load impact data sequences from different dimensions during ship operation;

[0007] A sliding analysis window is obtained based on the distribution characteristics of the spectral data of the load impact data sequence; different types of state characteristic values ​​and state characteristic reference values ​​are obtained based on the data change characteristics within the sliding analysis window of the load impact data sequence; the degree of deviation is obtained based on the difference between the state characteristic values ​​corresponding to the sliding analysis window and the state characteristic reference values.

[0008] Clustering is performed based on the difference in the degree of deviation between any two sliding analysis windows to obtain window clusters; critical representativeness is obtained based on the distribution characteristics of the degree of deviation within the window clusters; the deviation boundary value of the degree of deviation and the normal distribution ratio of the load impact data sequence are obtained based on the critical representativeness; a new normal distribution ratio is obtained based on the state characteristic value of the future new sliding analysis window; and a difference coefficient is obtained based on the difference between the new normal distribution ratio and the normal distribution ratio.

[0009] The initial learning rate step size of the load impact data sequence in the LSTM prediction model is adjusted according to different types of difference coefficients to obtain an adaptive LSTM prediction model; the ship load status is predicted according to the adaptive LSTM prediction model, and the power distribution is optimized according to the ship load status.

[0010] Further, the step of obtaining the sliding analysis window based on the distribution characteristics of the spectral data of the load impact data sequence includes:

[0011] The spectral data of the load impact data sequence is obtained by short-time Fourier transform; different amplitudes in the spectral data are used as weights of corresponding frequencies, and the weighted average of all frequencies in the spectral data is calculated to obtain frequency characteristic values; the reciprocal of the frequency characteristic values ​​is calculated and rounded to obtain the length of the window; half of the window length is used as the window sliding step size, and the window is slid across the load impact data sequence, obtaining a sliding analysis window with each slide.

[0012] Furthermore, the step of obtaining different types of state characteristic values ​​and state characteristic reference values ​​based on the data change characteristics within the sliding analysis window of the load impact data sequence includes:

[0013] The state characteristic values ​​include trend characteristic values ​​and degree of change characteristic values; the average value of the first difference of the data within the sliding analysis window is calculated to obtain the trend characteristic value; the average value of the second difference of the data within the sliding analysis window is calculated to obtain the degree of change characteristic value.

[0014] The average of the absolute differences between the state feature values ​​of any sliding analysis window and all other sliding analysis windows is calculated and negatively correlated to obtain the confidence level of the arbitrary sliding analysis window; the state feature value of the sliding analysis window corresponding to the maximum confidence level is used as the state feature reference value.

[0015] Furthermore, the step of obtaining the degree of deviation based on the difference between the state feature value corresponding to the sliding analysis window and the state feature reference value includes:

[0016] The degree of deviation is obtained by calculating the absolute value of the difference between the state feature value of the sliding analysis window and the state feature reference value of the same type.

[0017] Furthermore, the step of clustering based on the difference in the degree of deviation between any two sliding analysis windows to obtain window clusters includes:

[0018] In the formula This represents the probability of mutual selection between the i-th and j-th sliding analysis windows; Represents a linear normalization function. This indicates the degree of deviation in the i-th sliding analysis window. The deviation degree of the j-th sliding analysis window is represented; the mutual selection probability value is used as the similarity between sliding analysis windows in the K-means clustering algorithm to obtain different window clusters.

[0019] Further, the step of obtaining the critical representativeness based on the distribution characteristics of the corresponding deviation degree within the window cluster includes:

[0020] Calculate the average absolute value of the difference in the degree of deviation between any two sliding analysis windows within the window cluster and perform a negative correlation mapping to obtain the degree of concentration; calculate the average degree of deviation of all sliding analysis windows within the window cluster to obtain the average degree of deviation; calculate the product of the degree of concentration and the average degree of deviation to obtain the critical representativeness of the window cluster.

[0021] Furthermore, the step of obtaining the deviation boundary value of the deviation degree and the normal distribution ratio of the load-affected data sequence based on the critical representativeness includes:

[0022] Calculate the average value of the state characteristic values ​​of the sliding analysis window within the window cluster corresponding to the maximum value of the critical representativeness to obtain the deviation boundary value; calculate the ratio of the number of state characteristic values ​​in the load influence data sequence whose deviation degree is less than the deviation boundary value to the total number of state characteristic values ​​to obtain the normal distribution ratio.

[0023] Further, the step of obtaining the difference coefficient based on the difference characteristics between the new normal distribution ratio and the normal distribution ratio includes:

[0024] The difference between the new normal distribution ratio and the normal distribution ratio is calculated to obtain the difference coefficient.

[0025] Furthermore, the step of adjusting the initial learning rate step size of the load impact data sequence in the LSTM prediction model according to different types of difference coefficients to obtain an adaptive LSTM prediction model includes:

[0026] In the formula, X represents the adaptive learning rate step size. This indicates taking the minimum value. This indicates the preset maximum learning rate step size. This indicates the preset minimum learning rate step size. This represents the preset initial learning rate step size, and Q represents the difference coefficient. This represents the maximum absolute value of the difference coefficient;

[0027] If the difference coefficients of all types of load-affected data sequences are not negative, the LSTM prediction model will not be adjusted; if the difference coefficients of any type of load-affected data sequences are negative, the preset initial learning rate step size corresponding to the load-affected data sequences will be replaced with the adaptive learning rate step size to obtain the adaptive LSTM prediction model.

[0028] The beneficial effects of the embodiments of the present invention compared with the prior art are as follows:

[0029] This invention utilizes a sliding analysis window to adaptively determine the main change cycle of the load impact data sequence, thereby enabling more accurate analysis of data change characteristics. It acquires state feature values ​​to characterize the data change characteristics within the sliding analysis window, which can then be used to determine whether the data change pattern has changed. It acquires state feature reference values ​​as a benchmark for measuring the deviation of state feature values. It acquires the degree of deviation to characterize the extent of deviation in state feature values. It acquires window clusters to group sliding analysis windows with similar deviation degrees into the same cluster, thus determining the deviation boundary value. It acquires the deviation boundary value and the normal distribution ratio to reflect whether the data change pattern of the load impact data sequence has changed. It acquires the difference coefficient to reflect whether the change pattern of new data differs from historical data, thus determining whether the LSTM prediction model needs parameter adjustment. Finally, it adjusts the initial learning rate step size of the load impact data sequence in the LSTM prediction model based on different types of difference coefficients to obtain an adaptive LSTM prediction model. This results in higher prediction accuracy for the adaptive LSTM prediction model, avoiding situations where the load prediction effect is poor due to changes in the new data change pattern, ultimately leading to higher accuracy in ship power distribution optimization. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1This is a flowchart of a dynamic optimization method for a ship intelligent power distribution system based on dynamic perception of ship load, provided in Embodiment 1 of the present invention. Detailed Implementation

[0032] Embodiments of this disclosure are described in detail below, with examples of these embodiments illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this disclosure, and should not be construed as limiting it.

[0033] It should be noted that the terms "first," "second," etc., used in this disclosure and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure.

[0034] To illustrate the technical solution of the present invention, specific embodiments are described below.

[0035] The specific scenario addressed by this invention is as follows: During ship operation, the power load is affected by many factors, and the rapid changes in load demand lead to low accuracy of existing prediction methods. Therefore, this invention analyzes the changes in load data to adaptively adjust the parameters of the prediction model, thereby improving the accuracy of load prediction and power distribution optimization.

[0036] This invention provides a dynamic optimization method for a ship intelligent power distribution system based on dynamic sensing of ship load, such as... Figure 1 As shown, the method includes the following steps:

[0037] Step S1: Obtain load impact data sequences in different dimensions during ship operation.

[0038] First, load impact data sequences from different dimensions during ship operation are acquired. Shipboard sensors are used to collect multi-source operational data in real time, including electrical load information such as main power distribution bus current and voltage data, and power data from each generator; ship operating status information such as speed, main propeller speed and propulsion power, and hull attitude parameters (pitch angle, roll angle, and their rate of change); and external environmental information such as wind speed, wind direction, wave height, and heading changes. All data from all dimensions are collected and recorded synchronously with a unified time reference. Different load impact data sequences are constructed for different monitoring objects. To avoid the influence of data units, all data dimensions undergo linear normalization. The acquired load impact data sequences reflect changes in ship operating status and load conditions. Implementers can determine the data collection objects according to the implementation scenario.

[0039] Step S2: Obtain a sliding analysis window based on the distribution characteristics of the spectral data of the load impact data sequence; obtain different types of state characteristic values ​​and state characteristic reference values ​​based on the data change characteristics within the sliding analysis window of the load impact data sequence; obtain the degree of deviation based on the difference between the state characteristic values ​​and state characteristic reference values ​​corresponding to the sliding analysis window.

[0040] During ship operation, electrical load is affected by multiple factors such as ship speed, propeller condition, hull attitude, and sea state, resulting in rapidly changing and highly sudden load demands. Therefore, only through the fusion analysis of multi-dimensional load impact data sequences can a more comprehensive mapping relationship between operating status and load demand be constructed. This ensures that the power distribution system can be rationally scheduled according to actual load demand, improving distribution efficiency and reducing the possibility of overload or energy waste. Traditional load forecasting methods, such as BP neural networks or LSTM neural networks, often rely on historical load data or single electrical quantity modeling. Furthermore, the parameters are fixed after model training. When the variation patterns of new data differ significantly from historical data, the prediction results may deviate considerably, making it difficult for the distribution coefficient to accurately match load demand. Therefore, to improve the accuracy of distribution optimization, it is necessary to detect the characteristics of the input data in real time during the forecasting process, determine whether its variation pattern conforms to historical patterns, and trigger the optimization mechanism of the prediction model based on the data characteristics to dynamically update the model parameters. In this embodiment of the invention, an existing LSTM neural network model is used to construct a ship load forecasting model. This model can automatically learn the linear and non-linear correlations between the input multi-dimensional load impact data sequences, thereby achieving the prediction of ship load. In the prediction process, the collected load impact data sequences from different dimensions are first denoised using an adaptive Kalman filter algorithm to initially improve data accuracy. The denoised data is then used as input to construct a ship load prediction model based on an LSTM neural network. The output is the ship's electrical power consumption. The model is trained using the mean square error between the predicted and actual values ​​as the loss function, gradually bringing the model parameters closer to their optimal state, ultimately yielding an applicable LSTM prediction model. It should be noted that this algorithm is existing technology, and the specific steps are not detailed here.

[0041] Furthermore, during the ship's operation, if the changing pattern of newly collected data differs significantly from the historical data changing pattern used in the model training phase, continuing to use fixed model parameters for prediction can easily lead to increased prediction errors. Therefore, a dynamic update mechanism for model parameters needs to be introduced. When a mismatch between new data features and historical patterns is detected, the model parameters need to be re-optimized to improve the prediction accuracy of the prediction model for load changes under new operating conditions. First, the data characteristics of the load impact data sequence are analyzed. Therefore, a sliding analysis window is obtained based on the distribution characteristics of the spectral data of the load impact data sequence. Preferably, in this embodiment of the invention, the step of obtaining the sliding analysis window includes: obtaining the spectral data of the load impact data sequence through short-time Fourier transform; using different amplitudes in the spectral data as weights for corresponding frequencies, with higher amplitudes corresponding to higher frequency weights; calculating the weighted average of all frequencies in the spectral data to obtain frequency characteristic values; these frequency characteristic values ​​characterize the main changing frequencies of the load impact data sequence, thereby enabling the determination of the main changing period duration of the data. The reciprocal of the frequency characteristic value is calculated and rounded to obtain the window length, which can be used as the length for data feature extraction. Half of the window length is used as the window sliding step size, and the window is slid across the load impact data sequence, obtaining a sliding analysis window with each slide. Obtaining the sliding analysis window can match the data's own variation patterns, avoiding the problem of poor scale adaptability and flexibility of fixed time windows for data of different dimensions. After obtaining the sliding analysis window, its local data characteristics can be analyzed. Therefore, based on the data variation characteristics within the sliding analysis window of the load impact data sequence, different types of state characteristic values ​​and state characteristic reference values ​​are obtained.

[0042] Preferably, in this embodiment of the invention, the steps of obtaining state characteristic values ​​and state characteristic reference values ​​include: state characteristic values ​​including trend characteristic values ​​and degree characteristic values; calculating the average value of the first-order difference of the data within the sliding analysis window to obtain the trend characteristic value; and calculating the average value of the second-order difference of the data within the sliding analysis window to obtain the degree characteristic value. During ship navigation, the change characteristics of data in different dimensions within a short time window can directly reflect the stability of the current operating conditions and the load change trend; for example, when the speed remains basically constant and the sea state is stable, the average value of the first-order difference of the generator output power and the bus current is close to 0, and the average value of the second-order difference is small, indicating that the load change trend is gentle and the system is operating stably; when the ship accelerates or the wind increases, the average value of the first-order difference of the propeller speed and propulsion power increases significantly, and the average value of the second-order difference corresponding to the roll angle and its rate of change increases significantly, indicating that the load demand is showing a continuous increase and increased fluctuation. Therefore, by calculating the trend characteristic value and the degree characteristic value of change, the trend characteristics and drastic degree of load change under the ship's operating state can be effectively characterized, providing a reliable basis for the dynamic adjustment of subsequent model parameters; each sliding analysis window corresponds to two different types of state characteristic values, and subsequent steps will analyze the two types of state characteristic values ​​independently.

[0043] Since the characteristics and distributions of data changes may differ across dimensions, reference values ​​for the state characteristic values ​​of load impact data sequences of any dimension are needed. The average of the absolute differences between the state characteristic values ​​of any sliding analysis window and all other sliding analysis windows is calculated and negatively correlated to obtain the confidence level of any sliding analysis window; the state characteristic value of the sliding analysis window corresponding to the maximum confidence level is used as the state characteristic reference value. In this embodiment of the invention, through... Perform negative correlation mapping, where w represents the mapping object. This represents an exponential function with the natural constant as its base. Higher confidence levels mean a smaller difference between the state characteristic value and other state characteristic values ​​in the sliding analysis window, indicating that the state characteristic value is more representative and more likely to serve as a reference value for state characteristics.

[0044] Furthermore, the degree of deviation can be obtained based on the difference between the state feature value corresponding to the sliding analysis window and the state feature reference value. Preferably, in this embodiment of the invention, the step of obtaining the degree of deviation includes: calculating the absolute value of the difference between the state feature value of the sliding analysis window and the state feature reference value of the same type to obtain the degree of deviation. The greater the degree of deviation, the more the state feature value deviates from the state feature reference value.

[0045] Step S3: Cluster the windows based on the difference in the degree of deviation between any two sliding analysis windows to obtain window clusters; obtain the critical representativeness based on the distribution characteristics of the degree of deviation within the window clusters; obtain the deviation boundary value of the degree of deviation and the normal distribution ratio of the load-affected data sequence based on the critical representativeness; obtain the new normal distribution ratio based on the state characteristic value of the new sliding analysis window; obtain the difference coefficient based on the difference between the new normal distribution ratio and the normal distribution ratio.

[0046] After obtaining the degree of deviation of the state characteristic values ​​of different sliding analysis windows, the distribution of state characteristic values ​​and deviation boundary values ​​in the load influence data sequence can be analyzed; calculating the deviation boundary values ​​can be used to judge the applicability of the model when new data is added later; therefore, clustering is performed based on the difference characteristics of the degree of deviation of any two sliding analysis windows to obtain window clusters; preferably, in this embodiment of the invention, the step of obtaining window clusters includes:

[0047]

[0048] In the formula, This represents the probability of mutual selection between the i-th and j-th sliding analysis windows; Represents a linear normalization function. This indicates the degree of deviation in the i-th sliding analysis window. This represents the degree of deviation of the j-th sliding analysis window. The mutual selection probability value is used as the similarity between sliding analysis windows in the K-means clustering algorithm to obtain different window clusters. A larger mutual selection probability value means a smaller difference in the degree of deviation between the two sliding analysis windows, and the more likely their corresponding sliding analysis windows belong to the same cluster.

[0049] Furthermore, after obtaining different window clusters, for a load impact data sequence of any dimension, the greater the deviation of its state characteristic values ​​and the more densely distributed rather than discrete the characteristics, the more likely this deviation is to be a deviation boundary value. Therefore, the critical representativeness can be obtained based on the distribution characteristics of the corresponding deviation within the window cluster. Preferably, in this embodiment of the invention, the step of obtaining the critical representativeness includes: calculating the average of the absolute values ​​of the differences in deviations between any two sliding analysis windows within the window cluster and performing a negative correlation mapping to obtain the concentration degree; the greater the concentration degree, the denser the distribution of the corresponding deviation within the window cluster, and the more likely it is to be a true deviation situation. Calculate the average deviation of all sliding analysis windows within the window cluster to obtain the average deviation degree; the greater the deviation degree, the more obvious the deviation characteristics of the state characteristic values ​​within the window cluster, and the more likely it is to be a deviation boundary value. Calculate the product of the concentration degree and the average deviation degree to obtain the critical representativeness of the window cluster; the greater the critical representativeness, the more likely the state characteristic values ​​within the window cluster are to be deviation boundary values. Then, the deviation boundary value of the deviation degree and the normal distribution ratio of the load impact data sequence can be obtained based on the critical representativeness.

[0050] Preferably, in this embodiment of the invention, the steps of obtaining the deviation boundary value and the normal distribution ratio include: calculating the average deviation of the sliding analysis window within the window cluster corresponding to the maximum value of the critical representativeness to obtain the deviation boundary value; the deviation boundary value characterizes the maximum deviation of the state feature value in accordance with the current ship navigation state, and can subsequently be used to determine whether the change pattern of the dimensional data has changed significantly. The ratio of the number of state feature values ​​in the load impact data sequence whose deviation is less than the deviation boundary value to the total number of state feature values ​​is calculated to obtain the normal distribution ratio. The normal distribution ratio characterizes the proportion of the load impact data sequence that meets the condition of being less than the deviation boundary value in the current ship navigation state. Furthermore, when the load impact data sequence is updated, a new normal distribution ratio can be obtained based on the state feature values ​​of the new sliding analysis window. When the length of the new data satisfies a sliding analysis window, the state feature value of the new sliding analysis window is calculated, and its corresponding deviation is compared with the deviation boundary value to obtain the latest new normal distribution ratio. If the new normal distribution ratio is less than the previously obtained normal distribution ratio, it means that the new data characteristics of the data sequence affected by the load have changed significantly from the historical data change pattern. At this time, the prediction model needs to be retrained to improve the prediction accuracy. Therefore, the difference coefficient is obtained based on the difference between the new normal distribution ratio and the normal distribution ratio. Preferably, in this embodiment of the invention, the step of obtaining the difference coefficient includes: calculating the difference between the new normal distribution ratio and the normal distribution ratio to obtain the difference coefficient.

[0051] Step S4: Adjust the initial learning rate step size of the load impact data sequence in the LSTM prediction model according to different types of difference coefficients to obtain an adaptive LSTM prediction model; predict the ship load status according to the adaptive LSTM prediction model, and optimize power distribution according to the ship load status.

[0052] When the difference coefficient is negative, the proportion of the new normal distribution is less than the previously obtained proportion of the normal distribution, meaning that the new data characteristics of the load-affected data sequence differ significantly from the historical data's change pattern, thus requiring adjustment of the model parameters. Conversely, when the difference coefficient is not negative, it means that the proportion of the new normal distribution is not less than the previously obtained proportion of the normal distribution, the change pattern of the new data has not changed significantly, and high accuracy is achieved without adjusting the model parameters. Therefore, the initial learning rate step size of the load-affected data sequence in the LSTM prediction model can be adjusted according to different types of difference coefficients to obtain an adaptive LSTM prediction model. Preferably, in this embodiment of the invention, the steps for obtaining the adaptive LSTM prediction model include:

[0053]

[0054] In the formula, X represents the adaptive learning rate step size. This indicates taking the minimum value. This indicates the preset maximum learning rate step size. This indicates the preset minimum learning rate step size. This represents the preset initial learning rate step size, and Q represents the difference coefficient. This represents the maximum absolute value of the difference coefficient. The learning rate controls the learning speed and convergence of the prediction model. The difference coefficient can be used as a learning rate step size adjustment factor. Implementers can determine the preset maximum and minimum learning rate step sizes according to the implementation scenario. The preset initial learning rate step size is the parameter in the trained LSTM prediction model. In this embodiment of the invention, two types of state feature values ​​correspond to two difference coefficients. If the difference coefficients of all types of load impact data sequences are not negative, the LSTM prediction model is not adjusted. If the difference coefficients of any type of load impact data sequence are negative, the preset initial learning rate step size corresponding to the load impact data sequence is replaced with an adaptive learning rate step size to obtain an adaptive LSTM prediction model. By adaptively adjusting the learning rate step size of load impact data sequences of different dimensions, the adaptive LSTM prediction model can more accurately predict ship load, thereby improving the accuracy of power distribution optimization. Finally, the ship load status can be predicted according to the adaptive LSTM prediction model, and the power distribution can be optimized according to the ship load status. Implementers can determine the method of power distribution optimization based on the prediction results according to the implementation scenario, which is not limited here.

[0055] In summary, this invention provides a dynamic optimization method for a ship intelligent power distribution system based on dynamic perception of ship load. It obtains a sliding analysis window based on the spectral data of the load impact data sequence; obtains state feature values ​​and state feature reference values ​​based on the data change characteristics within the sliding analysis window; obtains the degree of deviation based on the state feature values ​​and state feature reference values; performs clustering based on the degree of deviation; obtains deviation boundary values ​​and normal distribution ratios based on the degree of deviation within the window clusters; obtains a new normal distribution ratio based on the state feature values ​​of future new sliding analysis windows; and obtains a difference coefficient based on the new normal distribution ratio and the normal distribution ratio. This invention adjusts the initial learning rate step size in the LSTM prediction model based on the difference coefficient to obtain an adaptive LSTM prediction model and predict the ship load state. Power distribution is optimized based on the ship load state, improving the accuracy of load prediction and power distribution optimization.

[0056] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A dynamic optimization method for a ship intelligent power distribution system based on dynamic perception of ship load, characterized in that, The method includes the following steps: Obtain load impact data sequences from different dimensions during ship operation; A sliding analysis window is obtained based on the distribution characteristics of the spectral data of the load impact data sequence; different types of state characteristic values ​​and state characteristic reference values ​​are obtained based on the data change characteristics within the sliding analysis window of the load impact data sequence; the degree of deviation is obtained based on the difference between the state characteristic values ​​corresponding to the sliding analysis window and the state characteristic reference values. Clustering is performed based on the difference in the degree of deviation between any two sliding analysis windows to obtain window clusters; critical representativeness is obtained based on the distribution characteristics of the degree of deviation within the window clusters; the deviation boundary value of the degree of deviation and the normal distribution ratio of the load impact data sequence are obtained based on the critical representativeness; a new normal distribution ratio is obtained based on the state characteristic value of the future new sliding analysis window; and a difference coefficient is obtained based on the difference between the new normal distribution ratio and the normal distribution ratio. The initial learning rate step size of the load impact data sequence in the LSTM prediction model is adjusted according to different types of difference coefficients to obtain an adaptive LSTM prediction model; the ship load status is predicted according to the adaptive LSTM prediction model, and the power distribution is optimized according to the ship load status.

2. The dynamic optimization method for a ship intelligent power distribution system based on dynamic perception of ship load as described in claim 1, characterized in that, The step of obtaining the sliding analysis window based on the distribution characteristics of the spectral data of the load impact data sequence includes: The spectral data of the load impact data sequence is obtained by short-time Fourier transform; different amplitudes in the spectral data are used as weights of corresponding frequencies, and the weighted average of all frequencies in the spectral data is calculated to obtain frequency characteristic values; the reciprocal of the frequency characteristic values ​​is calculated and rounded to obtain the length of the window; half of the window length is used as the window sliding step size, and the window is slid across the load impact data sequence, obtaining a sliding analysis window with each slide.

3. The dynamic optimization method for a ship intelligent power distribution system based on dynamic perception of ship load as described in claim 1, characterized in that, The steps of obtaining different types of state characteristic values ​​and state characteristic reference values ​​based on the data change characteristics within the sliding analysis window of the load impact data sequence include: The state characteristic values ​​include trend characteristic values ​​and degree of change characteristic values; the average value of the first difference of the data within the sliding analysis window is calculated to obtain the trend characteristic value; the average value of the second difference of the data within the sliding analysis window is calculated to obtain the degree of change characteristic value. The average of the absolute differences between the state feature values ​​of any sliding analysis window and all other sliding analysis windows is calculated and negatively correlated to obtain the confidence level of the arbitrary sliding analysis window; the state feature value of the sliding analysis window corresponding to the maximum confidence level is used as the state feature reference value.

4. The dynamic optimization method for a ship intelligent power distribution system based on dynamic perception of ship load according to claim 1, characterized in that, The step of obtaining the degree of deviation based on the difference between the state feature value corresponding to the sliding analysis window and the state feature reference value includes: The degree of deviation is obtained by calculating the absolute value of the difference between the state feature value of the sliding analysis window and the state feature reference value of the same type.

5. The dynamic optimization method for a ship intelligent power distribution system based on dynamic perception of ship load as described in claim 1, characterized in that, The step of clustering based on the difference in the degree of deviation between any two sliding analysis windows to obtain window clusters includes: In the formula This represents the probability of mutual selection between the i-th and j-th sliding analysis windows; Represents a linear normalization function. This indicates the degree of deviation in the i-th sliding analysis window. The deviation degree of the j-th sliding analysis window is represented; the mutual selection probability value is used as the similarity between sliding analysis windows in the K-means clustering algorithm to obtain different window clusters.

6. The dynamic optimization method for a ship intelligent power distribution system based on dynamic perception of ship load as described in claim 1, characterized in that, The step of obtaining the critical representativeness based on the distribution characteristics of the corresponding deviation degree within the window cluster includes: Calculate the average absolute value of the difference in the degree of deviation between any two sliding analysis windows within the window cluster and perform a negative correlation mapping to obtain the degree of concentration; calculate the average degree of deviation of all sliding analysis windows within the window cluster to obtain the average degree of deviation; calculate the product of the degree of concentration and the average degree of deviation to obtain the critical representativeness of the window cluster.

7. The dynamic optimization method for a ship intelligent power distribution system based on dynamic perception of ship load according to claim 1, characterized in that, The steps of obtaining the deviation boundary value of the deviation degree and the normal distribution ratio of the load-affected data sequence based on the critical representativeness include: Calculate the average deviation of the sliding analysis windows within the window cluster corresponding to the maximum value of the critical representativeness to obtain the deviation boundary value; calculate the ratio of the number of state feature values ​​in the load impact data sequence whose deviation is less than the deviation boundary value to the total number of state feature values ​​to obtain the normal distribution ratio.

8. The dynamic optimization method for a ship intelligent power distribution system based on dynamic perception of ship load according to claim 1, characterized in that, The step of obtaining the difference coefficient based on the difference characteristics between the new normal distribution ratio and the normal distribution ratio includes: The difference between the new normal distribution ratio and the normal distribution ratio is calculated to obtain the difference coefficient.

9. The dynamic optimization method for a ship intelligent power distribution system based on dynamic perception of ship load according to claim 1, characterized in that, The step of adjusting the initial learning rate step size of the load impact data sequence in the LSTM prediction model according to different types of difference coefficients to obtain an adaptive LSTM prediction model includes: In the formula, X represents the adaptive learning rate step size. This indicates taking the minimum value. This indicates the preset maximum learning rate step size. This indicates the preset minimum learning rate step size. This represents the preset initial learning rate step size, and Q represents the difference coefficient. This represents the maximum absolute value of the difference coefficient; If the difference coefficients of all types of load-affected data sequences are not negative, the LSTM prediction model will not be adjusted; if the difference coefficients of any type of load-affected data sequences are negative, the preset initial learning rate step size corresponding to the load-affected data sequences will be replaced with the adaptive learning rate step size to obtain the adaptive LSTM prediction model.

Images