Research On Interval Estimation Method Of Small Sample Flight Fuel Consumption Data

To check the reasonableness of the fuel consumption data reported by aircraft operators,the current third-party carbon emission monitoring agencies usually verify fuel consumption based on the point estimation method,using the predicted fuel consumption value corresponding to the range as the threshold value.But it is reasonable for the range to be larger or smaller than the predicted value within a certain interval,so it is more reliable to use the interval estimation method to give a reasonable prediction range.At the same time,the fuel consumption sample size of some flight segments is relatively scarce,and the traditional fuel consumption prediction model needs to rely on a large number of QAR datasets.To address the above problems,Mixed Kernel Gaussian Process Regression(MK-GPR)and Multi-Source Domain Adaptation(MS-DA),which have the ability of interval quantification,are selected.The former improves the prediction accuracy of the model under the limited training data,while the latter reduces the number of training samples required by the model from the data level.By combining the two methods,the multi-source domain adaptation interval estimation model(MS-DA-GPR)is proposed to improve the quality of small-sample aircraft fuel consumption intervals from two perspectives simultaneously.The main work is as follows:(1)The analysis of small-sample fuel consumption data shows that fuel consumption has strong nonlinearity as well as uncertainty,while several characteristic covariates with high correlation with fuel consumption are identified and used in the subsequent model prediction session.Several interval estimation methods and inter-domain distance measures are analyzed and compared,and Gaussian process regression(GPR)is selected to quantify the interval and maximum mean difference(MMD)is selected to measure the inter-domain distance.(2)To address the problem that there is no uniform standard in the current Gaussian process regression method when selecting the kernel,combined with its data characteristics,by systematically analyzing the structure of the kernel and the hybrid law,the sum of linear kernel and square exponential kernel is selected as the hybrid kernel,and the hyperparameters in the kernel are trained and solved using the gradient descent method to obtain the posterior distribution,to establish the hybrid kernel Gaussian process regression model(MK-GPR),which enhances the generalization ability of the model while ensuring the local learning ability compared with the single kernel,and can better capture the overall linear features and local nonlinear features in the fuel consumption data.(3)To address the problem that the traditional coverage width criterion CWC cannot reflect the degree of deviation of points outside the interval,the improved comprehensive width criterion ICWC is established by adding the interval mean relative deviation index MRD,which can effectively ensure the reliability of the interval estimation index and improve the model prediction accuracy.(4)To address the problem that traditional transfer learning is prone to "negative transfer",the problem of improving the similarity between source and target domains is refined into the domain adaptation problem of reducing the marginal probability distribution and conditional probability distribution between them by the joint probability distribution adaptation algorithm(JDA)based on the algorithm idea of MMD,and the semi-supervised transfer component Analysis(SSTCA)method is used to solve the optimal feature transformation matrix and construct the shared subspace of source and target domains.By using the MMD method to calculate the degree of difference between adjacent segments and set the transfer weight of each source domain,we propose a combined multi-source domain adaptation interval estimation model(MS-DA-GPR),which can effectively reduce the possibility of "negative transfer" and improve data resource utilization and model training efficiency compared with single-domain adaptation.This model can effectively reduce the possibility of "negative transfer" and improve the data resource utilization and model training efficiency compared with single domain adaptation,and help reduce the number of training samples required by the model for the target domain.The two methods are applied to a small-sample fuel consumption data set using actual A330 operation data from a domestic airline.Two sets of comparison experiments are designed to verify the feasibility of the proposed method: first,to compare the effects of single kernel function and mixed kernel function on the prediction results in the GPR model;second,to compare the effects of single-sector domain adaptation and multi-sector domain adaptation combination models on the prediction results.The results show that the prediction results obtained based on the MS-DA-GPR model have higher interval coverage,smaller mean interval width,cumulative deviation,comprehensive evaluation coefficient in general,and better prediction results.