Marine BMS state prediction method and system based on machine learning

Abstract

The invention relates to the technical field of battery management systems, and discloses a marine BMS state prediction method and system based on machine learning, and the method comprises the steps: collecting and preprocessing ship BMS parameter data and environment data, and obtaining the preprocessed data; constructing a battery compartment temperature field model and calculating a temperature stratification index TLI; automatically dividing data segments according to the fluctuation characteristics of the preprocessed data, and configuring differential prediction parameters; constructing a multivariable grey prediction model considering the influence of temperature stratification, and dynamically adjusting the parameters of the grey prediction model by using a temperature stratification index TLI; dynamically selecting an optimal prediction model combination and generating a final prediction result; according to the method, a temperature stratification evaluation and compensation mechanism is introduced, and a multivariable grey prediction model is combined, so that the prediction error is reduced in an environment with obvious temperature stratification.

Description

Technical Field

[0001] The present invention relates to the technical field of battery management systems, and more specifically, to a method and system for predicting the state of a marine BMS based on machine learning. Background Art

[0002] With the development of new energy technologies and the improvement of environmental protection requirements, electric propulsion ships are being more and more widely used in the maritime field. In electric propulsion ships, the battery management system (BMS), as the core control unit, the accurate assessment and prediction of its health state are crucial for ensuring the safe navigation of the ship. Different from the land application environment, ships face complex and changeable environmental factors during long-term voyages at sea, such as extreme sea conditions, temperature fluctuations, and high humidity. These factors pose great challenges to the measurement and prediction of the BMS state.

[0003] Existing BMS state prediction technologies mainly rely on statistical methods, machine learning algorithms, or deep learning models, such as neural networks, support vector machines, etc. These methods perform well in the land-fixed environment, but have obvious deficiencies in the complex marine environment of ships: on the one hand, traditional statistical methods are difficult to handle the high-noise, few-sample, and multi-variable data characteristics in the marine environment, resulting in low prediction accuracy; on the other hand, although complex deep learning models theoretically have higher accuracy, they pose challenges to the limited computing resources of ships and are difficult to achieve real-time and efficient prediction.

[0004] Even more prominent is the fact that existing prediction methods generally ignore two key characteristics of the ship environment: one is the difference in the fluctuation characteristics of BMS parameters under different sea conditions, and traditional single prediction models cannot adapt to such changes in fluctuation characteristics; the other is the phenomenon of temperature stratification commonly existing in the battery compartments of large ships, where batteries at different positions are affected by different temperatures, and existing methods usually regard the entire battery system as a uniform whole, ignoring the impact of this stratification effect on prediction accuracy.

[0005] Therefore, there is an urgent need for a method for predicting the state of a marine BMS that can adapt to complex marine environments, consider the influence of temperature stratification, and achieve high-precision prediction with limited computing resources, providing reliable technical support for the safety management and navigation of electric propulsion ships. Summary of the Invention

[0006] The present invention provides a method and system for predicting the state of a marine BMS based on machine learning, which solves the technical problems of low accuracy, poor adaptability, and high computing resource consumption in BMS state prediction in complex marine environments in related technologies.

[0007] The present invention provides a method for predicting the state of a marine BMS based on machine learning, including: Collect ship BMS parameter data and environmental data and pre-process them to obtain pre-processed data; Based on the preprocessed data, the battery compartment temperature field model is constructed and the temperature stratification index TLI is calculated; Automatically divide data segments and configure differentiated prediction parameters based on the volatility characteristics of preprocessed data; Based on the temperature stratification index TLI and differentiated prediction parameters, a multivariate grey prediction model considering the influence of temperature stratification is constructed, and the temperature stratification index TLI is used to dynamically adjust the parameters of the grey prediction model. Based on the current sea state level and temperature stratification index TLI, the optimal prediction model combination is dynamically selected and the final prediction result is generated.

[0008] Furthermore, the collecting of ship BMS parameter data and environmental data and preprocessing the data includes: Use wavelet transform to eliminate extreme fluctuation noise in raw data; Standardize the denoised data; The key influencing factors were identified through correlation analysis and principal component analysis.

[0009] Furthermore, the calculation formula of the temperature stratification index TLI is: ; in is the temperature stratification index; and Sensors and sensors Temperature value; is the distance weight function, For sensor and sensors The spatial distance between is the number of temperature sensors, Represents the summation symbol.

[0010] Furthermore, the step of automatically dividing the data segments according to the fluctuation characteristics of the preprocessed data includes: Calculate the standard deviation of time series data within the sliding window as a volatility assessment indicator; Setting multiple adaptive thresholds to divide the data into multiple segments; A smaller prediction time window is used for segments with sharp fluctuations, and differentiated prediction parameters are configured for different data segments.

[0011] Furthermore, the construction of a multivariate grey prediction model considering the influence of temperature stratification includes: Construct the grey differential equation: ; Among them is the development coefficient, representing the driving parameter for the internal development of the system, is the driving coefficient, representing the influence degree of the th factor on the system, represents the derivative with respect to time , represents the cumulative sequence of the variable to be predicted, represents the cumulative sequence of the th factor, represents the total number of influencing factors, represents the summation symbol; Introduce the temperature stratification compensation function, and dynamically adjust the model parameters according to the temperature stratification index TLI: ; ; Among them represents the adjusted development coefficient, is the development coefficient, is the sensitivity coefficient of temperature stratification to the development coefficient, represents the adjusted th driving coefficient, is the original th driving coefficient, is the sensitivity coefficient of temperature stratification to the driving coefficient, is the temperature stratification index.

[0012] Furthermore, it also includes establishing a multi-layer recursive residual correction system, and the steps are as follows: Calculate the residual sequence between the predicted value and the actual value; Apply the grey prediction model to the residual sequence again to obtain the residual predicted value; Add the original predicted value and the residual predicted value to obtain the corrected prediction result.

[0013] Furthermore, the steps of the dynamic selection of the optimal prediction model combination include: Calculate the prediction error evaluation function considering the influence of temperature stratification: ; Among them is the predicted value of the model for the data at the time point , is the actual value at the time point , is the temperature stratification influence coefficient, is the evaluation time window length, For the prediction error considering the influence of temperature stratification, denotes the temperature stratification index, denotes the model index, denotes the dataset index, denotes the summation symbol; Select the optimal prediction model or model combination according to the error evaluation function.

[0014] Furthermore, if multi-model fusion is selected, the fusion weights are dynamically determined based on the historical performance of each model and the current environmental conditions: ; where is the prediction error of model within the most recent evaluation window, is the total number of models, is the model fusion weight, denotes the model index, denotes the th model's prediction error within the most recent evaluation window, denotes the summation symbol; In addition, the weights can be further adjusted according to the sea state level and the temperature stratification index: ; where is the adjusted fusion weight of model ; is the fusion weight of model ; denotes the temperature stratification index; denotes the sea state level; and are the sensitivity coefficients of model to the sea state and temperature stratification, obtained through historical data analysis.

[0015] Furthermore, the final prediction result includes the predicted values of the main state parameters of the BMS and the visual output of the prediction information.

[0016] The present invention provides a marine BMS state prediction system based on machine learning for performing the above-mentioned marine BMS state prediction method based on machine learning, including: A data acquisition and processing unit for collecting and preprocessing the ship BMS parameter data and environmental data; A temperature field modeling unit for constructing a battery compartment temperature field model and calculating the temperature stratification index; An adaptive segmentation unit for automatically dividing data segments according to the BMS data fluctuation characteristics and configuring differential prediction parameters; Grey prediction model unit, used to construct a multivariate grey prediction model considering the influence of temperature stratification and dynamically adjust the model parameters according to the temperature stratification index; The multi-model fusion unit is used to dynamically select and fuse prediction models based on the current sea conditions and temperature stratification status to generate the final prediction results.

[0017] The beneficial effects of the present invention are as follows: by introducing a temperature stratification evaluation and compensation mechanism and combining it with a multivariate grey prediction model, the prediction error is reduced in an environment with obvious temperature stratification; when the sample size is reduced, the prediction accuracy is improved compared with the traditional neural network prediction model, and the state change of the marine BMS can be predicted more accurately; The innovative introduction of an adaptive segmentation system can automatically adjust prediction parameters and strategies according to different sea conditions and data fluctuation characteristics, and improve prediction stability under highly volatile sea conditions, which is particularly suitable for battery management needs of ships under extreme sea conditions; The multi-model fusion system built based on grey prediction theory has lower computing resource requirements than traditional deep learning methods. It can achieve real-time prediction updates on the computing hardware of standard ship configuration, perfectly adapting to the limited computing resource environment of ships. Through multi-model fusion technology, potential BMS anomalies can be predicted in advance with improved accuracy, providing crew members with sufficient response time and decision-making support, reducing the risk of maritime failures.

[0018] The method combines the characteristics of the actual operating environment of ships and can be seamlessly integrated with the existing ship BMS without additional hardware investment. It is implemented through software modules, which greatly improves the promotion and application value of the technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of a method for predicting a state of a marine BMS based on machine learning of the present invention; Figure 2 is a flow chart of step 1 in the present invention; Figure 3 is a flow chart of step 2 in the present invention; Figure 4 It is a flow chart of step 3 in the present invention; Figure 5 is a flow chart of step 4 in the present invention; Figure 6 It is a flow chart of step 5 in the present invention. DETAILED DESCRIPTION

[0020] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein, and the functions and arrangements of the elements discussed can be changed without departing from the scope of protection of the content of this specification. Each example can omit, substitute, or add various processes or components as needed. Additionally, the features described in some examples can also be combined in other examples.

[0021] In at least one embodiment of the present invention, a method for predicting the state of a marine BMS based on machine learning is disclosed, as Figures 1 to 6 shown, including: Step 1, collect ship BMS parameter data and environmental data and perform preprocessing to obtain preprocessed data; In this step, a shipborne sensor network is used to collect multi-source BMS parameter data and environmental data, and a dataset for prediction is generated through denoising and normalization processing.

[0022] Step 1.1, collect multi-source BMS parameter data and environmental data; The multi-source BMS parameter data includes electrical parameters such as the voltage of each battery cell ( ), battery current ( ), the temperature of each battery cell ( ), state of charge (SOC), state of health (SOH), etc.; The environmental data includes sea state level ( ), environmental temperature ( ), humidity ( ), etc.; Where represents the battery cell number, , is the total number of battery cells; Step 1.2, wavelet transform denoising; Apply wavelet transform to the original data to eliminate extreme fluctuation noise, select an appropriate wavelet basis function for multi-scale decomposition to obtain different frequency components; For the signal , its wavelet transform is: ; Where is the coefficient after wavelet transform, is the scale parameter, is the translation parameter; is the complex conjugate of the wavelet basis function , is the time variable, is the original signal, Represents the radical sign, Indicates the Integral from negative infinity to positive infinity; Step 1.3, data standardization processing; Perform standardization processing on the data after wavelet denoising so that parameters of different types and different dimensions can be uniformly processed: ; Where Is the standardized data, Is the original data, Is the Mean of the Is the Standard deviation of the Step 1.4, correlation analysis to identify key influencing factors; Identify key influencing factors through correlation analysis and principal component analysis, calculate the correlation coefficients of each parameter with the BMS status index, and screen out the parameters whose absolute value of the correlation coefficient exceeds the threshold As key factors.

[0023] In some embodiments, the wavelet basis function can be selected as Daubechies wavelet (db4), symlet wavelet (sym8) or Coiflet wavelet (coif3), and the most suitable wavelet basis is selected according to the characteristics of BMS data.

[0024] Optionally, in extreme cases where the sea state level exceeds 7, a double denoising strategy can be adopted, that is, first apply wavelet denoising, and then combine the empirical mode decomposition (EMD) method to further reduce the noise impact.

[0025] Step 2, based on the preprocessed data, construct a battery compartment temperature field model and calculate the temperature stratification index TLI; In this step, a three-dimensional temperature field model of the battery compartment is constructed, and the temperature stratification index is calculated to quantitatively evaluate the temperature non-uniformity degree, and the temperature stratification index and the temperature field distribution map are output.

[0026] Step 2.1, arrange a multi-point temperature sensor array; Deploy Temperature sensors at key positions inside the battery compartment to collect spatial distribution temperature data in real time: ; Where , , Respectively represent the temperature values measured by the 1st, 2nd, th temperature sensors, Indicates the total number of temperature sensors; Step 2.2, construct a three-dimensional temperature field interpolation model; Based on multi-point temperature data, construct a three-dimensional temperature field interpolation model. For any position inside the cabin of the temperature value, use the distance weighted interpolation method to calculate: ; Among them represents the point to the sensor of the Euclidean distance, is the distance attenuation coefficient, is the position at the temperature value, is the th sensor measures the temperature value; is the total number of temperature sensors; is the weight function, and the calculation formula is: ; Among them represents the point to the sensor of the distance of the power; Step 2.3, calculate the temperature stratification index TLI; The temperature stratification index (Temperature Layering Index, TLI) quantitatively evaluates the non-uniformity of the temperature distribution inside the battery compartment: ; Among them is the temperature stratification index, indicating the non-uniformity of the temperature distribution; , is the sensor and of the temperature value, is the sensor and between the spatial distance, and respectively represent the sensor numbers at different positions; is the number of temperature sensors, represents the summation symbol; is the distance weight function, defined as: ; Among them is the base of the natural logarithm, is the distance attenuation coefficient, used to control the influence degree of the distance on the weight, is the sensor and The spatial distance between; Step 2.4, generate a temperature field distribution map and time series data; Generate a temperature field distribution map of the battery compartment and time series data of the temperature stratification index for input to the subsequent prediction model.

[0027] In some embodiments, for large ships with more than 100 battery units, optionally, a hierarchical and zonal temperature modeling method is adopted. The battery compartment is divided into multiple regions, and a temperature field model is established for each region separately. Then, the models of each region are connected through boundary conditions to reduce the computational complexity. Additionally, optionally, the principles of thermal fluid dynamics are used to assist in the construction of the temperature field. Especially in a high temperature difference environment, the thermal convection effect is considered to improve the accuracy of the temperature field.

[0028] In a specific application scenario, such as a polar navigation ship, the calculation of the temperature stratification index can adopt a segmented weighting strategy: when the temperature difference exceeds a preset threshold (usually 15 °C), the weighting function adopts non-linear enhancement: ; where is the enhancement coefficient, is the temperature difference threshold, represents the part where the temperature difference exceeds the threshold, and are the temperature values of sensors and respectively, is the enhanced distance weighting function; is the base of the natural logarithm, is the distance attenuation coefficient used to control the influence degree of distance on the weight, is the spatial distance between sensors and ; This method can more effectively identify extreme temperature difference situations.

[0029] Step 3, automatically divide the data segments and configure differential prediction parameters according to the fluctuation characteristics of the preprocessed data; This step automatically divides the data segments and configures differential prediction parameters based on the fluctuation characteristics of the BMS data, and outputs the optimal segmentation result and parameter configuration.

[0030] Step 3.1, calculate the fluctuation characteristic index of the BMS parameters; For time series data , calculate the standard deviation within the sliding window as the fluctuation degree evaluation index; where represents the window length, represents the current time point.

[0031] Step 3.2, set the adaptive threshold to divide the data segments; Set multiple adaptive thresholds: ; where 、 、 represent the 1st, 2nd, and the th adaptive thresholds respectively, represents the total number of thresholds; Divide the data into segments according to the degree of fluctuation, satisfying: ; where represents the th data segment, represents the standard deviation of the data within the window; and represent the th and the th thresholds respectively; represents the data value at the time point ; represents the window length, represents the current time point.

[0032] Step 3.3, configure different prediction parameters for different data segments; For the segments with large fluctuations, adopt a smaller prediction time window and configure different prediction parameters for different data segments, including but not limited to: prediction step , background value generation coefficient , regularization parameter ; Step 3.4, construct a sea condition - parameter mapping table; Use historical data to train to obtain the optimal parameter combinations for each segment and construct a sea condition - parameter mapping table: ; where represents the sea condition - parameter mapping table, is the optimal background value generation coefficient, is the optimal regularization parameter, is the optimal prediction step, is the sea condition level, is the temperature stratification index; Realize the adaptive selection of parameters under different sea conditions and temperature stratification conditions.

[0033] Optionally, when the ship is navigating in complex sea areas (such as areas where multiple ocean currents converge), the wave characteristic evaluation can be combined with the sample entropy index to form a two-dimensional evaluation system with the standard deviation index: ; where represents the eigenvector of the -th data point, represents the standard deviation of the data within the window, is the sample entropy; represents the time point data value; represents the window length, represents the current time point.

[0034] The K-means clustering algorithm based on two-dimensional indicators automatically determines the optimal number of segments and segment boundaries, improving the segmentation accuracy.

[0035] In specific shipping application scenarios, for example, for long-distance voyages of ocean freighters, an adaptive mechanism for the navigation stage can be optionally introduced. The navigation process is divided into stages such as the departure stage, the ocean voyage stage, and the port approach stage. Specialized segmentation systems are constructed respectively according to the characteristics of BMS data in different stages. For example, the battery usage pattern is relatively stable during the port approach stage, and a larger time window can be adopted; while the load changes frequently during the departure stage, a smaller time window is adopted to improve the system adaptability.

[0036] Step 4: Based on the temperature stratification index TLI and the differential prediction parameters, construct a multivariate grey prediction model considering the influence of temperature stratification, and use the temperature stratification index TLI to dynamically adjust the parameters of the grey prediction model; It should be understood that in this step, a GM(1, N) multivariate grey prediction model considering the influence of temperature stratification is constructed, and the BMS state prediction result is output.

[0037] Step 4.1: Construct a GM(1, N) multivariate grey prediction model; Construct a multivariate grey prediction model GM(1, N), where 1 represents a differential equation and N represents that the equation contains N variables. For the original data sequence and the related factor sequences: ; where , , respectively represent the original sequences of the 1st, 2nd, -th influencing factors, represents the total number of influencing factors; First, perform a first-order accumulation generation to obtain and , Denotes the serial number of the factor sequence. In the state prediction of marine BMS, denotes the original data sequence, including key state indicators such as battery health status or remaining life; while denotes the key factors affecting the BMS state, such as depth of discharge, number of cycles, temperature stratification index, etc.

[0038] Specifically, for any sequence: ; where denotes the original data sequence, , , respectively denote the data values at the 1st, 2nd, th time points in the original sequence, denotes the sequence length; Its first-order accumulated generating sequence is: ; where denotes the accumulated value at the th time point, denotes the value at the th time point in the original sequence, denotes the time point, denotes the sequence length, denotes the index of accumulation.

[0039] Similarly, for the factor sequence the same first-order accumulated generating operation is also performed to obtain ; Step 4.2, establish a grey differential equation model; The calculation formula is: ; where is the development coefficient, representing the driving parameter of the internal development of the system, is the driving coefficient, representing the influence degree of the th factor on the system, represents 's derivative with respect to time , denotes the accumulated sequence of the variable to be predicted, denotes the accumulated sequence of the th factor, denotes the total number of influencing factors, denotes the accumulated sequence of the variable to be predicted, denotes the summation symbol.

[0040] To discretize and solve the above differential equation, a mean generation sequence is used: ; where represents the value of the mean generation sequence at the -th time point, represents the cumulative value at the -th time point, represents the cumulative value at the -th time point, represents the time point, represents the sequence length.

[0041] After discretizing the differential equation, we get: ; where represents the value at the -th time point in the original sequence, is the development coefficient, represents the value of the mean generation sequence at the -th time point, is the driving coefficient of the -th factor, represents the cumulative value of the -th factor at the -th time point, represents the total number of influencing factors, represents the time point, represents the sequence length.

[0042] Step 4.3, introduce the temperature stratification compensation function; The temperature stratification compensation function corrects the prediction result. Define the temperature compensation function: ; where represents the temperature compensation function, represents the temperature value of the -th sensor, represents the temperature value of the -th sensor, represents the spatial distance between sensors and , is the -th basis function, represents the weight coefficient of the -th basis function, represents the index of the basis function, is the total number of basis functions.

[0043] In practical applications, the Radial Basis Function (RBF) can be selected as : ; where represents the th radial basis function, represents the temperature value of the th sensor, represents the temperature value of the th sensor, represents the spatial distance between sensors and ; , and respectively represent the temperature center, distance center, and width parameter of the th radial basis function, represents the natural exponential function, represents the width parameter of the th radial basis function.

[0044] Based on the temperature stratification index dynamically adjust the parameters and : ; ; where represents the adjusted development coefficient, is the development coefficient, is the sensitivity coefficient of temperature stratification to the development coefficient, represents the adjusted th driving coefficient, is the original th driving coefficient, is the sensitivity coefficient of temperature stratification to the driving coefficient, is the temperature stratification index.

[0045] For marine BMS, when the temperature stratification index is relatively large, the development rate of the system usually accelerates (the battery ages faster), so usually takes a positive value, ranging from 0.01 to 0.05; while determines its sign and magnitude according to the correlation with temperature stratification based on specific factors; Step 4.4, solve the model parameters and construct a residual correction system; Use the least squares method to solve the corrected coefficients and . Construct the parameter matrix: ; where represents the adjusted development coefficient, , , respectively represent the corrected first, second, th driving coefficients, represents the parameter matrix, represents the matrix transpose; The data matrix is: ; ; where represents the data matrix, containing the values of the original sequence from the second to the th time points; represents the coefficient matrix, containing the cumulative sequences of the mean generation sequence and each influencing factor; represents the value of the original sequence at the th time point, represents the value of the mean generation sequence at the th time point, represents the cumulative value of the th influencing factor at the th time point, represents the sequence length, represents the total number of influencing factors.

[0046] Solve for the parameters using the least squares method: ; where represents the parameter matrix, represents the transpose of the matrix , represents the inverse matrix of the matrix , represents the data matrix.

[0047] When there is noise or multicollinearity in the data, a regularization term can be introduced and the ridge regression method can be used to solve: ; where represents the parameter matrix, represents the transpose of the matrix , is the regularization parameter in ridge regression, is the identity matrix, represents the inverse matrix of the matrix , represents the data matrix.

[0048] Build a prediction model and make predictions. Establish a multi-layer recursive residual correction system for non-linear fluctuations, and improve the prediction accuracy through iterative calculations. Specifically, for the residual between the predicted value and the actual value at time point : ; where represents the residual at the -th time point, represents the actual value at the -th time point of the original sequence, represents the predicted value at the -th time point.

[0049] Construct a residual sequence: ; where represents the residual sequence, , , respectively represent the residuals at the 2nd, 3rd, and -th time points, represents matrix transpose, represents the sequence length.

[0050] Apply the grey prediction model to the residual sequence again to obtain the residual predicted value , and finally add the original predicted value and the residual predicted value to obtain the corrected prediction result: Apply the grey prediction model to the residual sequence again to obtain the residual predicted value , and finally add the original predicted value and the residual predicted value to obtain the corrected prediction result: ; where represents the prediction result at the -th time point after correction, represents the original predicted value, represents the residual predicted value, represents the prediction step size, represents the length of the original sequence.

[0051] In some embodiments, for ships sailing in tropical regions, the temperature stratification effect is more obvious. Optionally, a piecewise linear function is used to replace the simple linear relationship to adjust the parameters and : ; where represents the development coefficient after using the piecewise linear function, represents the development coefficient, and represent the sensitivity coefficients when the temperature stratification index is below and above the threshold respectively, represents the temperature stratification index, represents the temperature stratification threshold, represents the smaller value between and represents the part exceeding the threshold.

[0052] In a specific application scenario, such as the battery management system of military ships, a multi-time scale prediction strategy is optionally combined. For short-term prediction (within 24 hours), a first-order linear form of the grey prediction model is adopted; for medium-term prediction (1 - 7 days), a non-linear correction term is introduced; for long-term prediction (more than 1 month), a hybrid prediction system is constructed by combining the battery aging model and the grey prediction model to meet the prediction requirements of different time scales.

[0053] Step 5: Dynamically select the optimal prediction model combination based on the current sea state level and the temperature stratification index TLI, and generate the final prediction result; In addition, this step realizes dynamically selecting the optimal prediction model combination according to the current sea state and temperature distribution state, and generating the final prediction result.

[0054] Step 5.1: Select candidate models from the model library; According to the current sea state level and the temperature stratification index , select suitable candidate models from the pre-trained model library: ; where , , represent the 1st, 2nd, th candidate prediction models respectively, represents the total number of candidate models; In the application scenario of marine BMS, the model library includes but is not limited to the following types of models: Basic GM(1, 1) grey prediction model: suitable for stable sailing periods with small data fluctuations and strong trends; Non-equidistant GM(1, 1) model: suitable for cases where data sampling is uneven; Residual correction GM(1, N) model: the core model of this application, suitable for scenarios with multi-factor influences and temperature stratification phenomena; Fractional-order GM(r, N) model: suitable for battery degradation prediction with long-term memory characteristics; ARIMA time series model: As a supplementary model, it is suitable for short-term prediction; Step 5.2, calculate the prediction error evaluation function; The prediction error evaluation function takes into account the influence of temperature stratification, and its calculation formula is: ; Where is the prediction value of the model for the data at the time point , is the actual value, is the temperature stratification influence coefficient, is the evaluation time window length, is the prediction error considering the influence of temperature stratification, represents the temperature stratification index, represents the model index, represents the dataset index.

[0055] In the marine BMS system, according to historical data analysis, when the temperature stratification index exceeds the threshold (usually 5°C), the prediction error will increase. Therefore, is usually set between 0.05 - 0.15 to achieve a reasonable amplification of the prediction error in the case of severe temperature stratification, so that the evaluation function can more accurately reflect the actual performance of the model under different temperature stratification conditions; Step 5.3, select the optimal prediction model; Select the optimal prediction model or model combination according to the error evaluation function: ; Where is the weight of the th dataset, is the number of datasets, is the optimal model, represents the model index that minimizes the objective function ; is the prediction error considering the influence of temperature stratification, represents the temperature stratification index, represents the th dataset, represents the th prediction model; Step 5.4, dynamically determine the fusion weight; If multi-model fusion is selected, then based on the historical performance of each model and the current environmental conditions, dynamically determine the fusion weight . For the state prediction of the marine BMS, the inverse error weighting method is used to calculate the fusion weight: ; where is the prediction error of the model within the most recent evaluation window, is the total number of models, is the model 's fusion weight, represents the model index, represents the th model's prediction error within the most recent evaluation window.

[0056] In addition, the weights can be further adjusted according to the sea state level and the temperature stratification index: ; where is the adjusted fusion weight of the model , is the fusion weight of the model ; represents the temperature stratification index; represents the sea state level; and are the sensitivity coefficients of the model to the sea state and temperature stratification, obtained through historical data analysis.

[0057] For example, for the residual correction GM(1,N) model suitable for handling the influence of temperature stratification, its value will be larger, while for the fractional order GM(r,N) model suitable for handling sea state fluctuations, its value will be larger.

[0058] Step 5.5, generate the final prediction result; The final prediction result includes the predicted values of the main state parameters of the BMS and the visual output of the prediction information, providing decision support for the crew. The ship control system can automatically adjust the charging and discharging strategy according to the prediction result, optimize the battery service life, and give an early warning when an abnormal state is predicted, reducing the risk of sea failures.

[0059] The final prediction result is calculated as: ; where is the final prediction result at time point , is the predicted value of the model at time point , while ensuring (achieved through normalization); is the adjusted weight.

[0060] In some embodiments, a task-adaptive weight adjustment method is optionally adopted for different types of ship tasks. For example, for ships performing military tasks, more attention is paid to the stability and reliability of predictions, so the weights of models with higher stability are increased; while for commercial transport ships, more attention is paid to the real-time and accuracy of predictions, and the weights of each model are adjusted accordingly.

[0061] In specific extreme sea condition application scenarios, such as when sailing in a typhoon area, a time-varying weight strategy is optionally adopted to adjust the model fusion weights in real time as the sea conditions change. Before and after the sudden change of sea conditions, the weights of models with strong adaptability are increased; during the stable period of sea conditions, the weights of models with high accuracy are increased to achieve the dynamic optimization of model weights. In addition, ship navigation trajectory prediction is also optionally combined, and the model combination is pre-adjusted in advance according to the sea conditions and temperature distribution on the predicted route to achieve active prediction adjustment.

[0062] A marine BMS state prediction system based on machine learning, which is used to execute the above-mentioned marine BMS state prediction method based on machine learning, includes: A data acquisition and processing unit, which is used to collect and preprocess the ship BMS parameter data and environmental data; A temperature field modeling unit, which is used to construct a battery compartment temperature field model and calculate the temperature stratification index; An adaptive segmentation unit, which is used to automatically divide data segments according to the BMS data fluctuation characteristics and configure differential prediction parameters; A grey prediction model unit, which is used to construct a multi-variable grey prediction model considering the influence of temperature stratification and dynamically adjust the model parameters according to the temperature stratification index; A multi-model fusion unit, which is used to dynamically select and fuse prediction models based on the current sea conditions and temperature stratification status to generate the final prediction result.

[0063] Here, the present invention provides an implementation example: In the actual application on a certain large electric propulsion cargo ship, this method is used to predict the state change of the ship's battery management system. The cargo ship is equipped with a 1200 kWh lithium battery pack, which is composed of 1500 single cells connected in series and parallel and distributed in three independent battery compartments.

[0064] The system collected parameters including the voltages of 800 battery cells, battery current, data of 150 distributed temperature sensors, sea conditions, environmental temperature, etc. The db4 wavelet was used to denoise the original data to eliminate the high-frequency noise interference during navigation. Through correlation analysis, 12 key factors affecting the BMS state were determined, including the average battery temperature, maximum temperature difference, average SOC, discharge depth, cumulative cycle times, etc.

[0065] A three-dimensional temperature field model of the battery compartment was constructed based on data from 150 temperature sensors. In a case of a voyage across the equator, the system monitored a temperature difference of up to 12 °C between the upper and lower parts of the battery compartment, and the calculated temperature stratification index TLI was 8.3. The conventional BMS system could not consider this temperature stratification effect, resulting in a SOC prediction deviation of up to 15%. However, this method adjusts the prediction model by combining the temperature stratification index and controls the prediction deviation within 4%.

[0066] Based on the fluctuation characteristics of 30-day voyage data, the system automatically classifies the data into three categories: stable voyage segments, medium fluctuation segments, and high fluctuation segments. For the voyage record encountering sea state 9, the system identifies this segment of data as a high fluctuation segment, automatically shortens the prediction time window to 2 hours, and adjusts the parameters of the grey prediction model, improving the prediction stability by 59%.

[0067] A GM(1,7) multivariate grey prediction model considering the influence of temperature stratification was constructed, which includes 6 key influencing factors. In the case of obvious temperature stratification in the battery compartment, the model parameters are dynamically adjusted through the temperature stratification compensation function. When the temperature difference reaches 15 °C, the development coefficient increases by 32%, accurately reflecting the accelerated battery aging effect caused by temperature stratification. The average error between the SOH value predicted by the model and the actual measured value is only 2.3%, while the error of the traditional model is 7.8%.

[0068] The system dynamically selects and fuses multiple prediction models according to the sailing conditions. During a 45-day ocean voyage, the system automatically adjusts the model weights at different sailing stages: the weight of the GM(1,1) model is 0.6 during stable sailing, the weight of the residual correction GM(1,7) model increases to 0.75 in high fluctuation sea conditions, and the weight of the ARIMA model increases to 0.4 during the docking stage. The RMSE (root mean square error) of the SOC predicted by the fused model is 1.8%, which is 33% lower than 2.7% of a single model.

[0069] Through this practical application, the high precision and stability of this method in complex marine environments were verified. Before an unexpected storm arrived, the system successfully predicted the abnormal trend of the BMS, issued an early warning 96 hours in advance, giving the ship enough time to adjust its course to avoid the storm center and preventing possible battery system failures and safety risks. In addition, this system runs on the ship's standard computing platform, with an average CPU occupancy rate of only 12% and a memory occupancy of 108 MB, meeting the requirements of the ship's limited computing resources.

[0070] The embodiments of the present invention have been described above. However, these embodiments are not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative rather than restrictive. Under the inspiration of these embodiments, those of ordinary skill in the art can also make more equivalent embodiments in various forms, all of which fall within the protection scope of these embodiments.

Claims

1. A method for predicting the state of a marine BMS based on machine learning, characterized in that, Including: Collecting ship BMS parameter data and environmental data and performing preprocessing to obtain preprocessed data; Based on the preprocessed data, constructing a battery compartment temperature field model and calculating the temperature stratification index TLI; Automatically dividing data segments according to the fluctuation characteristics of the preprocessed data and configuring differential prediction parameters; Based on the temperature stratification index TLI and differential prediction parameters, constructing a multivariate grey prediction model considering the influence of temperature stratification, and dynamically adjusting the grey prediction model parameters using the temperature stratification index TLI; Based on the current sea state level and temperature stratification index TLI, dynamically selecting the optimal prediction model combination and generating the final prediction result.

2. The method for predicting the state of a marine BMS based on machine learning according to claim 1, characterized in that, The preprocessing in the collecting ship BMS parameter data and environmental data and performing preprocessing includes: Using wavelet transform to eliminate extreme fluctuation noise in the original data; Performing normalization processing on the denoised data; Identifying key influencing factors through correlation analysis and principal component analysis.

3. The method for predicting the state of a marine BMS based on machine learning according to claim 1, characterized in that, The calculation formula of the temperature stratification index TLI is: ； wherein is the temperature stratification index; and are the temperature values of sensors and sensor respectively; is the distance weight function, is the spatial distance between sensor and sensor ; is the number of temperature sensors, represents the summation symbol.

4. The method for predicting the state of a marine BMS based on machine learning according to claim 1, characterized in that, The steps of automatically dividing data segments according to the fluctuation characteristics of the preprocessed data include: Calculating the standard deviation of the time series data within the sliding window as the fluctuation degree evaluation index; Setting multiple adaptive thresholds to divide the data into multiple segments; Adopting a smaller prediction time window for the severely fluctuating segments and configuring differential prediction parameters for different data segments.

5. The method for predicting the state of a marine BMS based on machine learning according to claim 1, characterized in that, The constructing a multivariate grey prediction model considering the influence of temperature stratification includes: Constructing a grey differential equation: ； Among them is the development coefficient, representing the driving parameter for the internal development of the system, is the driving coefficient, representing the influence degree of the th factor on the system, represents the derivative with respect to time , represents the accumulated sequence of the variable to be predicted, represents the accumulated sequence of the th factor, represents the total number of influencing factors, represents the summation symbol; Introducing a temperature stratification compensation function and dynamically adjusting the model parameters according to the temperature stratification index TLI: ； ； Among them represents the adjusted development coefficient, is the development coefficient, is the sensitivity coefficient of temperature stratification to the development coefficient, represents the th adjusted driving coefficient, is the original th driving coefficient, is the sensitivity coefficient of temperature stratification to the driving coefficient, is the temperature stratification index.

6. The method for predicting the state of a marine BMS based on machine learning according to claim 5, characterized in that, It also includes establishing a multi-layer recursive residual correction system, and the steps are: Calculating the residual sequence between the predicted value and the actual value; Applying the grey prediction model to the residual sequence again to obtain the residual predicted value; Adding the original predicted value and the residual predicted value to obtain the corrected prediction result.

7. A method for predicting the state of a marine BMS based on machine learning according to claim 1, characterized in that The steps of dynamically selecting the optimal prediction model combination include: Calculating the prediction error evaluation function considering the influence of temperature stratification: ； Among them is the model for the prediction value of the data at the time point . is the actual value at the time point . is the temperature stratification influence coefficient is the evaluation time window length is the prediction error considering the influence of temperature stratification represents the temperature stratification index represents the model index represents the data set index represents the summation symbol; Selecting the optimal prediction model or model combination according to the error evaluation function.

8. A method for predicting the state of a marine BMS based on machine learning according to claim 7, characterized in that It also includes dynamically determining the fusion weight: ； where is the prediction error of the model within the most recent evaluation window, is the total number of models, is the model fusion weight, represents the model index, represents the th model's prediction error within the most recent evaluation window, represents the summation symbol; In addition, the weight can be further adjusted according to the sea state level and temperature stratification index: ； Among them is the fusion weight of the adjusted model , is the fusion weight of the model ; represents the temperature stratification index; represents the sea state level; and are the sensitivity coefficients of the model to the sea state and temperature stratification, which are obtained through historical data analysis.

9. A method for predicting the state of a marine BMS based on machine learning according to claim 1, characterized in that The final prediction result includes the predicted values of the main state parameters of the BMS and the visual output of the prediction information.

10. A system for predicting the state of a marine BMS based on machine learning, characterized in that A marine BMS state prediction method based on machine learning for executing any one of claims 1-9, including: A data acquisition and processing unit for collecting and preprocessing ship BMS parameter data and environmental data; A temperature field modeling unit for constructing a battery compartment temperature field model and calculating the temperature stratification index; An adaptive segmentation unit for automatically dividing data segments according to the BMS data fluctuation characteristics and configuring differential prediction parameters; A grey prediction model unit for constructing a multivariate grey prediction model considering the influence of temperature stratification and dynamically adjusting the model parameters according to the temperature stratification index; A multi-model fusion unit for dynamically selecting and fusing prediction models based on the current sea state and temperature stratification state and generating the final prediction result.

Images