59results about How to "Improve forecasting performance" patented technology

The invention discloses a medium and long term hydrologic forecasting method based on empirical mode decomposition. The medium and long term hydrologic forecasting method based on empirical mode decomposition comprises the steps that firstly, a hydrologic forecasting model is built according to the following procedures of (101) empirical mode decomposition, wherein empirical mode decomposition is carried out on a hydrologic time sequence s(t) of a forecast drainage basin, (102) kernel principle component analysis, wherein kernel principle component analysis is carried out on n intrinsic mode function components Fj obtained through empirical mode decomposition and a trend item rn, and p main components F'k are extracted, (103) building of a training sample set, wherein the training sample set is built according to the extracted p main components F'k, (104) building of a support vector machine model, and (105) training of the support vector machine model; secondly, annual runoff data of the years needing forecasting are forecast through the built hydrologic forecasting model. The medium and long term hydrologic forecasting method based on empirical mode decomposition is simple in step, convenient to realize, easy and convenient to operate, good in use effect and capable of effectively resolving the problem of low forecasting accuracy of an existing hydrologic forecasting method.

The invention belongs to the technical field of data processing, and discloses an LSTM neural network cycle hydrological forecasting method based on mutual information, which comprises the following steps: screening and classifying original data through mutual information analysis, and taking rainfall, reservoir water level and flow hydrological characteristics as input characteristics of a long-term and short-term memory cycle forecasting model; the long-term change of flood is reflected by simulating rainfall process training and determining the structure of the LSTMC model; and verifying the output of the model by using the actual flood data. According to the method, the data set is analyzed by adopting a mutual information-based method, the flow at the current moment and each hydrological characteristic of the previous longer time period are fully captured, and the input characteristics of the model are dynamically selected. According to the method, the deep learning algorithm is utilized, the cyclic prediction model based on the LSTM neural network is adopted, when the method is used for flood flow time series prediction, the problem that the hydrological change process is greatly influenced by factors in the earlier stage is solved, and effective features can be automatically captured well.

The present invention discloses a climate change prediction method and system for fitting various climate modes based on a modified BP neural network. The method comprises the following steps: collecting rainfall data that characterizes a climate change feature; performing downscaling processing on climate mode prediction grid data, so as to obtain a long series of data with the same resolution as measured data; performing average correction on a downscaled climate mode prediction series; establishing a modified BP neural network model and determining a network model structure; training and inspecting the established BP neural network model; and predicting a future climate spatial-temporal change using a model output. According to the method and system disclosed by the present invention, output / input data processing of the BP neural network is modified, and various climate modes are fit to predict the future climate change, so that reliability of the prediction result is increased, and the defects that the deviation of a single mode is relatively large during an assessment and that basin space nonuniformity is not considered during a multi-mode collection assessment and the like are made up.

The invention discloses a transient disturbance weather map and low-frequency disturbance weather map manufacturing method based on atmospheric variable physical decomposition and application of the method in weather forecast. The method includes that formerly observed, recently observed and medium-term numerical weather predication modes are used for output atmospheric space three-dimensional multi-variable and time continuous grid point data, climate components and synoptic scale disturbance components of daily cycle and annual cycle are obtained through physical decomposition, and the synoptic scale disturbance components are used for manufacturing a transient disturbance weather map and a low-frequency disturbance weather map. According to the method, after climate seasonal variation field and day-by-day synoptic scale transient disturbance field physical decomposition is performed on meteorological data of the Northern Hemisphere in the past 30 years, synoptic scale transient disturbance components predicated through the recently observed and medium-term numerical weather predication modes are used for drawing a transient disturbance weather map and a low-frequency disturbance weather map. The transient disturbance weather map is suitable for manufacturing 1 to 3 days short-term weather forecast and 4 to 9 days medium-term weather forecast, the low-frequency disturbance weather map is suitable for manufacturing medium-term and 10 to 30 days elongating stage weather process forecast. A novel tool for weather forecast is provided, and accuracy of the weather forecast can be improved.

The invention relates to the technical field of photovoltaic power generation, and in particular to a radiation attenuation-considered photovoltaic power short-term prediction method. The method comprises the following steps of: carrying out training by adoption of an indirect prediction method so as to obtain a sunny day surface radiation prediction model; obtaining an attenuation coefficient ofhistory daily surface radiation, establishing a surface radiation attenuation coefficient prediction model according to the attenuation coefficient, and establishing a cloud coverage coefficient prediction model; carrying out training by taking history real surface radiation, temperature and humidity as inputs of a meteorological factor and taking photovoltaic power as an output, so as to obtain aphotovoltaic power prediction model; and predicting photovoltaic power generation power by utilizing the photovoltaic power prediction model by taking a predicted value of the surface radiation as input of a surface radiation value and taking meteorological data of weather forecast as input of the meteorological factor. According to the method, the influence degree of cloud on the prediction precision can be reduced, the process of image analysis is saved and the algorithm is simpler and more efficient.

The invention discloses a marine environmental prediction product parallel processing method applied to tiled map services. The method comprises the following steps that (1) the longitude and latitude of metadata of a marine environmental prediction product are read, a WGS84 coordinate is converted into a WebMercator projection coordinate; (2) a data depth layer is cut into n parts according to the kernel number of a parallel machine, an MATLAB parallel strategy is designed; (3) a map grading number is set; (4) according to the current map level interpolation, an MATLAB drawing instruction is used for generating graded section files; (5) drawing of all parameters of the current map grade is completed; (6) an LOD mode is used for organizing the section files, and drawing of all parameters of all grades is completed; (7) after all parallel computers complete processing, metadata are generated. The marine environmental prediction product parallel processing method applied to tiled map services has a good application prospect in the informatization research field of marine environmental sciences. According to the method, a parallel computing technology is used for guaranteeing quick processing of marine environmental prediction data, and the release cycle of the tiled map services is shortened.

The invention provides a feature extraction hydrological forecasting method based on deep learning, and belongs to the field of water resource efficient utilization and hydrological forecasting. The method comprises the following steps: firstly, obtaining a watershed hydrological forecasting characteristic factor set by utilizing watershed historical information; secondly, training the characteristic factor set by utilizing a data mining algorithm, and obtaining a plurality of groups of session flood process sets with similar magnitudes and process forms under the action of different factors;then, carrying out parameter calibration of all models and methods in traditional hydrological forecasting based on a deep learning algorithm, forming a model library and a method library matched withthe models, the methods and parameter schemes, and finally completing hydrological forecasting calculation in combination with clustering analysis. Compared with an existing method, the method has the advantages that the defects that a traditional hydrological forecasting method is low in forecasting precision, short in effective forecasting period and the like are effectively overcome, the forecasting precision can be obviously improved and the forecasting period can be obviously prolonged when hydrological forecasting is carried out, good applicability and feasibility are achieved, and an effective technical method is provided for basin hydrological forecasting.

The invention provides a data processing method for forecasting heavy precipitation weather. The data processing method for forecasting the heavy precipitation weather comprises the steps that firstly, thermodynamic factors, dynamic factors and a moisture index influencing occurrence of heavy precipitation are collected; secondly, according to the precipitation occurrence and development nonlinear theory, the thermodynamic factors, the dynamic factors and the moisture index are integrated, so that the physical quantity index THP is obtained; finally, according to the synoptic situations at 500hPa and at 700hPa of upper air, whether a good lifting mechanism exists or not is judged, and then the belt of the heavy precipitation and the magnitude of the heavy precipitation can be forecasted, wherein the area with a large value of the index THP is the large-probability area of precipitation in next six hours, the maximum value of the THP corresponds to the maximum amount of precipitation in next six hours, when the value of the THP is increased along with time, the amount of precipitation is increased, and otherwise the amount of precipitation is reduced. By the adoption of the data processing method for forecasting heavy precipitation weather, the belt of heavy precipitation and the development and change of precipitation can be more objectively forecasted, the magnitude of precipitation can be estimated quantitatively, and automatic operation can be achieved through a computer.

The invention discloses a bus station passenger flow volume prediction method based on a GCN-LSTM combination model, and belongs to the field of traffic prediction. According to the method, spatial correlation of passenger flow volume data is extracted by using a graph convolutional neural network, the data with spatial correlation information is input into a long-term and short-term memory cycleunit for time feature extraction, and finally, a prediction result is obtained through a full connection layer. According to the method, the spatiotemporal features of the traffic data are simultaneously concerned, so that the defect that the spatiotemporal features of the traffic data cannot be fully extracted by the existing method is overcome, the passenger flow volume data prediction precisionis improved, and the problem of short-time passenger flow volume prediction can be better solved.

The invention discloses a method for forecasting the temperature of a slab of a hot rolling heating furnace, comprising the following steps of: firstly, carrying out simplified hypothesis on heat transfer of radiation; secondly, drawing a grid division diagram on a one-dimensional unsteady heat transfer problem, dividing a billet into 8 layers i.e. 9 nodes, and obtaining explicit difference equations with 9 nodes by utilizing a one-dimensional unsteady difference equation; and thirdly, determining an unknown thermal property coefficient f by adopting a genetic algorithm, and forecasting a temperature value of a next time by the explicit difference equations based on a temperature value of each node of a last time after determining the unknown thermal property coefficient f. By adopting the invention, the billet tapping temperature in a heating furnace can be accurately forecasted, and the tediousness and the complexity of calculation can be simplified.

The invention relates to an ocean wave model prediction method based on an active disturbance rejection state observer and belongs to the data processing and prediction technology field. The method comprises steps that firstly, rapid Fourier transform of heave displacement signals at a measurement time segment is carried out, amplitude and phase of a signal corresponding to each simple harmonic wave are solved, the main frequency component quantity of the heave motion and a harmonic wave parameter corresponding to each mode are identified through amplitude peak value detection, a Kalman observer is designed, the Kalman observer is utilized to observe each measured mode simple harmonic wave and each mode simple harmonic wave of main frequency components, a harmonic wave prediction parameter corresponding to each mode simple harmonic wave of the main frequency components is estimated and updated online, the heave motion of synthetic marine equipment in the future prediction time segment is predicted based on the prediction parameters, phase rectification compensation for a wave compensation system is carried out in advance, and the prediction effect is comprehensive, strict and accurate.

The embodiment of the invention discloses an image generation method and device, an image generation model training method and device, equipment and a medium, and belongs to the field of image processing. The method comprises the steps: acquiring a first original image, first structure information corresponding to the first original image and second structure information corresponding to a secondoriginal image; inputting the first original image, the first structure information and the second structure information into an encoder to obtain an apparent feature vector and a structure feature vector output by the encoder; and decoding the apparent feature vector and the structural feature vector through a decoder to obtain a target image output by the decoder, the target image including a target object, the target object having apparent features corresponding to the first object, and the target object having structural features corresponding to the second object. The method can be used for amplifying an image set, and is beneficial for increasing the number of images of a sample set or a training set in an image key point prediction task, so as to improve the prediction effect of a key point prediction model.

The invention discloses a parameter-optimized machine learning fault diagnosis system. The system is used for fault diagnosis of the ethylene cracking process and comprises a data preprocessing module, a principal component analysis module, a machine learning module and a swarm Intelligence algorithm module. Fault diagnosis is performed on important parameter indexes in the chemical process of ethylene cracking, the defects that existing chemical fault diagnosis technologies are low in instrument prediction accuracy and are easily affected by human factors are overcome, the swarm Intelligencealgorithm module is introduced to optimize parameters of a support vector machine, so that the parameter-optimized machine learning fault diagnosis system is obtained. The system has good prediction effect under the condition of small samples and easily finds the global optimal solution.

The invention discloses a traffic flow speed predicting method and system. The traffic flow speed predicting method comprises the steps that many sample data is obtained, wherein the sample data is traffic flow speed data acquired at a moment t; a fuzzy neural network model is improved, wherein the step specifically includes sample data classification; parameters in subordinate functions of fuzzy variables in the fuzzy neural network model are determined according to a sample data classification result; a least square estimation method is adopted to optimize a rule function in the fuzzy neural network model; the improved neural network model is adopted to train the sample data, and a neural network prediction value of traffic flow speed at the moment t is obtained. The traffic flow speed predicting method and system have the advantages of high prediction precision.

The invention relates to a wind speed forecasting method and device based on G-L mixed noise characteristic kernel ridge regression technology. The method comprises the following steps: 1) obtaining wind speed data set D1; using the Bayesian principle for the loss function of Gaussian-Laplace mixed noise characteristic; 2) through the use of the theories of statistical learning and optimization and in combination with the loss function obtained in step 1), establishing the original problem of the kernel ridge regression model based on the Gauss-Laplace mixed noise; deducing and solving the dual problem of the kernel ridge regression model; 3) determining the optimal parameters of the dual problem of the kernel ridge regression model; selecting the kernel function; constructing the decision function of the kernel ridge regression model; and 4) constructing the wind speed forecasting model of the kernel ridge regression model; and using this forecasting mode to forecast and analyze the wind speed value. The device of the invention includes a loss function obtaining module, a dual problem solving module, a decision function constructing module and a wind speed forecasting module. The method and invention meet practical application in wind power generation, agricultural production, and etc. which are demanding in terms of wind speed forecasting accuracy.

The invention relates to a precipitation forecasting system for a landing tropical cyclone process, and the system comprises a generalized initial value construction module which constructs generalized initial values of a plurality of variables having an influence on a forecast amount, and transmits the generalized initial values to an initial value similarity discrimination module; the initial value similarity discrimination module discriminates the similarity of each single variable contained in the generalized initial value; sequentially calculating path similarity area indexes of the target TC path and the historical TC path in the similar region; comparing the time of the starting point of the target TC with the time when the historical TC generates rainfall to the land for the firsttime and marking the historical TC with the difference not exceeding a certain time, and comparing the intensity of the target TC with the intensity of the historical TC and marking the historical TCwith the difference not exceeding a certain intensity level; arranging the marked historical TC numbers from small to large according to the TSAI values to obtain the sequence of the marked historicalTC, and determining m optimal similarity initial values to be sent to the ensemble forecasting module; and the ensemble forecasting module acquires the corresponding forecast quantity of the optimalsimilar initial value and ensembles the forecast quantity.

The invention discloses a method for designing an advertisement click rate prediction system in a mobile application. The method comprises the following steps: carrying out integration, cleaning and transformation preprocessing on advertisement data; generating features by using the preprocessed advertisement data and constructing a feature complete set; carrying out effective feature selection onthe feature complete set, and selecting an effective feature set; and training an advertisement click rate prediction model by using the effective feature set. According to th method for predicting an advertisement click rate in a mobile application, the long-tail data in the advertisement data is classified according to the similarity and the data valuing frequency, so that the defect that information hidden in the long-tail data cannot be effectively utilized in the prior art is overcome, the information in the long-tail data is fully utilized, and the prediction effect is improved.

The invention relates to the technical field of ferrous metallurgy, and provides a converter tapping silicon-manganese alloy addition amount determination method based on yield prediction. The methodcomprises the following steps of: S1, collecting converter multi-heat production data and carrying out normalization processing; S2, determining a model input variable; S3, determining factors influencing the alloy yield in a converter tapping process and adopting the same as the input variable of a model; S4, establishing a monotonicity constrained BP artificial neural network Mn element yield prediction model; S5, adjusting model parameters to obtain an optimized prediction result; and S6, determining the predicted addition amount of converter tapping silicon-manganese alloy. According to the method, the BP artificial neural network is improved by adopting a monotonicity constraint mode, so that the BP artificial neural network can be combined with a metallurgical reaction mechanism to predict the yield of the Mn element at the smelting end point of a converter, and a prediction effect better than that of a common BP artificial neural network can be obtained; The method has good accuracy and economic benefits, and can provide beneficial guidance for the addition of alloys in the field production process.

The invention discloses a rainfall forecast displacement correction algorithm based on object diagnosis. According to the algorithm, the defects of an MODE method and a CRA method are overcome, a reasonable object displacement deviation calculation method is designed, the thought of an object diagnosis method and the thought of an observation similarity method are combined, the defects of a point-to-point diagnosis method are overcome, and the rainfall forecasting performance of the mode is directly improved from the perspective of object displacement deviation correction. Data experiments andstatistical cross tests show that the prediction performance of different prediction periods of validity after displacement correction is improved, the critical success index and the detection rate are improved, and the false alarm rate is reduced.

The invention relates to a landing tropical cyclone daily rainfall forecasting system. The system comprises: a generalized initial value construction module for receiving paths of historical TC, obtaining a forecast path of a target TC at a certain path report starting moment, and combining observation paths of the forecast path of the target TC before the path report starting moment into a targetTC path, performing processing to obtain TC moving speed information of the target TC and the historical TC in a specific day scale time period; the initial value similarity discrimination module which constructs a daily scale similar region, identifies the nearest point and the shortest distance of the historical TC, calculates the similarity degree of the historical TC path and the target TC path, marks the historical TC of which the moving speed difference with the target TC reaches a threshold value, marks the historical TC of which the shortest distance is greater than a certain threshold value, and selecting m optimal similar historical TCs to send to an ensemble forecasting module; the ensemble forecasting module which obtains and ensembles the specific daily precipitation field ofthe optimal historical TC. The system has good forecasting performance for TC daily precipitation in China.

The invention discloses a short-term wind speed combined forecasting method for a wind turbine cabin of a wind power plant. The method comprises the steps of analyzing the similarity of wind speed subsequences of a forecasted wind turbine cabin and all wind turbine cabins within a certain time period day by day by adopting a dynamic time warping method and a correlation coefficient method separately, extracting wind speed data of a plurality of subsequences with the most similar evolution, separately establishing a generalized regression neural network sub-model forecasting unit based on the dynamic time warping method and the correlation coefficient method, and globally optimizing specific parameters of each sub-model by adopting a particle swarm optimization, wherein the mean of the forecasting results of the two sub-models is used as a final forecasting result of the combined forecasting method. The method realizes fine forecasting of the cabin wind speed of each wind turbine in the wind power plant, thereby effectively improving the short-term output forecasting level of the whole wind power plant.

The invention provides a temperature ensemble forecasting method and system. The method starts from a plurality of numerical mode forecasting fields through temperature ensemble forecasting. dynamic error correction and forecast effect inspection are carried out on each forecast field according to past 15-day forecast fields and observation fields; the temperature abrupt change condition is adjusted; the weight of each mode is calculated by adopting a gradient descent method; the optimal ground temperature forecasting field is obtained through weighted average; According to the temperature ensemble forecasting method and system provided by the invention, the system error caused by the error of the numerical mode forecasting, especially the mode error under the condition of sudden temperature rise or temperature drop, is improved, a gradient descent method is used for solving the weight of each mode, so that a better forecasting effect can be obtained, and a reliable reference is provided for weather forecasting.

The invention discloses a prediction method of convection weather. The method comprises steps of according to real conditions and physical data of convection weather of many years, acquiring multiplerepresentation physical quantities and thresholds of the representation physical quantities; according to representation meanings of the representation physical quantities, carrying out grouping; dividing the grouped representation physical quantities into five grades according to the thresholds in a piecewise linearity manner; determining representation physical quantities related to weather phenomena, wherein the weather phenomena are related to the representation physical quantities; comparing the representation physical quantities of actual data with thresholds of the representation physical quantities, if meeting thresholds of all representation physical quantities of a certain weather phenomenon, predicting that the weather phenomenon will occur; and or else, predicting that the weather phenomenon will not occur. The prediction method can be applied to different value weather prediction modes; convection weather can be quite objectively predicted; and broadcasting effects are quite good.

The invention provides a credit data processing method, a credit data processing device, a storage medium and an electronic device. The method comprises the steps of acquiring the information relatedto credit repayment, wherein the information related to credit repayment comprises the asset scale of a platform, the weighting relation of a neural network model, historical loan repayment statistical information, industry loan repayment historical conditions and a future specified time period; inputting the information related to the credit repayment into the neural network model; and obtaininga first output result of the neural network model, the first output result being used for indicating a loan repayment condition of the platform in the future specified time period. According to the invention, the problems of loan return amount estimation and asset flow management based on experience and benchmark values of operation management personnel in a credit platform in the prior art are solved.

The invention discloses an ensemble prediction method for a short-term wind speed of a wind power plant. The ensemble prediction method is based on a WRF mode and a random forest algorithm. The methodcomprises the following steps: based on the WRF mode, choosing 6 different boundary layer parameterization schemes to predict meteorological factors of wind speeds, wind directions and the like at the 70m height; then utilizing varieties of boundary layers parameterization schemes to predict the wind speeds in an ensemble mode; applying the wind speeds predicted by all the single boundary layer parameterization schemes and wind speed data actually detected by an anemometer tower to the random forest algorithm to establish an ensemble prediction model; predicting the wind speed of the wind power plant. The ensemble prediction method disclosed by the invention is a scientific and effective method for predicting the short-term wind speed of the wind power plant, has the characteristics of strong generalization ability, good stability and high prediction accuracy, improves wind power prediction accuracy, is favorable for dispatch and operation of an electric power system and has certain practical value.

The invention discloses a leaf powdery mildew potential forecasting method for grapes under protected cultivation. The method includes the steps of firstly, collecting the indoor and outdoor meteorological factor data of the greenhouse of the grapes under protected cultivation; secondly, using manual-control experiments to investigate the leaf powdery mildew emergence grade of the grapes under protected cultivation under different meteorological conditions; thirdly, analyzing the relation between the powdery mildew emergence and development grade and the meteorological conditions so as to determine the key meteorological factors influencing the emergence and development of the powdery mildew of the grapes under protected cultivation; fourthly, building a forecasting model of the key meteorological factors influencing the emergence and development of the powdery mildew of the grapes under protected cultivation in the greenhouse; fifthly, determining the meteorological index range of different grades of powdery mildew, and building a model for forecasting the leaf powdery mildew potential of the grapes under protected cultivation. By the method which is simple, practicable and good in forecasting effect, the emergence and development grades of the leaf powdery mildew of the grapes under protected cultivation can be forecasted accurately.

The invention discloses an LncRNA and disease association relationship prediction method based on Naive Bayes, which comprises: according to known data sets of an MiRNA and disease association relationship, an MiRNA and LncRNA association relationship and an LncRNA and disease association relationship, constructing a complex network based on the three association relationships; in the complex network, finding common neighbor nodes of LncRNA nodes and disease nodes; by a probability model based on Naive Bayes, calculating a probability that the LncRNA nodes and the disease nodes, which have thecommon neighbor nodes, are connected so as to obtain a similarity value of LncRNA node and disease node pairs. According to the invention, a plurality of association relationships of a plurality of databases are integrated so as to establish more connections for the to-be-predicted LncRNA nodes and disease nodes in the constructed complex network, thereby improving a prediction effect of the LncRNA and disease association relationship.

The invention discloses a dynamic statistics combined sub-seasonal prediction method based on low-frequency increment space-time coupling, and the method comprises the steps: selecting tropical and tropical outside atmosphere abnormal signals as a prediction factor variable, taking the low-frequency increment of the variable as a prediction object and a prediction factor, and eliminating the interference of a weather change rate and a seasonal change rate. On one hand, a synchronous physical relationship between a forecast factor and a forecast quantity increment is considered, a singular value decomposition statistical method is utilized to find a high coupling mode of a synchronous forecast factor increment and a forecast quantity increment, and a multiple linear regression method is adopted to establish a sub-season prediction model based on a physical mechanism. On the other hand, by means of the advantage that the dynamic mode has a good forecasting effect on the sub-season tropical and extra-tropical atmosphere abnormal modes, the time coefficient (namely, the forecasting factor) of the tropical and extra-tropical atmosphere abnormal high coupling modes predicted by the dynamic mode is substituted into the forecasting model, and a power-statistics combined sub-season prediction model is further constructed to predict meteorological elements.

The invention discloses a tropical cyclone landing process gale forecasting system based on an ensemble forecasting model. The system comprises the following steps that a generalized initial value construction module constructs generalized initial values of a plurality of variables which have influence on a forecast quantity, and sends the generalized initial values to an initial value similarity judgment module; an initial value similarity judgment module discriminates the similarity of each single variable contained in the generalized initial value, determines n optimal similar TC and sends the n optimal similar TC to a gale ensemble forecasting module; the gale ensemble forecasting module obtains the n optimal similar TC to form n TC process gale fields, and the n TC process gale fields are gathered to obtain process gale information of the target TC; and the optimal forecasting scheme selection module calculates a forecasting accuracy TS score in a selected gale speed grade threshold by using the process gale information of the target TC, and obtains an optimal forecasting scheme model of the process gale of the target TC according to the forecasting accuracy TS score. The system has good forecasting performance on the gale in the TC landing process.

The invention relates to an improved ELM algorithm-based capillary quality forecasting method. The method comprises the steps of collecting a plurality of sets of historical field data during the capillary perforation process so as to construct a training set; according to collected field data, determining the input layer, the output layer and the hidden layer of an integrated ELM network; in combination with a plurality of common excitation functions, determining an excitation function of the integrated ELM network in the weight setting manner; optimizing each weight in the excitation function of the integrated ELM network by adopting a genetic algorithm, and obtaining an optimal excitation function; adopting the training set for training the integrated ELM network, and completing the construction of the integrated ELM network; inputting data obtained during the actual production process into each sub-network of the integrated ELM network, and obtaining the output result of each sub-network so as to obtain the output prediction result of the integrated ELM network, namely the quality prediction result of a capillary tube. According to the invention, the rapid performance of an ELM model and the robustness of an integration method are inherited. As a result, the quality of the capillary tube can be forecasted more accurately.