40 results about "Mean absolute percentage error" patented technology

The invention discloses a battery capacity prediction and state evaluation method and system based on multi-model collaborative learning, and belongs to the technical field of battery management. The method comprises the following steps: constructing a database containing multiple lithium ion battery long-term cycle data, and classifying according to a capacity attenuation trend; cleaning and preprocessing short-term cycle data of the to-be-tested battery; matching the to-be-tested data with the long-term attenuation trend in the database by using a clustering algorithm, and determining an optimal matching trend; distributing weights for the data in the matching trend by adopting a correlation algorithm, and generating initial capacity attenuation prediction; performing sequence correction on the preliminary prediction in combination with meta-learning and a related model, and generating a smooth future attenuation trend conforming to a physical law; and outputting a capacity prediction and health state evaluation result, and evaluating the prediction precision through a root-mean-square error and an average absolute percentage error. The method significantly improves the precision and generalization ability of long-term capacity prediction, and is suitable for various scenes such as electric vehicles, energy storage systems, consumer electronics and the like.

The invention discloses a public building cold load short-time prediction method fusing physical information, and the method comprises the steps: collecting and preprocessing the historical cooling capacity, indoor environment, outdoor weather and equipment operation state data of a public building at a fixed time interval, and obtaining multi-dimensional input features; respectively establishing a workday sub-model and a holiday sub-model according to workday and holiday scene division; the workday sub-model and the holiday sub-model jointly form a cold load prediction model, the workday sub-model adopts a long short-term memory (LSTM) network, and the holiday sub-model adopts a light gradient elevator (Light GBM); a physical constraint loss function based on building energy balance and heat conduction residual error is introduced in the training process, and the physical constraint loss is fused into a total loss function according to a weighting coefficient so as to constrain that the output of each sub-model accords with the law of energy conservation and thermal inertia; monitoring the prediction error MAPE in real time and performing online calibration; and outputting a short-time cold load prediction result.

The invention discloses a key parameter long-time-sequence prediction method for a complex process industry, and the method comprises the steps: collecting multivariable sensor data in the process industry, and constructing a high-dimensional long-time-sequence prediction data set; constructing a PatchConvRNN prediction model by combining time slice embedding, dimension decoupling convolution, depth separable convolution and a recurrent neural network based on a sequence-to-sequence normal form; a point value-statistical mixed loss function is adopted, the point prediction precision, the sequence mean value and the standard deviation consistency are optimized at the same time, a prediction model is trained in combination with an optimization algorithm, and network model parameters are adjusted; and comprehensively evaluating the prediction model through a root-mean-square error, an average absolute percentage error and a standard deviation average absolute error. According to the method, high-precision prediction and fluctuation maintenance of the key time sequence variables under the complex working condition of the industrial process are achieved, and powerful support is provided for quality control and predictive maintenance of the production process.

The invention provides a traffic flow prediction method combining a dynamic graph neural network and a Mama mechanism. The future short-term traffic flow is predicted by using historical traffic data. According to the method, firstly, normalization preprocessing is carried out on traffic state data collected by multiple sensors, a training sample is generated by adopting a sliding window, and a traffic flow value in the next one hour is predicted according to data in the past 24 hours. On the basis of the model structure, a time modeling module composed of multiple layers of MambaBlocks is constructed and used for capturing historical time sequence dependence; constructing a spatial modeling module of dynamic graph convolution, and combining a static adjacency matrix of the road network with a learnable adaptive adjacency structure to extract spatial association; and finally, the outputs of the modules are fused, and a prediction result is obtained through a prediction output module. In the training process, a Huber loss function is used as an optimization target, and evaluation indexes such as a mean absolute error (MAE), a root mean square error (RMSE) and a mean absolute percentage error (MAPE) are used for evaluating the performance of the model. According to the method, the traffic space-time dynamic characteristics are effectively mined, and the long-range dependence modeling capability and the prediction precision are improved.

The invention discloses a tunnel ventilation system and a control method thereof, and relates to the technical field of tunnel ventilation energy-saving control. According to the method, traffic data, environment data and equipment starting data in a tunnel are preprocessed and divided into data sets, a traffic prediction model based on LSTM and an environment prediction model based on full connection are constructed, and training is carried out through a root-mean-square error and an average absolute percentage error; and sequentially predicting traffic and environment data by using the model, and solving an optimal ventilation equipment combination through a sequential quadratic programming algorithm by taking energy consumption minimization as a target and environmental standard reaching as a constraint, and adjusting operation. The method has the advantages that advanced regulation and control are achieved through the two-stage prediction model, pollutants are prevented from exceeding the standard, and the air quality is guaranteed; dynamic training, multi-parameter optimization and a fault tolerance mechanism are combined, energy conservation and environment regulation and control are accurately balanced, meanwhile, the operation reliability and the intelligent level of the system are improved, and manual intervention is reduced.

According to the logging-while-drilling data lag compensation method based on deep learning provided by the invention, the prediction precision of the lag logging-while-drilling data is improved by combining the drilling engineering data and the historical logging-while-drilling data. According to the method, an encoder-decoder neural network model based on a self-attention mechanism is utilized, lag logging-while-drilling data is predicted in combination with drilling engineering features, and prediction precision is improved through a convolutional layer. Results show that when the method is used for predicting the gamma curve in the logging-while-drilling data, the root mean square error within the prediction range of 30 meters is lower than 0.589, and the average absolute percentage error is lower than 1.587. The problem of lag of the logging while drilling data is effectively solved, the operation risk under the complex stratum condition is reduced, the drilling strategy can be optimized in real time, and the safety and efficiency of the drilling process are improved.

The invention relates to the technical field of lithium ion battery prediction and health management, and discloses a lithium ion battery early life prediction method for a marked data scarcity scene, and the method comprises the steps: extracting the degradation characteristics of a preset dimension through the charging and discharging data of the first 70 cycle periods of a lithium ion battery, and carrying out the subsequent prediction; training a plurality of prediction models by using the labeled samples, and obtaining pseudo labels of the unlabeled samples by using the trained prediction models; clustering the unlabeled samples to obtain corresponding cluster labels; calculating a confidence score of each unlabeled sample based on the pseudo labels and the cluster labels, and selecting the unlabeled samples with high scores and the corresponding pseudo labels to add into an initial training set; and meanwhile, through the change of the average absolute percentage error of the training set, a preset confidence threshold and a preset sampling proportion are dynamically adjusted, and the training set is updated, so that a more accurate target prediction model is trained, and the early life of the lithium ion battery is predicted.

The invention discloses a flight landing time prediction method based on machine learning, and relates to the technical field of flight pre-judgment, and the method comprises the steps: carrying out the preprocessing of data based on historical flight data and historical meteorological data, and outputting feature sample data; performing data cleaning and standardization processing on the feature sample data, and outputting a training data set and a test data set in combination with actual landing and planned landing moment information of historical flights; establishing a landing prediction model, and performing optimization by combining reinforcement learning and a generative adversarial network; using the average absolute percentage error to test the landing prediction model and output a trained landing prediction model; and outputting a dynamically updated landing prediction model by adopting a concept drift detection algorithm, online gradient lifting and adaptive weighted fusion. Through combination of the deep Q network and multi-agent reinforcement learning, the prediction result is optimized in real time, and dynamic adjustment can be performed according to changes of actual flights.

The invention relates to a Nusselt number prediction method, program and equipment based on a heat transfer associated hybrid model of a shell-and-tube heat exchanger and a storage medium, and belongs to the technical field of intelligent design and performance optimization of industrial heat exchange equipment. The objective of the invention is to solve the problems of low prediction precision, poor adaptability and lack of physical interpretability of a traditional empirical correlation under complex working conditions. Through an adaptive weight mechanism, the final hybrid model can automatically select an optimal prediction strategy according to different working conditions, and in four typical working condition cases, the average decision coefficient is up to 0.9930, and the average absolute percentage error (MAPE) is only 1.66%, which is significantly superior to that of a single model. According to the system, high-precision and high-robustness prediction of the heat transfer performance of the shell-and-tube heat exchanger is achieved, more importantly, the corrected correlation formula has clear physical significance and high interpretability, and a revolutionary tool is provided for intelligent design, real-time performance monitoring and energy efficiency optimization of the heat exchanger.

The invention relates to the technical field of ocean remote sensing, in particular to an offshore water quality remote sensing inversion and classification method based on small sample deep learning, which comprises the following steps: acquiring offshore in-situ observation, satellite remote sensing and ocean reanalysis data; carrying out atmospheric correction, carrying out space-time reconstruction on the missing remote sensing reflectivity by adopting a data interpolation empirical orthogonal function, carrying out mathematical transformation on input features, and screening out an optimal feature subset by adopting a strategy based on an average absolute percentage error; constructing a deep learning network based on a generative small sample to invert the offshore soluble inorganic nitrogen and inorganic phosphorus concentration, and determining the water quality grade; and analyzing an inversion mechanism by using an SHAP method. According to the method, the problem of data space-time discontinuity is solved through the data interpolation empirical orthogonal function, the capture capability and inversion precision of the nonlinear relation under the small sample condition are remarkably improved by utilizing the generative network, the model interpretability is realized in combination with SHAP analysis, and scientific support is provided for offshore water quality fine management.

The invention relates to a numerical control machine tool reliability modeling method and system considering a fault trend. The method comprises the following steps: checking the homogeneity of a numerical control machine tool based on the fault trend; constructing an AMSAA model of the multiple samples subjected to truncation at unequal time; evaluating the precision of the reliability model; aiming at the problem of insufficient reliability model precision of a multi-sample condition of a numerical control machine tool, a multi-sample AMSAA model modeling method of the numerical control machine tool considering a fault trend is provided, sample data classification is carried out through trend inspection and homotype inspection, and fault data of same-type machine tools are preprocessed by adopting a fault total time method; adopting a maximum likelihood method to estimate AMSAA model parameters of the same-type machine tool, and using Cramer-Von Mises to test the goodness of fit of the model; evaluating the precision of the reliability model by taking an average absolute percentage error (MAPE) of an instantaneous MTBF point estimation value and an MTBF observation value as an index; compared with a multi-sample AMSAA model established by a direct maximum likelihood method, the method is higher in prediction precision.

The present application belongs to the technical field of photonic device design, and specifically relates to a micro-ring resonator reverse design optimization method based on deep learning. The present application comprises: obtaining different micro-ring resonator structure parameters and corresponding free spectral range and quality factor, and processing and dividing the data set; building a cascade neural network model, wherein the reverse network takes the target performance index as the input to predict the structure parameter, the forward network inputs the structure parameter and outputs the performance index, and is used to constrain and correct the reverse prediction result; after training, the test set is evaluated to measure the reverse effect by the average absolute percentage error index, the weight in the loss function is changed, and multiple rounds of training are performed to determine the optimal weight configuration and final model of the reverse design. The present application has the advantages of high calculation efficiency, excellent design precision and strong physical realizability, can reduce the dependence on large-scale simulation and multiple iteration optimization, effectively shorten the device design cycle, and has high application value.

The present invention discloses a configuration generation method of an adjustable phononic crystal based on deep learning, a computing device, a computer program product and a storage medium. The configuration generation method of an adjustable phononic crystal based on deep learning is executed in a computing device. The method includes: inputting a first eigenfrequency vector indicating the dispersion relation of the phononic crystal configuration to be generated into a configuration generation model for processing to obtain at least one candidate configuration image; performing image preprocessing on each candidate configuration image to obtain multiple similar images; inputting each similar image into a dispersion relation prediction model for processing to obtain a second eigenfrequency vector representing the dispersion relation of each similar image; calculating the average absolute percentage error value of the first eigenfrequency vector and each second eigenfrequency vector respectively, and selecting the similar image indicated by the second eigenfrequency vector corresponding to the smallest average absolute percentage error value as the configuration image of the phononic crystal to be generated.

The application discloses a cold-rolling friction coefficient prediction method based on a physical graph topology and a graph convolutional neural network, and comprises the following steps: defining a graph topology structure establishment rule for friction coefficient analysis in plate strip cold continuous rolling; selecting analysis parameters according to plate strip cold continuous rolling process characteristics and production experience, determining a node set, a physical relationship function set and an edge set in the graph topology structure; constructing an adjacency matrix and a corresponding self-adjacency matrix of the graph convolutional network with physical weights; introducing a degree matrix to perform normalization processing on the self-adjacency matrix to generate a normalized self-adjacency matrix; constructing a physical feature graph convolutional neural network based on the above settings and a basic form of the graph convolutional network; selecting cold-rolling product samples, obtaining analysis parameters and friction coefficient prior values, and training the physical feature graph convolutional neural network; and adopting a determination coefficient, a mean square error, a mean absolute error and a mean absolute percentage error to evaluate the performance of the physical feature graph convolutional neural network.

The invention discloses an eggshell spot detection method based on multi-scale feature fusion and an attention mechanism, and relates to the technical field of eggshell spot detection, and the method comprises the steps: collecting an eggshell surface image, carrying out the data labeling, generating a spot region binary mask, and calculating a spot area proportion true value; carrying out data preprocessing on the image, wherein the data preprocessing comprises size adjustment, data enhancement and standardization; an end-to-end deep convolutional neural network model is constructed, the model extracts and fuses multi-scale features through a texture feature encoder, key region features are enhanced through a speckle attention module, and a speckle area proportion prediction value is output through a regression prediction module; training the model, and optimizing parameters by adopting a smooth average absolute percentage error loss function; and inputting a to-be-detected eggshell image to the trained model, and outputting a spot area proportion evaluation result. According to the invention, automatic, high-precision, high-efficiency and lightweight prediction of the eggshell spot area proportion is realized.

The present application relates to the technical field of lithium ion battery prediction and health management, and discloses a lithium ion battery early life prediction method for a labeled data scarce scene, which comprises the following steps: using the charging and discharging data of the first 70 cycle periods of the lithium ion battery, extracting degradation features of a preset dimension, and performing subsequent prediction; using labeled samples to train a plurality of prediction models, using the trained prediction models to obtain pseudo-labels of unlabeled samples; clustering the unlabeled samples to obtain corresponding cluster labels; based on the pseudo-labels and the cluster labels, calculating the confidence score of each unlabeled sample, and selecting the unlabeled sample with a high score and its corresponding pseudo-label to add to the initial training set; and simultaneously, by the change of the average absolute percentage error of the training set, dynamically adjusting the preset confidence threshold and the preset sampling ratio, updating the training set, and training a more accurate target prediction model to predict the early life of the lithium ion battery.

The invention discloses a part quality dynamic early warning system based on a hierarchical federal learning framework, and the system comprises a part quality dynamic early warning platform which comprises a data processing unit, a part quality prediction unit and a dynamic early warning unit, and the data processing unit is used for preprocessing data; the part quality prediction unit is used for performing part life prediction by utilizing a CNN-Informer model on the basis of the data preprocessed by the data processing unit, and optimizing model hyper-parameters by taking a minimum symmetric average absolute percentage error SMAPE as a target function and adopting an improved big sugarcane-mouse algorithm; the dynamic early warning unit is used for deciding whether to give an early warning or not based on the prediction result; the system security protection platform is used for providing computing support through the cloud computing platform, verifying a user and guaranteeing data security; the enterprise user interaction platform is used for receiving the early warning result of the dynamic early warning unit and giving user feedback suggestions; the method can effectively improve the comprehensiveness and accuracy of part life prediction.

The invention relates to a digital data processing technology of a power management system, in particular to a power demand analysis method and system under multi-source data fusion. The analysis method comprises the following steps: collecting and converging multi-source heterogeneous data; performing data fusion and standardization processing; carrying out multi-dimensional scene load characteristic modeling and trend analysis; performing layered and classified electricity demand prediction; and outputting a hierarchical classification prediction report. The analysis system comprises a multi-source data acquisition module, a data fusion processing module, a feature modeling and analysis module, a hierarchical prediction and optimization module and a visual display module. Compared with the prior art, the method has the advantages that a complete power utilization analysis data panorama is constructed, and the load feature recognition integrity is improved by more than 40%; the average absolute percentage error MAPE of short-term load prediction can be reduced to be within 3% and is improved by about 35% compared with a general modeling analysis method; and the efficiency of converting the prediction result into the action can be improved by 50%.

This invention proposes a ship heave prediction system and method based on a Conv-Bi-LSTM model. The method involves acquiring historical ship motion data, establishing a motion information dataset, preprocessing the dataset, randomly splitting it into training and test datasets, constructing a ship heave prediction model using a Conv-Bi-LSTM model, training the Conv-Bi-LSTM network with the training dataset, and inputting the test dataset into the ship heave prediction model for prediction. The prediction accuracy of the test data is then obtained, and the model's performance is evaluated using mean error (MAE), mean square error (MSE), root mean square error (RMSE), and mean absolute percentage error (MAPE). This invention uses historical motion information with multiple degrees of freedom as input to the model for comprehensive prediction of ship heave, fully utilizing the forward and reverse time state information of the motion time series, thus improving the prediction accuracy of heave motion.

This invention discloses a tunnel ventilation system and its control method, relating to the field of tunnel ventilation energy-saving control technology. The method first preprocesses and divides the tunnel's traffic data, environmental data, and equipment operation data into datasets. It then constructs an LSTM-based traffic prediction model and a fully connected environmental prediction model, trained using root mean square error and mean absolute percentage error, respectively. The models sequentially predict traffic and environmental data. With energy minimization as the objective and environmental compliance as the constraint, a sequential quadratic programming algorithm is used to solve for the optimal combination of ventilation equipment and adjust its operation. Its advantages lie in the two-level prediction model's ability to achieve proactive regulation, preventing pollutant exceedances and ensuring air quality. Combined with dynamic training, multi-parameter optimization, and fault tolerance mechanisms, it accurately balances energy saving and environmental control while improving system reliability and intelligence, reducing manual intervention.

The invention relates to the technical field of machine learning, and discloses a training method for an electrical performance prediction model of a semiconductor tape-out, and computer equipment, and the method comprises the steps: obtaining training data which comprises the technological parameters of a plurality of semiconductor tape-out pieces and the real electrical performance parameters of the semiconductor tape-out pieces; and inputting the training data into a preset electrical performance prediction model, and training the electrical performance prediction model by using a fusion loss function composed of a mean square error, a mean absolute error and a mean absolute percentage error to obtain the trained electrical performance prediction model. The method has the beneficial effect that the prediction accuracy and robustness of the electrical performance prediction model are improved.

The application provides a laser cladding process molten pool morphology information prediction method and system based on VMD and improved DA-RNN, comprising the following steps: S1: continuously collecting molten pool morphology information in the molten pool dynamic video to obtain molten pool morphology information data; S2: pre-processing the molten pool morphology information data by VMD (variational mode decomposition) to decompose the original data into K different subsequences; S3: taking the decomposed different subsequences as data characteristic variables to construct a VMD data set, and dividing the VMD data set into a training set, a validation set and a test set; S4: establishing a DA-RNN prediction network model, and training and predicting on the data set; S5: quantitatively evaluating the DA-RNN prediction network model by using mean absolute percentage error and mean square error; and S6: predicting the laser cladding molten pool morphology information by using the prediction network model. The application can not only effectively extract data characteristics and reduce the complexity of data, but also effectively improve the molten pool morphology prediction model precision.

The invention relates to the technical field of new energy vehicles, and particularly discloses a new energy vehicle SOC estimation system and method based on a fractional order neural network, and the method depends on the fractional order calculus characteristics of the fractional order neural network, a fractional order operator has a memory effect, and can accurately capture the historical change correlation of voltage, temperature and voltage drop in the charging and discharging process of a battery. According to the method, the fitting precision of the dynamic attenuation process of the battery after long-term circulation can be improved, error accumulation caused by neglecting a historical state is avoided, meanwhile, the lagging response characteristic of the battery can be quickly responded, and the adaptability to time-varying nonlinearity is higher through dynamic adaptation of fractional order numbers; after the fractional order neural network is iterated by the sparrow search optimization algorithm, the weight, the threshold and the order parameter can be accurately matched with the battery characteristics, a dynamic voltage drop characteristic weight adjustment mechanism is matched, the average absolute percentage error of SOC prediction is reduced, prediction precision reduction caused by battery aging is avoided, and the SOC estimation precision is improved.

The application discloses a cloud server performance degradation prediction method based on time sequence segmentation, first extracts performance resource time sequence data on a cloud server, decomposes the obtained time sequence data by adopting a DTW-BU time sequence segmentation algorithm, respectively constructs LSTM models for segmented subsequences, and predicts cloud server resource time sequence data, verifies model precision by using root mean square error and average absolute percentage error, predicts system performance degradation trend according to time sequence prediction data of the LSTM model, tests data fitting degree, and determines a software regeneration time node according to prediction data threshold value; the application can improve the precision of cloud server performance degradation prediction results, avoid overfitting phenomenon in the prediction process, and solves the problem of how to perform software regeneration at an optimal time point for cloud server performance degradation.

This invention belongs to the field of carbon emission prediction technology, specifically involving a dynamic correction method, system, and storage medium based on carbon emission prediction. First, a carbon emission prediction model is pre-trained using historical and associated feature data and deployed to online prediction nodes. Then, prediction-measured data pairs are collected, and the real-time prediction relative deviation is calculated by correcting the mean absolute percentage error. A multi-factor dynamic deviation threshold formula with boundary constraints is used to determine whether correction is triggered. After correction is triggered, an incremental training dataset is constructed. Features are jointly filtered using SHAP values ​​and Pearson correlation coefficients to complete lightweight adaptive updates. Then, incremental training with bottom-level freezing and top-level adaptive learning rate fine-tuning yields the corrected model. Finally, the corrected model is generalized and validated. If effective, the online model is updated, and the basic threshold and weight coefficients of the dynamic deviation threshold formula are iterated.

The invention discloses a thermal power plant demand prediction method based on holt-winters. The method comprises the steps that S1, a sales settlement data matrix in a preset historical period is acquired, rows in the matrix represent years, and columns in the matrix represent months; s2, respectively substituting the historical data in the sales settlement data matrix into a holt-winters addition model and a multiplication model; s3, traversing all historical data in the sales settlement data matrix, continuously fitting values of three hyper-parameters including a horizontal item, a trend item and a seasonal item in the holt-winters model, and determining an optimal parameter combination according to the hyper-parameter corresponding to the minimum average absolute percentage error in the model fitting process; and S4, predicting the electric quantity demand and the heat demand of the target thermal power plant in a future preset month according to the optimal combination. According to the prediction method, the trend and seasonal characteristics of the electric heating demand are captured at the same time, the limitation of single historical data reference in the same period is solved, and reliable data support is provided for production scheduling plan making.

This invention belongs to the field of coal mine safety production warning technology and discloses a coal and gas outburst early warning method based on real-time on-site data. By continuously acquiring real-time environmental data on gas and wind speed through gas concentration sensors and wind speed sensors installed at the tunneling face, the gas emission rate is calculated. In-depth analysis of the gas emission rate yields early warning indicators for coal and gas outbursts, combining dynamic changes in gas concentration with reverse identification based on gas emission rate prediction and evaluation indicators. Moving average, deviation rate, dispersion rate, volatility, root mean square error, and mean absolute percentage error early warning indicators are established. The weights of each indicator are determined using the analytic hierarchy process (AHP), a fuzzy comprehensive early warning model is constructed, and a comprehensive index of the differences among the indicators is calculated for coal and gas outburst early warning. This invention solves the problem of existing methods' difficulty in real-time data acquisition, which hinders effective on-site decision-making.

This invention relates to the field of tourist volume interval prediction technology, specifically a multi-objective tourist volume interval prediction method and system based on frequency mixing. The prediction method uses weekly tourist volume intervals as the prediction objective, and daily search index, AQI, and rainfall as high-frequency influencing variables. A standardized variable matrix is ​​constructed through frequency mixing data alignment. The matrix is ​​then reconstructed into low, medium, and high-frequency feature subsequences using wavelet packet decomposition combined with sample entropy. A multi-type candidate model library is then built. The optimal sub-model for each subsequence is selected based on adaptive weighted interval evaluation indicators and 5-fold cross-validation. Finally, a sparrow optimization algorithm is used to search for the optimal weights by minimizing the interval mean absolute percentage error and root mean square error as dual objectives. The prediction results of the optimal sub-models are weighted and fused to obtain the final value, thereby effectively preserving high-frequency data features, achieving multi-scale feature decoupling of the sequence, and significantly improving the accuracy and reliability of tourist volume interval prediction.

The application discloses a ship motion posture prediction method based on multiple combinations, selects hidden layer node number, initial learning rate and maximum training times as hyperparameters of a neural network to be optimized, determines initialization parameters and fitness functions of a whale optimization algorithm, trains the GRU neural network by using the whale optimization algorithm, obtains an optimized WGRU neural network model, decomposes original ship posture data by using complete set empirical mode decomposition, trains the WGRU network by using the decomposed data, and obtains a CWGRU combined prediction model; the prediction result of the CWGRU combined prediction model is minimum in terms of root mean square error and mean absolute percentage error, and the correlation coefficient is the highest; since the CEEMD algorithm is used to decompose the original ship posture data, the algorithm has better prediction effect on the ship motion posture with non-stationary and nonlinear characteristics.

The application discloses a soybean protein gelation multi-step prediction method based on ultrasound. The ultrasound waveform of the gelation process is organized into a time sequence, and a sequence-to-sequence neural network model based on an attention mechanism is combined. The original waveform of all sampling points is adaptively weighted through a gated attention mechanism, the encoder hidden state is aggregated through an additive attention mechanism with a linear proximate bias, and information is dynamically retrieved from the encoding history for each future prediction step through a scaled dot-product cross-attention mechanism. The model outputs the gelation degree prediction value of the next 5 time steps through self-recurrence, and the gelation endpoint is determined in advance accordingly. Under leave-one-out cross-validation, the overall average accuracy of the method to the unseen gelation conditions reaches 95.72%, and the average absolute percentage error in short-term prediction in the later stage of gelation is as low as 1.02%, which provides intelligent technical support for real-time tracking and early prediction of the gelation endpoint in the industrial production of bean products.