44results about How to "Reduce forecast bias" patented technology

The invention discloses a drug target relation prediction method based on a drug substructure and molecule character description information. Drug substructure information, molecule character description information and known drug target relations are obtained through a database; similarity matrixes among drugs are independently established according to the drug substructure information, the molecule character description information and the known drug target relations; the various established similarity matrixes are integrated into a final drug similarity matrix according to a weight set; and the drug target relations are predicted based on the feature that targets of the similar drug targets are also similar. According to the method, the similarities are established only according to the drug molecule character description information and the substructure information independent of information such as the sequences of the targets, the target relation prediction can be carried out on new drug compounds, and massive manpower and material resources consumed by biochemistry experiments are avoided. An experiment result shows that according to the method, the drug target relations of can be predicted accurately.

The invention discloses a queuing theory-based data transmission bandwidth prediction method in the intelligent power distribution and consumption business, aiming at the problem that a great deviation exists between a predicted value of a bandwidth and an actual demand value easily caused by the traditional bandwidth prediction method. The method comprises the following steps of firstly acquiring intelligent power distribution and consumption business QoS (quality of service) requirement parameters from an application layer as basic data for bandwidth prediction, and converting the QoS parameters of an upper layer into queuing theory model parameters through the queuing theory; secondly determining a queuing theory-based transmission bandwidth prediction model according to the QoS constraint condition of the electric power telecommunication business, and meanwhile converting the prediction model into a linear constraint nonlinear programming solving model to obtain the optimal solution; finally obtaining a smallest predicted transmission bandwidth capable of ensuring the QoS requirement of the system business. According to the method, the predicted band width is obtained by modeling and solving through intelligent power distribution and consumption business QoS parameter mapping and the queuing theory, so that the bandwidth prediction deviation is reduced, and the system bandwidth utilization rate is optimized.

The invention discloses a wind power prediction system for an extreme scene in the field of wind power prediction, and the system comprises a data collection server, a database server, a switch, a PCworkstation, a wind power prediction server, and a power grid dispatching center. The data acquisition server is connected with the numerical weather forecasting system, the anemometer tower, the booster station SCADA server, the fan SCADA server and the three-dimensional laser radar measurement system through a communication network, and the data acquisition server is connected with the PC workstation, the wind power prediction server and the database server through a switch through the communication network; and the switch is connected with the power grid dispatching center through the datainterface server. By increasing the input information amount, reducing the prediction deviation and adopting a physical, statistical and learning combined prediction method, a combined prediction model with small prediction error and high calculation efficiency is established, the calculation of economic dispatching is enhanced by selecting an extreme scene of a load, and a central scene sample isconsidered, so that the economy of the system is ensured.

The invention discloses a building electricity consumption prediction method and system based on Stacking model fusion, and belongs to the field of building electricity consumption prediction. According to the method, multiple regression models are integrated by adopting a Stacking model fusion algorithm, an electricity consumption Stacking integrated model is constructed, the advantages of the multiple models are integrated, and prediction deviation is reduced; for buildings with unstable electricity consumption, the electricity consumption Stacking integrated model is trained by utilizing multiple influence factors such as historical electricity consumption, temperature, wind power, humidity and time information, so that the prediction accuracy is improved, managers of the buildings caneffectively control the energy consumption of the buildings, and the situation that the difference between the electricity consumption and the estimated electricity consumption is too large is avoided; according to the invention, reasonable estimation and purchase are carried out when a building manager participates in electricity market transaction, so that the building manager can effectively control electric charge expenditure, electricity selling arrangement of an electric power department or an electricity selling company is facilitated, the effects of energy conservation and emission reduction can be achieved, and good social benefits and economic benefits are achieved.

The invention relates to the technical field of data processing, and provides a time sequence anomaly point detection method and device. The method comprises the steps of: training the regression model of a time sequence through a training set; predicting a sequence value at the current time according to the regression model obtained by training and an input time sequence before the current time,and performing anomaly detection on a sequence value at the current time obtained by observation according to the sequence value at the current time obtained by prediction; and, according to an anomaly detection result, when the sequence value at the current time obtained by observation is considered as anomaly, replacing the sequence value at the current time obtained by observation by the sequence value at the current time obtained by prediction, and continuously performing anomaly point detection on the next time of the time sequence. In a time sequence point anomaly detection task, a regression prediction method is adopted; an abnormal value is replaced by a prediction value; prediction deviation is reduced as much as possible; and the detection accuracy rate is increased.

The invention discloses a seasonal commodity demand prediction method based on time sequence decomposition. The method comprises the following steps: firstly, separating a peak sequence s1 and a conventional value sequence s2 from historical demand data based on a statistical method; secondly, based on the peak sequence s1, marking whether the training data is a peak demand or not; predicting a peak occurrence probability p by using a composite classifier consisting of two classifiers; and calculating a peak probability threshold value alpha by using recent historical data, carrying out regression strategy selection based on the peak prediction probability p and the peak probability threshold value, if p is greater than alpha, carrying out peak demand prediction by using a K neighbor model, otherwise, carrying out non-peak demand regression prediction by using a random forest model. According to the method, through seasonal peak probability modeling, seasonal demands are predicted by utilizing a plurality of regression models, sudden seasonal commodity peak values are effectively coped with, meanwhile, the accuracy of peak value prediction is greatly improved, and favorable supportis provided for enterprises to purchase seasonal commodities.

The invention provides an energy consumption prediction method and device based on complementary fuzzy neural network. The method comprises steps of: acquiring energy consumption historical data corresponding to energy to be tested; classifying and structuring the energy consumption historical data; performing gray treatment on the structured energy consumption historical data and screening out valid historical data; normalizing the valid historical data; performing fuzzy processing on the normalized data; inputting the data subjected to the fuzzy processing into the fuzzy neural network which predicts an energy consumption predicted value corresponding to the data subjected to the fuzzy processing; performing anti-normalization processing on the energy consumption predicted value; performing whitening processing on an anti-normalization processing result to obtain and output a target predicted value. The energy consumption prediction method and device based on complementary fuzzy neural network may increase the accuracy of the energy consumption prediction results.

The invention provides an air quality prediction method for multi-task learning based on multi-dimensional secondary feature extraction. The air quality prediction method comprises the steps of data acquisition, data preprocessing, pollutant selection, establishment of a convolutional neural network model and a long-term and short-term memory network model for multi-dimensional secondary feature extraction, establishment of a multi-task learning model for multi-dimensional secondary feature extraction and verification. According to the air quality prediction method, the problem that only timeinternal correlation and space internal correlation are considered during traditional spatio-temporal data modeling, and spatio-temporal correlation is not considered is solved. According to the air quality prediction method, influence information related to pollutant values is considered from three perspectives of space, time and space-time, and the prediction deviation is reduced by learning themutual influence among multiple time and space tasks through multi-task learning, so that the prediction precision of time and space models is more accurate.

The embodiment of the invention discloses a sample acquisition method and a device applied to model training. The method comprises the following steps: randomly sampling a data source to obtain a training sample set; training the model according to the training sample set to obtain a prediction error rate of the training sample set; re-determining the sampling rate of the data source according tothe prediction error rate, and sampling the data source according to the re-determined sampling rate to obtain an updated training sample set; iteratively executing the steps of training the model according to the updated training sample set, re-determining the sampling rate of the data source according to the obtained prediction error rate, and obtaining the updated training sample set accordingto the re-determined sampling rate; and obtaining the updated training sample set as a target sample set. According to the method of the embodiment of the invention, the target sample set close to thefeature distribution of the prediction sample set can be selected from the data source, and the target sample set is used for training the model.

The invention provides a load prediction method and device based on a recurrent neural network and a meta-learning strategy, and the method comprises the steps: receiving to-be-predicted target time of a target energy system; obtaining historical data of the target energy system before the target time, and extracting time sequence characteristics of the historical data by adopting a recurrent neural network; selecting a load prediction algorithm matched with the time sequence characteristics from a plurality of algorithm models according to a classification model; and outputting the load valueof the target energy system at the target time by using the load prediction algorithm. Through the method and the device, the technical problems of high time consumption and low accuracy in attempting various prediction algorithms when the load prediction algorithm is adopted to predict the energy load in the prior art are solved.

The invention provides a whole road network traffic state prediction method and device, and relates to the technical field of intelligent traffic, and the method comprises the steps: obtaining the traffic flow data of a detection road section and a historical travel OD matrix of a whole road network; determining predicted traffic flow data of the detection road section in a preset time interval according to the traffic flow data, and determining a traffic flow deduction operation state of the whole road network in the preset time interval according to the historical travel OD matrix; determining an optimization model according to the predicted traffic flow data and the deduced traffic flow operation state; and determining a prediction OD according to the optimization model and a preset constraint condition, and performing state prediction of the whole road network. According to the invention, the problem that only a single type of traffic flow parameters can be predicted in the prior art is overcome, and the traffic state prediction problem of a road section without detection data in the road network is solved. In addition, mutual verification of prediction results of various traffic states such as OD, speed and flow can be realized, prediction deviation of a single type of data is reduced, and prediction accuracy of the traffic states is improved.

The invention relates to a steel plate roller abrasion loss prediction method and system, and relates to the field of steel rolling. According to the method, training sample data corresponding to influence factors influencing rolling kilometers in the sample data is determined through a grey relational analysis method, and input variables of the neural network are optimized; an extreme learning machine neural network is trained by using training sample data, a neural network model is established, the prediction precision of the rolling kilometers is improved, the rolling kilometers of each setof working rollers are predicted by using the trained neural network model, and then the predicted rolling kilometers are used for predicting the abrasion loss of the working rollers. The predictiondeviation of the roller abrasion loss is reduced, and the poor plate shape caused by excessive abrasion of the working roller is avoided.

The invention provides a load forecasting method and a device based on a neural network, wherein, the method comprises the following steps of: receiving a time period to be forecasted; inputting the time period to a neural network model for predicting an energy load, wherein the neural network model is a radial basis function neural network trained based on a hybrid particle swarm optimization algorithm; predicting an energy load value at the time period using the neural network model. The invention solves the technical problem in the prior art that the accuracy rate is low when the single load forecasting algorithm is used for forecasting the energy load.

The method is suitable for the technical field of image / video compression, and provides an image-level JND threshold prediction method and device, equipment and a storage medium. The method comprisesthe following steps: performing perception distortion discrimination on the to-be-detected image and the compressed image in the compressed image set corresponding to the to-be-detected image throughthe trained multi-classification perception distortion discriminator. Obtaining a perception distortion discrimination result set; performing fault-tolerant processing on the perception distortion discrimination result set through an image-level JND search strategy; according to the method, the image-level JND threshold of the to-be-detected image is obtained through prediction, so that the prediction deviation of the image-level JND threshold is reduced, the prediction accuracy of the image-level JND threshold is improved, and the predicted JND threshold is closer to the perception of a humaneye vision system on the quality of the whole image.

The invention provides a load forecasting method and device based on a neural network, wherein, the method comprises the following steps of: receiving a time period to be forecasted; Inputting the time period to a neural network model for predicting an energy load, wherein the neural network model is composed of a linear prediction model and a non-linear prediction model; Predicting an energy loadvalue at the time period using the neural network model. The invention solves the technical problem in the prior art that the accuracy rate is low when the single load forecasting algorithm is used for forecasting the energy load.

The invention discloses an optimization method for predicting a refraction adjustment value in an SMILE refractive surgery and system. The method comprises the following steps: an optimal machine learning algorithm is used, the preoperative parameters of training samples and the post-operative refraction are used as training attributes, and a prediction model is trained and generated; as for a to-be-predicted new case, the post-operative refraction is used as a target attribute and set as an ideal value, and in cooperation with the preoperative parameters of the to-be-predicted new case, the prediction model is used to generate a Nomogram prediction value. The method can effectively improve the prediction accuracy of the Nomogram prediction value, dependence on experts in a preoperative plan making process can be reduced, the professional threshold for preoperative plan making is reduced, and the preoperative plan making accuracy is further enhanced.

The invention discloses an age prediction method based on ensemble learning of an intestinal flora prediction model, which comprises the following steps: acquiring original data of human intestinal flora metagenomics, performing quality control on the acquired genome data, acquiring an abundance table of intestinal flora species composition and metabolic pathway composition, and constructing a sample data set; performing feature selection on the sample data set; constructing a multi-class age prediction model by using the screened features in combination with host regional information, determining hyper-parameters which enable the prediction model to be higher in precision by using grid search, and training and predicting each optimal prediction model to obtain an integrated age prediction method; and finally, predicting the age of the sample by using the determined intestinal flora characteristics and an integrated age prediction method, and determining age-related key species and pathways through characteristic interpretation. According to the invention, an integrated learning method is adopted, so that the age prediction accuracy is effectively improved; the regulation of intestinal flora can be directionally guided, so that anti-aging intervention is realized.

The invention discloses a method for predicting the service life of an electron multiplier of a cesium atomic clock, and the method comprises the steps: obtaining the change rule of the working voltage of the electron multiplier of the cesium atomic clock based on the analysis of the long-term operation data of the cesium atomic clock, and obtaining the service life of a newly developed electron multiplier of the cesium atomic clock through the analysis and calculation according to the rule. The blank of the method for predicting the service life of the electron multiplier is filled, the prediction deviation is less than 10%, and the service life of the cesium atomic clock can be directly deduced.

The embodiment of the invention provides an air rudder thermal environment modeling method and device and a storage medium, and the method comprises the steps of obtaining the thermal environment parameters of an air rudder in a flight trajectory; calculating the size of an air rudder interference area according to the thermal environment parameters, the shape of the air rudder and the position ofthe air rudder on the rocket body; calculating air rudder grid scale parameters according to the size of the air rudder interference area and a preset air rudder grid division model; and meshing thesurface of the air rudder according to the mesh scale parameters. According to the air rudder thermal environment modeling method and device and the storage medium provided by the embodiment of the invention, the thermal environment prediction deviation can be reduced, so that the reasonable grid amount is ensured, and the aerial rudder thermal environment simulation precision and calculation efficiency are improved.

The invention provides a load prediction method and device based on an auto-encoder and a meta-learning strategy, and the method comprises the steps: receiving to-be-predicted target time of a targetenergy system; obtaining historical data of the target energy system before the target time, and extracting time sequence characteristics of the historical data by adopting the auto-encoder; selectinga load prediction algorithm matched with the time sequence characteristics from a plurality of algorithms according to a classification model obtained by meta-learning; and outputting the load valueof the target energy system at the target time by using the load prediction algorithm. Through the method and the device, the technical problems of high time consumption and low accuracy in attemptingvarious prediction algorithms when the load prediction algorithm is adopted to predict the energy load in the prior art are solved.

The invention relates to a duck sebum character living body prediction method and application thereof. The method comprises the following steps: feeding a batch of ducklings in stages, recording the weight and feed intake of each duck in each stage, calculating the residual feed intake and feed conversion rate of each duck after the feeding is completed, obtaining a body size index and a serum biochemical index of each duck, selecting a batch of ducks from the bred ducks as a to-be-detected group, slaughtering the ducks to obtain sebum traits of the ducks in the to-be-detected group, calculating a multiple linear regression equation of the sebum traits of the ducks by using statistical analysis software, and predicting the residual group sebum traits of the raised ducks by using the multiple linear regression equation. According to the method, the sebum traits of the live ducks can be quickly and accurately estimated, the elimination probability of excellent individuals in breeding is greatly reduced, and the breeding efficiency is greatly improved.

The invention provides a fundus image processing method, a model training method and equipment, and the method comprises the steps: extracting a single-channel fundus image from a multi-channel fundus image; determining a maximum pixel value and a minimum pixel value in the single-channel eye fundus image; processing pixel values in the single-channel eye fundus image by using the maximum pixel value and the minimum pixel value; and synthesizing the processed single-channel fundus images into a multi-channel fundus image.

The embodiment of the invention provides a prediction method for a demand side heat supply load. The prediction method comprises the following steps that S1, collecting regional heat supply historicaldata; S2, performing data cleaning on the heat supply historical data to realize data abnormal value elimination and blank data filling; S3, establishing a heat supply load model; S4, performing heatsupply load prediction. According to the method of the invention, the calculation process is simplified under the condition that the precision is not lost, so that the lightweight mobile equipment can be embedded into the method, the prediction equipment can be conveniently installed and deployed on a demand side (a user side), and the prediction precision and credibility are improved. Predictionwith relatively high confidence can be realized only by depending on a load time sequence, and the operability and real-time performance of the algorithm in an industrial application environment areimproved.

The invention discloses a traffic demand forecasting method and system for a dynamic road network. According to the invention, multi-source traffic data is collected and integrated all the time; dynamic grid data is formed through overlapping traffic flow data, geographic information data and traffic management information on the road network; and a data forecasting model is established on the basis of dynamic road network data. The method has characteristics of good real time performance, high accuracy and the like, can improve jamming study and judge precision effectively, so as to provide reasonable traffic demand data (queue, flow, speed, occupancy rate and the like) for signal optimization. The system can improve an optimization model and relieves traffic jams.

According to the coal reservoir permeability prediction method, surrounding rock stress, gas pressure, temperature and compressive strength serve as input values, permeability serves as an output value, and data are divided into a training set and a test set after being cleaned; a multiple linear regression model, a BP neural network model and an SVM model are established by using the training set, and the test set is predicted; and then a joint prediction model is established, and the most suitable weight is obtained by using a particle swarm algorithm. According to the method, multiple machine learning models are combined, the advantages of all the models are extracted, a more accurate combined prediction model is obtained, prediction deviation is reduced, the combined prediction model has good robustness, and even if deviation occurs due to model assumption, only small influence can be generated on algorithm performance. No matter how the index number and the sample number of the data change, the permeability can be accurately predicted; and the requirement for the training data volume is loose, and a precise prediction model can be obtained through small sample data training.

The invention discloses a method for predicting the bandwidth of intelligent power distribution business data transmission based on queuing theory. Aiming at the problem that the existing bandwidth prediction method is likely to cause a large deviation between the predicted value of the bandwidth and the actual demand value, the method first obtains the bandwidth from The QoS requirement parameters of the intelligent power distribution service at the application layer are used as the basic data for bandwidth prediction, and the high-level QoS parameter indicators are converted into queuing theory model parameters through queuing theory; secondly, the transmission bandwidth prediction based on queuing theory is determined according to the QoS constraints of power communication services At the same time, the prediction model is converted into a linear constraint nonlinear programming solution model to obtain the optimal solution; finally, the minimum prediction transmission bandwidth that can guarantee the system service QoS requirements is obtained. The method obtains the predicted bandwidth through intelligent power distribution service QoS parameter mapping and queuing theory modeling, which can reduce bandwidth prediction deviation and optimize system bandwidth utilization.

The invention discloses a method and system for improving the applicability of a BJT device mismatch model. The method comprises the following steps of 1, adding mismatch selection factors mismod in different locating and wiring modes into the mismatch model; 2, according to the mismatch selection factors mismod, selecting different locating and wiring modes for simulation. By means of the methodand system, the applicability of the bipolar transistor BJT device mismatch model can be improved, and the deviation between actual measurement and prediction is reduced.

The present disclosure discloses a prediction method, device and automatic prediction system for estimating the arrival time of a track, wherein the method includes: obtaining real-time position information of aircraft entering the prediction area; triggering a pre-established prediction model according to the real-time position information of the aircraft The electronic fence in , calculate the first predicted time T that takes time from the current electronic fence where the aircraft is located to the time it takes to receive the signal to align with the runway d ;According to the alignment information of the aircraft to the runway, obtain the second predicted time T from receiving the alignment signal to actually landing. s ; According to the first prediction time T d and the second predicted time T s , and calculate the arrival time of the aircraft based on the sum of the current time. Therefore, by implementing the prediction scheme disclosed in the present disclosure, the prediction deviation caused by unexpected factors can be effectively reduced, the prediction accuracy of the arrival time of airport flights can be improved, the waste of airport security resources can be avoided, and the security risks caused by inaccurate prediction of arrival time can be eliminated. Hidden danger.

The invention discloses a low-carbon optimized operation method considering cooperation of a carbon capture power plant and photovoltaic power generation. The method comprises the following steps: 1, obtaining a photovoltaic power generation output scene; 2, clustering the photovoltaic power generation output scenes; 3, establishing a power system low-carbon optimization operation objective function considering cooperation of the carbon capture power plant and photovoltaic power generation; 4, determining a power system low-carbon optimization operation constraint condition considering cooperation of the carbon capture power plant and photovoltaic power generation; and 5, solving an electric power system low-carbon optimization operation model formed by the objective function and the constraint condition by using Cplex to obtain the starting and stopping states of an electric power system unit when the total efficiency coefficient is maximum. According to the method, the time sequence correlation of the photovoltaic prediction error can be considered, cooperative operation of the carbon capture power plant and photovoltaic power generation is promoted, and the overall operation efficiency of a power system is improved.

The invention provides a method for recommending a target object to a user. The method comprises that: an object feature of the target object is generated according to at least one of an object control feature and an object random feature of the target object; the object control features and the object random features are obtained after multi-task alternate training is performed on a control matching model and a random matching model; in the multi-task alternate training, a control sample is used as an input sample of a control matching model, a random sample is used as an input sample of a random matching model, and an optimization target is achieved by modifying an object control feature of a target object in the control sample or modifying a value of an object random feature of the target object in the random sample, wherein the optimization target comprises that the difference between the object control feature and the object random feature of the same target object is as large aspossible; and the user features and the object features of the target object are input into a matching model, and the target object recommended to the user is determined according to the matching degree of the user features and the object features output by the matching model.